University of Southern California Dissertations and Theses

Sensor-based mobile health approaches for personal air pollution and pediatric asthma studies

(USC Thesis Other)

Sensor-based mobile health approaches for personal air pollution and pediatric asthma studies

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Sensor-Based Mobile Health Approaches for Personal Air Pollution and
Pediatric Asthma Studies

by

Hua Hao

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
DOCTOR OF PHILOSOPHY
EPIDEMIOLOGY

December 2020

Copyright 2020 Hua Hao
DEDICATION
This dissertation is dedicated to the three most important people in my life:
my mother Hui Xia;
my father Manchao Hao; and
my husband Varun Pattisapu.

iii
ACKNOWLEDGEMENTS

I was only able to complete this dissertation because of the supports from my family and
friends. I would like to thank my parents Hui Xia and Manchao Hao, my husband Varun
Pattisapu, my parent-in-law Ravi Pattisapu and Shree Pattisapu, and the rest of all my family,
without your encouragement and support, I cannot complete my PhD.
I would like to start by thanking my doctoral dissertation committee chair and mentor Dr
Rima Habre for guiding, supporting and encouraging me through the most difficult periods.
Thank you so much for your kindness and patience! I also would like to thank my committee
members Dr Sandrah Eckel, Dr Theresa Bastain and Dr Laura Loyola for making sure I
conducted the best science I could and always being there for me when I needed help.
I would also thank you to all my beloved friends: Ziwei Song, Siyu Song, Kaili Ding,
Yang Yu, Lu Gao, Cuixiang Cao, and Shiying Gong. Thank you so much for being great
companions through the program and always being there with me.

iv
Table of Contents
DEDICATION ............................................................................................................................................... ii
ACKNOWLEDGEMENTS .......................................................................................................................... iii
LIST OF TABLES ...................................................................................................................................... viii
LIST OF FIGURES ...................................................................................................................................... ix
ABSTRACT .................................................................................................................................................. x
1 Chapter 1: Introduction .......................................................................................................................... 1
1.1 Introduction .................................................................................................................................. 1
1.2 References: ................................................................................................................................... 5
2 Chapter 2: Background Literature ......................................................................................................... 9
2.1 Pediatric asthma ........................................................................................................................... 9
2.2 The Burden of Pediatric Asthma ................................................................................................ 10
2.3 Asthma Control .......................................................................................................................... 11
2.3.1 Definition of Pediatric Asthma Control ................................................................................. 11
2.3.2 Risk Factors for Worsening Pediatric Asthma Control ......................................................... 12
2.3.3 Asthma Phenotypes ................................................................................................................ 13
2.3.4 Triggers for Acute Pediatric Asthma Exacerbation ............................................................... 13
2.4 Particulate Matter Air Pollution Exposure as a Risk Factor for Acute Asthma Exacerbation in
Children ................................................................................................................................................... 14
2.5 Personal Air Pollution Exposure Assessment ............................................................................ 15
2.5.1 Personal Air Pollution Exposure Assessment Methods ......................................................... 15
2.5.2 Microenvironmental Modeling Approach ............................................................................. 16
2.6 Mobile Health (mHealth) and Sensor-Based Approaches for Personalized Exposure and Health
Assessment .............................................................................................................................................. 18
2.6.1 Three Domains Framework ................................................................................................... 18
2.6.2 Ubiquitous Availability and Use of Smartphones ................................................................. 18
2.7 Conclusion .................................................................................................................................. 20
2.8 References .................................................................................................................................. 21
3 Chapter 3: Microenvironmental Time-Activity classification using smartphone sensor data for
personal air pollution exposure assessment ................................................................................................. 29
3.1 Abstract ...................................................................................................................................... 29
3.2 Introduction ................................................................................................................................ 30
3.3 Methods ...................................................................................................................................... 32
3.3.1 Data Collection ...................................................................................................................... 32
v
3.3.2 Data Pre-Processing ............................................................................................................... 34
3.3.3 Additional Feature Extraction ................................................................................................ 37
3.3.4 Machine Learning Algorithms Evaluated .............................................................................. 37
3.3.5 Initial Model Training and Testing ........................................................................................ 38
3.3.6 Model Prediction Accuracy Assessment ............................................................................... 39
3.3.7 Training Data Considerations and Feature Importance Ranking ........................................... 39
3.3.8 Additional Model Prediction Accuracy Assessment Under Various Scenarios .................... 40
3.3.9 Independent Subject (S2) Tests with Variable Degrees of Resemblance in Time-Activity
Patterns 40
3.3.10 Same Subject (S1) and Independent Subject (S2) Tests Considering Temporal
Autocorrelation .................................................................................................................................... 40
3.3.11 Same Subject (S1) and Independent Subject (S2) Tests Considering Spatial
Autocorrelation .................................................................................................................................... 41
3.3.12 Same Subject (S1) and Independent Subject (S2) Tests Considering Spatiotemporal
Autocorrelation .................................................................................................................................... 42
3.4 Results ........................................................................................................................................ 43
3.4.1 Optimal Machine Learning Model for Microenvironment Classification ............................. 43
3.4.2 Model Prediction Accuracy in Independent Subject (S2) Tests with Variable Degrees of
Resemblance in Time-Activity Patterns .............................................................................................. 45
Table 3.4. Model prediction accuracy in independent subject tests (S2) under natural living
(unscripted) and scripted time-activity patterns................................................................................... 46
3.4.3 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Temporal Autocorrelation ............................................................................................... 46
3.4.4 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Spatial Autocorrelation ................................................................................................... 47
3.4.5 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Spatiotemporal Autocorrelation ...................................................................................... 48
3.5 Discussion .................................................................................................................................. 48
3.5.1 Mobile Health (mHealth) Approaches to Microenvironment Classification Problems ......... 49
3.5.2 Optimal Model Performance and Feature Importance........................................................... 50
3.5.3 Generalizability of the Optimal Model Across People for Larger Scale Applications .......... 51
3.5.4 Cross Validation Strategies That Consider Spatiotemporal Autocorrelation Inherent in High
Resolution Sensor Data........................................................................................................................ 52
3.5.5 Limitations, Strengths and Future Directions ........................................................................ 53
3.6 References: ................................................................................................................................. 55
3.7 Supplement ................................................................................................................................. 61
vi
4 Chapter 4: Association of Personal and Ambient Air Pollution and Lung Function Decrements in
Asthmatic Children in Log Angeles ............................................................................................................ 72
4.1 Abstract ...................................................................................................................................... 72
4.2 Introduction ................................................................................................................................ 73
4.3 Materials and Methods ............................................................................................................... 74
4.3.1 Study Design and Population ................................................................................................. 74
4.3.2 Asthma Outcomes .................................................................................................................. 76
4.3.3 Exposure Assessment............................................................................................................. 77
4.3.4 Covariate Information ............................................................................................................ 78
4.3.5 Statistical Analysis ................................................................................................................. 79
4.4 Results ........................................................................................................................................ 80
4.4.1 Descriptive Summaries .......................................................................................................... 80
4.4.2 Exposure and Health Analyses .............................................................................................. 84
4.5 Discussion .................................................................................................................................. 92
4.6 References: ................................................................................................................................. 96
4.7 Supplement ............................................................................................................................... 101
5 Chapter 5. Association of Personal and Ambient Air Pollution and Asthma Exacerbation in
Asthmatic Children in Los Angeles ........................................................................................................... 105
5.1 Abstract .................................................................................................................................... 105
5.2 Introduction .............................................................................................................................. 106
5.3 Materials and Methods ............................................................................................................. 107
5.3.1 Study Design and Population ............................................................................................... 107
5.3.2 Asthma Outcomes ................................................................................................................ 109
5.3.3 Exposure Assessment........................................................................................................... 111
5.3.4 Covariate Information .......................................................................................................... 112
5.3.5 Statistical Analysis ............................................................................................................... 113
5.4 Results ...................................................................................................................................... 114
5.4.1 Descriptive Summaries ........................................................................................................ 114
5.4.2 Exposure and Health Analyses ............................................................................................ 119
5.5 Discussion ................................................................................................................................ 124
5.6 References: ............................................................................................................................... 129
5.7 Supplement ............................................................................................................................... 132
6 Chapter 6: Conclusions and Future Directions .................................................................................. 135
6.1 Conclusions and Future Directions .......................................................................................... 135
vii
6.2 References ................................................................................................................................ 138

viii
LIST OF TABLES
TABLE

PAGE
Table 3.1 Dates and times of scripted and natural living (unscripted) data collection
from subjects S1 and S2.
33
Table 3.2 Characteristics of the 31 measured and calculated AndroSensor features
for S1 (N=21,296) by sensor category.
34
Table 3.3 Prediction performance (reported as balanced accuracy, %) of five ML
classifiers based on S1 data using 31 features (M1), 93 features (M2), and
1,023 features (M3).
44
Table 3.4 Model prediction accuracy in independent subject tests (S2) under
natural living (unscripted) and scripted time-activity patterns.
46
Table 3.5 Model prediction accuracy in same subject (S1) tests accounting for
temporal autocorrelation (DFC) compared to conventional CV (stratified
7:3 random selection).
46
Table 4.1 Study Group Characteristics. 80
Table 4.2 Distributions of Person-Day Level Lung Function Outcomes. 82
Table 4.3 Descriptive Statistics of 24-Hour Average (6AM-6AM) Air Pollutant
and Meteorology Measurements.
83
Table 4.4 Exposure (24-Hour Average 6AM-6AM) Correlation Matrix. 84
Table 4.5 Change in %predicted morning, evening or daily FEV1 (forced
expiratory volume in 1 second [L/s]) per standard deviation change in
pollutant exposure.
86
Table 4.6 Change in %predicted morning, evening or daily PEF (peak expiratory
flow rate [L/s]) (per standard deviation change) in pollutant exposure.
90
Table 5.1 Study Group Characteristics. 114
Table 5.2 Distributions of Person-Day Level Health Outcomes. 116
Table 5.3 Descriptive Statistics of 24-Hour Average (6AM-6AM) Air Pollutant
and Meteorology Measurements.
117
Table 5.4 Exposure (24-Hour Average 6AM-6AM) Correlation Matrix. 118
Table 5.5 Associations of Rate Ratio (per standard deviation change) Between Air
Pollutant Exposures and Daily Count of Rescue Inhaler Use.
119
Table 5.8 Associations of Odds Ratio (per standard deviation change) Between Air
Pollutant Exposures and Asthma Symptoms on lag 0 day.
121

ix
LIST OF FIGURES
FIGURE

PAGE
Figure 3.1 DFC nested CV workflow to account for temporal autocorrelation in
prediction performance evaluations.
41
Figure 3.2 Nested spatiotemporal CV framework to account for spatiotemporal
autocorrelation in prediction performance evaluations, adapted and
modified from. The color represents assigned spatial blocks and the shape
represents different days of data collection denoted by D (e.g., D1, D2,
etc.).
42
Figure 3.3 Top 10 most important features for overall and microenvironment class
prediction ranked by permutation feature importance method from the
optimal, final model in S1 (random forests algorithm with 31 features).

45
Figure 3.4 Spatial autocorrelation ranges for six selected features (A) and the
corresponding spatial blocks created by setting block size equal to the
median spatial autocorrelation range across these features (B).
47
Figure 3.5 Illustration of location accuracy changes in space (A) and in time (B)
between indoor and outdoor microenvironments.
51
Figure 4.1 Comparison between Collected Forced Expiratory Volume in One
Second (FEV1) and Clinic Spirometer Tested FEV1.
83
Figure 4.2 Association of daily air pollution exposure (lag 0) with % predicted
FEV1 and PEF. Effect estimates and 95% confidence intervals are
scaled to a standard deviation change in exposure.
88
Figure 5.1 Results of single- and two-pollutant models (rate ratio and 95% CI per
standard deviation increase in traffic-related pollutant exposure) for
daily count of rescue inhaler use in relation to traffic-related PM2.5, NO,
and NOx exposure in the last 24 hours (lag 0).
121

x
ABSTRACT
The American Public Health Association recognizes that asthma is one of the leading
chronic diseases among children in the United States. Children with asthma are significantly
burdened by asthma morbidity, including emergency department visits, hospitalizations, and
even deaths. A large proportion of the burden of asthma is attributed to treating the consequences
of poor asthma control, so asthma control becomes an important aspect of the evaluation and
management of the disease. The mechanisms leading to asthma exacerbation in children are
complex, but environmental exposures are largely presumed to be important. Environmental
asthma triggers include indoor and outdoor allergens, such as dust mites, animal allergens, molds
and pollens, and indoor and outdoor pollutants, including environmental tobacco smoke (or
secondhand smoke), chemical, combustion by-products, and ozone and particulate matter. While
there are a multitude of known environmental triggers responsible for poor long-term control of
asthma in children, there is limited knowledge in how these exposures contribute to acute
worsening of disease on an hourly or daily basis.
Furthermore, many of the above epidemiological studies relied on ambient air pollution
measurements or estimates (for example, measurements at nearest central site, or models of
outdoor residential air pollution) as a surrogate of true personal exposure, potentially introducing
exposure measurement error which may weaken statistical power to detect effects and introduce
bias or distort the estimation of true risk.
Therefore, our goal in this work is to develop and apply sensor-based mobile Health
(mHealth) approaches to improve time-activity classification and investigate acute triggers of
asthma exacerbation in children. We first develop a novel mHealth approach to improve time-
activity classification by using smartphone sensor data and machine learning techniques. Then,
based on our work in the Los Angeles Pediatric Research Using Integrated Sensor Monitoring
xi
System (LA PRISMS) Center, where we developed an informatics platform for wearable sensor-
based environmental health studies and deployed it in a panel study of pediatric asthma, we aim
to estimate the day-level and within-day acute effects of air pollution exposures on asthma
exacerbations (lung function decrements, use of rescue medication, increased odds of symptoms)
in children with moderate to severe asthma.

1
1 CHAPTER 1: INTRODUCTION
1.1 INTRODUCTION
The American Public Health Association recognizes that asthma is one of the leading
chronic diseases among children in the United States
1
. Children with asthma are significantly
burdened by asthma morbidity, including emergency department visits, hospitalizations, and
even deaths
2,3
. A large proportion of the burden of asthma is attributed to treating the
consequences of poor asthma control
4
, so asthma control becomes an important aspect of the
evaluation and management of the disease. Asthma control assessment involves controlling
current impairment which comprises a multitude of factors such as daily and nightly symptoms,
frequency of reliever medication use, level of activity, and overall quality of life (QoL). It also
takes into account future risk of exacerbations, permanent lung function impairment, and adverse
treatment effects
5
. Uncontrolled asthma is also related to worse performance metrics in both
physical exercise and lung function
6
. Therefore, according to the National Asthma Education
and Prevention Program (NAEPP) Expert Panel Report 3 (EPR-3)
7
and the 2016 Global
Initiative for Asthma (GINA) guidelines
8
, once the treatment is established, achieving asthma
control becomes the goal.
The mechanisms leading to asthma exacerbation in children are complex, but
environmental exposures are largely presumed to be important
9
. Environmental asthma triggers
include indoor and outdoor allergens, such as dust mites, animal allergens, molds and pollens,
and indoor and outdoor pollutants, including environmental tobacco smoke (or secondhand
smoke), chemical, combustion by-products, and ozone and particulate matter
10
. Many previous
studies have found air pollution from ozone, sulfur dioxide, nitrogen dioxide and particulate
matter (PM) may induce or aggravate asthma
11–14
PM of aerodynamic diameter < 2.5 m
2
(PM2.5) represents a combination of small particles and liquid droplets from vehicular exhaust,
coal and wood-burning and industrial activities
15
. Exposure to PM2.5 has been shown to
exacerbate children’s asthma, including reduced percent predicted forced expiratory volume in 1
second (FEV1)
16
, cough and wheeze symptoms
17
, and rescue medication use
18
.
While there are a multitude of known environmental triggers responsible for poor long-
term control of asthma in children, there is limited knowledge in how these exposures contribute
to acute worsening of disease on an hourly or daily basis. Furthermore, asthma is a
heterogeneous diseases with several different phenotypes which respond to and are exacerbated
by different factors and conditions, such as allergic asthma and aspirin-intolerant asthma
19
.
These phenotypes require an individualized approach to counseling children about environmental
exposures and prescribing a medication regimen tailored to each child’s asthma subtype
20
.
Taken together, these studies suggest a need for a more tailored and personalized approach to
investigating acute asthma triggers and their association with subsequent exacerbations.
Furthermore, many of the above epidemiological studies relied on ambient air pollution
measurements or estimates (for example, measurements at nearest central site, or models of
outdoor residential air pollution) as a surrogate of true personal exposure, potentially introducing
exposure measurement error which may weaken statistical power to detect effects and introduce
bias or distort the estimation of true risk
21
. The gold standard for assessing air pollution exposure
in human health studies is personal monitoring
22,23
, which consists of sampling air in the
breathing zone of a participant, using a worn monitor. However, collecting personal air pollution
measures for large populations remains a significant challenge due to feasibility, cost and burden
on participants
23
. The microenvironmental modeling approach assumes that total personal
exposure can be accurately estimated by accounting for how much time a person spends in key
3
microenvironments during a day, and the concentration of air pollution encountered in those
microenvironments, where the air pollution levels are assumed to be homogeneous in each
microenvironment
24
. The microenvironmental approach has been widely used in the
epidemiological literature as a more scalable approach to assess personal exposure than personal
monitoring. However, applying the microenvironmental modeling approach requires accurate
time-activity data, or information on how much time an individual spent within each key
microenvironment encountered during a day (for example, indoors, outdoors and in-transit)
25–27
.
Previous research has most commonly relied on time-activity diaries and questionnaires to
classify time spent in major microenvironments; however, these approaches are often
burdensome (for example, having to report time-activity every 15 minutes) and can suffer from
recall bias leading to unreliable data. Because of recent advances in mobile technology,
environmental sensors embedded within smartphones provide a unique opportunity to capture
time-activity information passively and more accurately in large population studies, given how
ubiquitous and widespread smartphone use has become recently
28,29
. Therefore, developing
methods or algorithms to automatically classify time-activity patterns with high accuracy and
low user burden is highly desirable to allow for more accurate personal air pollution exposure
estimation at population scales, in conjunction with advances in miniaturized, wearable sensor
technology and air monitoring devices
30
.
Therefore, our goal in this work is to develop and apply sensor-based mobile Health
(mHealth) approaches to improve time-activity classification and investigate acute triggers of
asthma exacerbation in children. We first develop a novel mHealth approach to improve time-
activity classification by using smartphone sensor data and machine learning techniques. Then,
based on our work in the Los Angeles Pediatric Research Using Integrated Sensor Monitoring
4
System (LA PRISMS) Center, where we developed an informatics platform for wearable sensor-
based environmental health studies and deployed it in a panel study of pediatric asthma, we aim
to estimate the day-level and within-day acute effects of personal PM2.5 exposures on asthma
exacerbations (lung function decrements, use of rescue medication, increased odds of symptoms)
in children with moderate to severe asthma.
In Chapter 2 below, we first review the background literature for this work. In the first
section, childhood asthma is reviewed, as well as the epidemiological evidence on air pollution
exposure and childhood asthma exacerbation. In the second section, methods for improving
personal air pollution exposure assessment in health studies, including advances in wearable and
miniaturized sensing technology and data science approaches are reviewed. In Chapter 3, we
present our novel mHealth approach to improve time-activity classification for personal exposure
assessment by using mobile sensor data and machine learning techniques. In Chapters 4 and 5
respectively, we examine the association of personal and ambient air pollution with pediatric
asthma exacerbations. In Chapter 6, we summary this dissertation and discuss the future
directions.

5
1.2 REFERENCES:
1. Zahran HS, Bailey CM, Damon SA, Garbe PL, Breysse PN. Vital signs: Asthma in
children — United States, 2001-2016. Morb Mortal Wkly Rep. 2018.
doi:10.15585/mmwr.mm6705e1
2. van Dellen QM, Stronks K, Bindels PJE, Öry FG, Bruil J, van Aalderen WMC. Predictors
of asthma control in children from different ethnic origins living in Amsterdam. Respir
Med. 2007. doi:10.1016/j.rmed.2006.08.002
3. Akinbami L. The state of childhood asthma, United States, 1980-2005. Adv Data. 2006.
4. Barnes PJ, Jonsson B, Klim JB. The costs of asthma. Eur Respir J. 1996.
doi:10.1183/09031936.96.09040636
5. Alzahrani YA, Becker EA. Asthma control assessment tools. Respir Care. 2016.
doi:10.4187/respcare.04341
6. Lang A, Mowinckel P, Sachs-Olsen C, et al. Asthma severity in childhood, untangling
clinical phenotypes. Pediatr Allergy Immunol. 2010. doi:10.1111/j.1399-
3038.2010.01072.x
7. National Asthma Education and Prevention Program. Expert Panel Report 3 (EPR-3):
Guidelines for the Diagnosis and Management of Asthma-Summary Report 2007. J
Allergy Clin Immunol. 2007.
8. GINA C científico de l. Global Strategy for Asthma Management and Prevention ( 2016
update ). Guía Gina Actual. 2016.
9. Etzel RA. How environmental exposures influence the development and exacerbation of
asthma. In: Pediatrics. ; 2003.
10. Environmental N, Foundation E. National Environmental Education Foundation
6
Environmental Management of Pediatric Asthma Guidelines for Health Care Providers.
https://www.niehs.nih.gov/health/materials/environmental_management_of_pediatric_ast
hma_guidelines_for_health_care_providers_508.pdf.
11. Spann K, Snape N, Baturcam E, Fantino E. The Impact of Early-Life Exposure to Air-
borne Environmental Insults on the Function of the Airway Epithelium in Asthma. Ann
Glob Heal. 2016. doi:10.1016/j.aogh.2016.01.007
12. Bernstein JA, Alexis N, Barnes C, et al. Health effects of air pollution. J Allergy Clin
Immunol. 2004. doi:10.1016/j.jaci.2004.08.030
13. McConnell R, Islam T, Shankardass K, et al. Childhood incident asthma and traffic-
related air pollution at home and school. Environ Health Perspect. 2010.
doi:10.1289/ehp.0901232
14. Clark NA, Demers PA, Karr CJ, et al. Effect of early life exposure to air pollution on
development of childhood asthma. Environ Health Perspect. 2010.
doi:10.1289/ehp.0900916
15. Guevara M. Emissions of primary particulate matter. Issues Environ Sci Technol. 2016.
doi:10.1039/9781782626589-00001
16. Delfino RJ, Quintana PJE, Floro J, et al. Association of FEV1 in asthmatic children with
personal and microenvironmental exposure to airborne particulate matter. Environ Health
Perspect. 2004. doi:10.1289/ehp.6815
17. Habre R, Moshier E, Castro W, et al. The effects of PM2.5 and its components from
indoor and outdoor sources on cough and wheeze symptoms in asthmatic children. J Expo
Sci Environ Epidemiol. 2014. doi:10.1038/jes.2014.21
18. Williams AM, Phaneuf DJ, Barrett MA, Su JG. Short-term impact of PM 2.5 on
7
contemporaneous asthma medication use: Behavior and the value of pollution reductions.
Proc Natl Acad Sci U S A. 2019. doi:10.1073/pnas.1805647115
19. Chung KF, Wenzel SE, Brozek JL, et al. International ERS/ATS guidelines on definition,
evaluation and treatment of severe asthma. Eur Respir J. 2014.
doi:10.1183/09031936.00202013
20. Skloot GS. Asthma phenotypes and endotypes: A personalized approach to treatment.
Curr Opin Pulm Med. 2016. doi:10.1097/MCP.0000000000000225
21. Zou B, Wilson JG, Zhan FB, Zeng Y. Air pollution exposure assessment methods utilized
in epidemiological studies. J Environ Monit. 2009. doi:10.1039/b813889c
22. Larkin A, Hystad P. Towards Personal Exposures: How Technology Is Changing Air
Pollution and Health Research. Curr Environ Heal reports. 2017. doi:10.1007/s40572-
017-0163-y
23. Brauer M. How much, how long, what, and where: Air pollution exposure assessment for
epidemiologic studies of respiratory disease. In: Proceedings of the American Thoracic
Society. ; 2010. doi:10.1513/pats.200908-093RM
24. Georgopoulos PG, Lioy PJ. Conceptual and theoretical aspects of human exposure and
dose assessment. J Expo Anal Environ Epidemiol. 1994.
25. Dockery DW, Spengler JD. Personal exposure to respirable particulates and sulfates. J Air
Pollut Control Assoc. 1981. doi:10.1080/00022470.1981.10465205
26. Duan N. Models for human exposure to air pollution. Environ Int. 1982.
doi:10.1016/0160-4120(82)90041-1
27. Ott WR. Concepts of human exposure to air pollution. Environ Int. 1982.
doi:10.1016/0160-4120(82)90104-0
8
28. Purcell K. Half of adult cell phone owners have apps on their phones. 2011:1-33.
http://pewinternet.org/Reports/2011/Apps-update.aspx.
29. Preimesberger C. Ericsson Mobility Report Predicts 500M 5G Devices by 2022. eWeek.
2016.
30. Anagnostopoulos T, Garcia JC, Goncalves J, Ferreira D, Hosio S, Kostakos V.
Environmental exposure assessment using indoor/outdoor detection on smartphones. Pers
Ubiquitous Comput. 2017;21(4):761-773. doi:10.1007/s00779-017-1028-y

9
2 CHAPTER 2: BACKGROUND LITERATURE
In this chapter, we review the literature on pediatric asthma in relation to environmental
exposures – specifically air pollution and particulate matter (PM) – with a special emphasis on
studies of acute effects, existing air pollution exposure assessment methods and their strengths
and weaknesses, and new advances or developments in sensors and mobile Health (mHealth)
methods that are paving the way for more accurate and personalized exposure assessment and
health investigations.
2.1 PEDIATRIC ASTHMA
Asthma is a chronic disorder of both adults and children, which afflicts more than 25
million Americans accounting for 7.7 percent of adults and 8.4 percent of children
1
. Asthma
prevalence has been increasing over the last few decades in all age, sex, and racial groups, but
especially in children
2
. Asthma is characterized by increased inflammation in the lower
respiratory tract, bronchial constriction and hyperreactive airways that are sensitive to a wide
variety of triggers
3
. Current recommendations are tailored or targeted towards establishing and
maintaining asthma control, in order to reduce the chances of exacerbations and improve quality
of life (QoL)
4
.
The American Public Health Association recognizes asthma as one of the most prevalent
chronic diseases among children in the United States
5
. Children with asthma are significantly
burdened by asthma symptoms and account for higher rates of emergency department visits,
hospitalization, and even mortality
6
. Asthma control is positively associated with improved
quality of life. Uncontrolled asthma is related to worse performance metrics in physical exercise,
and reduced lung function increased the frequency of emergency department visits, school
absenteeism, and has a high financial cost to families and society
7
.
10
Asthma can be very disruptive to daily life and can lead to worsened symptoms that may
inhibit physical activity, lead to missed school days, and cause poor sleep in children
8
. Common
pediatric asthma symptoms include coughing, wheezing, dyspnea, and chest tightness, with
nighttime awakening and early morning symptoms being more prevalent
9
.
2.2 THE BURDEN OF PEDIATRIC ASTHMA
Asthma in the general population imposes a significant burden on morbidity, healthcare
costs, and quality of life with rising prevalence around the world with the number of new asthma
cases in the United States since 1980 rising 75% among all patients and 160% among children 0-
5 years
10
. There is a higher prevalence of pediatric asthma among children aged 11–17 years,
but the youngest children aged 0-4 years of age actually had the highest rates of asthma-related
health care
11
.
Patients with mild intermittent asthma used their reliever and control medications less
often, were less likely to be hospitalized or have urgent care visits, missed fewer days of school,
and were more satisfied with their quality of life compared to those with mild persistent disease,
who, in turn, were overall less burdened than those with moderate/severe persistent disease
10
.
When looking at the long-term trends in pediatric asthma burden in the United States from 1980-
2007, there is an overall improving trend in children with asthma aged 0 to 17 years with regards
to the rate of emergency visits, and an improvement starting from the 1990s in the rate of
hospitalizations and mortality
11
.
Asthma symptoms in all patients can include cough, wheezing, shortness of breath, and
chest tightness. In addition, pediatric asthma symptoms often occur at night, and about 40% of
asthmatic children present with nocturnal awakening due to symptoms
12
. Uncontrolled
symptoms can lead to increased urgent visits to doctors and more emergency department visits
11
which can occasionally lead to hospitalization and the need for mechanical ventilation. These
symptoms often negatively affect quality of life which can result in a myriad of consequences.
decreasing limits in a child’s daily activities, decreased school attendance, and worse academic
performance which can further lead to diminished maximal educational achievement, potentially
lower socioeconomic status and wider aggregate societal effects when accounting for the total
prevalence of pediatric asthma
11
.
Pediatric asthma not only affects the quality of life (QoL) in asthmatic children but also
decreases the QoL in their parents. A Dutch study showed that parents of chronically ill children
had a significantly lower health related QoL. The largest subgroup of their study was the parents
of children with asthma The study found that this parental subgroup had lower QoL related to
sleep, social functioning, daily activities, vitality, positive emotions, and depressive symptoms
13
.
2.3 ASTHMA CONTROL
2.3.1 Definition of Pediatric Asthma Control
A large proportion of the total cost of asthma is attributed to treating the consequences of
poor asthma control
14
. Therefore, improving asthma control is an important aspect of the
evaluation and management of the disease. Control of asthma includes control of the current
impairment, including daily/nocturnal symptoms, reliever use, level of activity, and quality of
life, and control of the future risks, including the risk of exacerbations, permanent impairment of
lung function, and the adverse effects of treatment. According to the National Asthma Education
and Prevention Program (NAEPP) Expert Panel Report 3 (EPR-3)
15
and the 2016 Global
Initiative for Asthma (GINA) guidelines
4
, achieving asthma control becomes the goal after
treatment is established. According to the GINA guidelines, controlled asthma in children ages
12
six and older is assessed by twice or less times daytime symptoms per week, none limitation of
actives, none nocturnal symptoms/awakenings, and twice or less times of need for reliever/rescue
inhaler use per week, and greater than 80% predicted or personal best lung function (peak
expiratory flow rate (PEFR) or forced expiratory volume in one second (FEV1)). Partly
controlled asthma is defined as any of the following measure presented, more than twice daytime
symptoms per week, any limitation of actives, any nocturnal symptoms/awakenings, and more
than twice times of need for reliever/rescue inhaler use per week, and less than 80% predicted or
personal best lung function (PEFR or FEV1). Uncontrolled asthma is defined as three or more
features of the partly controlled asthma.
2.3.2 Risk Factors for Worsening Pediatric Asthma Control
The mechanisms leading to poor control of asthma in children are complex and previous
studies identified several casual risk factors for worsening pediatric asthma control with a list of
examples as follows:
1) Race or ethnicity and socioeconomic status, where children who are African American, Puerto
Rican, or come from lower-income families, or poor insurance status are at higher risk of an
asthma attack in the United States
16,17
.
2) Genetics, with one susceptibility gene identified to be associated with recurrent asthma
exacerbation in children ages 2-6 years
18
.
3) Being overweight or obese has been associated with poor asthma control in children by many
studies
19,20
For example, a national study recruited 28,645 children with asthma showed that
obese children had 15% higher odds of developing an asthma exacerbation compared to normal-
weight children
21
.
13
2.3.3 Asthma Phenotypes
Another challenge in achieving asthma control in children is related to the heterogeneity
of the disease, where asthma symptoms manifest differently across different individuals and
response to treatment varies
22
. There are several different phenotypes each with unique factors
involved in exacerbation. For example, allergic asthma is associated with atopic conditions like
allergic rhinitis, eczema and increased immunoglobulin E levels that often begins in childhood
and features eosinophilic inflammation. Aspirin-exacerbated asthma is associated with nasal
polyps and is induced when taking non-steroidal anti-inflammatory drugs like aspirin and
ibuprofen. There is a specific eosinophilic type of asthma classified by elevated presence of
eosinophils in the sputum that responds well to corticosteroid therapy
23
. In addition, there is also
a non-eosinophilic asthma associated with obesity that often improves with weight loss
24
.
2.3.4 Triggers for Acute Pediatric Asthma Exacerbation
Preventing asthma exacerbation in children requires identification of risks on the
individual level so that personalized care protocols can be developed to prevent such
exacerbations
25
. Previous studies identified several triggers for asthma exacerbations in children:
1) Exercise is a common trigger of bronchial hyperresponsiveness and may cause cough, wheezing,
shortness of breath and chest tightness
26
. The sign of exercise-induced bronchial
hyperresponsiveness may begin during or soon after exercise, and if untreated, these symptoms
may last for 60 minutes or longer
27
.
2) Viral infections, which are related to almost 85% of severe asthma exacerbations in children most
commonly due to human rhinovirus and can activate neutrophils and eosinophils, increase airway
inflammation
2829
;
3) Second-hand smoking, or Environmental Tobacco Smoke (ETS), shown to be associated with
worsening lung function and emergency room visits
30
;
14
4) Exposure to outdoor air pollution such as particulate matter and ozone, including traffic-related
air pollution
31

32

33
;
5) Allergens and indoor air pollutants, such as dust mites, animal allergens, molds, pollens, tobacco
smoke (or second-hand smoke), chemical, and combustion by-products can all lead to poor
asthma control and asthma exacerbation in sensitized children
3435
;
2.4 PARTICULATE MATTER AIR POLLUTION EXPOSURE AS A RISK FACTOR FOR
ACUTE ASTHMA EXACERBATION IN CHILDREN
Exposure to both outdoor and indoor air pollutants has been associated with increased
asthma exacerbations, rates of hospitalization, and reduced lung function
31,32
. Particulate matter
(PM) of aerodynamic diameter < 2.5 m (PM2.5) consists of a combination of small particles and
liquid droplets from vehicular exhaust, coal and wood-burning and industrial activities
36
. Short-
term air pollution exposure increases the risk of asthma exacerbation among children
37
. A case-
crossover study from China suggested that weekly ambient exposure to PM2.5 was significantly
associated with pediatric outpatient department visits
38
. Another case-crossover study conducted
in Massachusetts with 33,387 children found those with low birthweight had increased odds of
having an asthma clinical visit due to higher ambient exposure of one day before PM2.5
39
.
Emergency department visits are not the only outcome of pediatric asthma exacerbation. A
recently published paper by using an asthma digital health platform on national scale concluded a
1 μg/m
3
(12%) increase in weekly exposure to PM2.5 increases weekly inhaler use by 0.82%
40
.
Averaged 24 hour personal continuous PM2.5 exposure was associated with reduced percent
predicted FEV1
41
, and daily indoor PM2.5 exposure was associated with cough and wheeze
symptoms
42
. Many of the above studies were using ambient PM2.5 level to represent personal
PM2.5 exposure which may introduce measurement error. Measurement error is an unavoidable
15
issue in air pollution epidemiology studies
43
. In general, exposure measurement error tends to
distort the sensitivity of epidemiologic studies to discover the true effects of air pollution on
public health
44
. For example, misclassification is one of the major types of measurement error in
air pollution epidemiology studies, and it usually biases a relative risk, regression coefficient, or
other effective measures towards the null value (no association)
45

46
. Therefore, the gold
standard for assessing air pollution exposure in human health studies is personal monitoring
47,48
.
However, collecting personal air pollution measures for large populations remains a significant
challenge due to feasibility, cost and burden on participants
48
, so better approach is suggested to
improve the personal air pollution exposure assessment.
2.5 PERSONAL AIR POLLUTION EXPOSURE ASSESSMENT
2.5.1 Personal Air Pollution Exposure Assessment Methods
Approaches for personal air pollution exposure assessment are usually divided into two
categories: direct methods and indirect methods
49
. Direct assessment methods include studies
where an individual is followed by an observer
50
, and measurements are taken on the individual
including biomarkers and personal monitoring. Indirect assessment methods include modeling
approaches where data from ambient monitors, microenvironmental modeling approach, surveys,
and mathematical models are used
53
.
Duan
54
introduced the concept that personal exposure can be calculated as an
accumulating process that is related to real-time pollutant concentration and duration of the
exposure as shown in Equation 1:
𝐸 = ∫ 𝐶 (𝑡 )𝑑𝑡
𝑡 2
𝑡 1
(1)
16
Where E is the total personal exposure, C(t) is the pollutant concentration level at time t, and dt
is the time duration that is associated with C(t).
For direct methods, the measurements are measured based on each subject individual
movement trajectory and pollution exposure concentration. There are several advantages for
direct methods. First, because the method measures personal exposure continuously for each
subject, the spatial resolution can be accurate to meter and the temporal resolution to second
depending on the logging frequency
55
. Finally, recent development in mobile positioning
technology, such as Global Positioning System (GPS), presents significant advances for
individual-based exposure assessment. However, individual-based assessment methods also have
several limitations, such as limited participation rates due to burden, high costs, and
inconvenience to use the equipment
56
.
For indirect methods, one fixed value is assigned to represent the air pollution
concentration for each region at a specific time
57
. However, the disadvantages of indirect
assessment methods should not be neglected. First, the assumption of homogeneous air pollution
distribution in a region does not capture the true spatial variability of most air pollutants. Also,
the assumption of different subjects in the same region have the same air pollution concentration
is incorrect, which would be an ecological fallacy.
2.5.2 Microenvironmental Modeling Approach
Indirect assessment methods such as ambient measurements from the monitor and
mathematical models fail to consider important differences between personal exposure and
ambient levels
46
. The microenvironmental modeling approach was developed to help address
this issue in estimating total personal exposure
58
. According to the microenvironmental
framework, a person's daily activities are related to a series of microenvironments, such as a
17
home, workplace, traveling route, and recreation location, and integrating exposure by
combining measurements of pollutant concentrations within each major microenvironment with
information about the duration of time people spend in them (through logs or diaries) can
provide an accurate estimate of total personal exposure
49,54,59
. Concentrations of air pollutants
are is assumed to be homogeneous in each microenvironment. Although human activities are
very complicated, the microenvironment approach assumes that daily activities follow a
somewhat repeatable routine such as commuting to work, staying at the workplace, commuting
back home, and staying home and can be simplified or approximated by capturing some of the
major microenvironments
60
. According to Sexton et al
61
, total personal exposure is calculated as
a time-weighted integration of average according to the following equation (1):
𝐸 𝑖𝑗
= ∑ 𝐶 𝑗 𝑡 𝑖𝑗
𝐽 𝑗 (1)
where 𝐸 𝑖𝑗
is the time-weighted integrated exposure for a person 𝑖 over the specific time period,
𝐶 𝑗 is the air pollution concentration in the 𝑗 th microenvironment, 𝑡 𝑖𝑗
is the aggregate time that
person i spends in microenvironment j; and J is the total number of microenvironments that
person i moves through during the specified time period.
Previous studies mainly relied on self-reported questionnaires or surveys to obtain the
microenvironment frequency distributions for subjects (Klepeis, 1999; N. Armstrong &
Welsman, 2006; Cho, Rodríguez, & Evenson, 2011; Lubans, Boreham, Kelly, & Foster, 2011).
However, reports by questionnaires yield recall bias and may be inconsistent because users may
forget to report. Therefore, more convenient and accurate method is advised to estimate the 𝑡 𝑖𝑗
in
the above equation (1).
18
2.6 MOBILE HEALTH (MHEALTH) AND SENSOR-BASED APPROACHES FOR
PERSONALIZED EXPOSURE AND HEALTH ASSESSMENT
2.6.1 Three Domains Framework
The ultimate goal of air pollution exposure assessment is to minimize errors in estimating
exposures for large-scale populations. Larkin and Hystad
47
presented a three-domain framework
to illustrate how available technologies and advances in data science can contribute to this goal.
se are the three important domains that contribute to advances in personal air pollution exposure
assessment. Each domain can provide unique information to personal exposure assessment, but
their interaction can provide novel exposure information that cannot be obtained from any single
domain. For example, single air pollution models cannot capture personal exposures without
detailed time-activity pattern information. Similarly, smartphone GPS-based models cannot
provide actual measurements of air pollution concentration. Even if we have personal
measurements, this data requires integration with both air pollution models and time-activity
patterns to build generalizable and accurate personal exposure models. This leads to a conclusion
that a strategy combines the above domains into a properly integrated framework, which will be
discussed in later chapters.
2.6.2 Ubiquitous Availability and Use of Smartphones
Recent advances in smartphone technology provide an opportunity to improve
geolocation and time-activity data in microenvironments to inform personal air pollution
exposure assessment on large-scale population-based research. About sixty-eight percent of the
U.S. population used a smartphone in 2017, and twenty-night percent of them have downloaded
an application related to health management or tracks
64
. More smartphone users start to use
sensors to help them measure and improve their health and behavior
65
. Smartphones and
19
mHealth (mobile health) are providing numerous possibilities for improving air pollution
exposure assessment and epidemiology studies.
Evolutions in smartphone and sensing technologies are revolutionizing the way health
studies are being conducted. Environmental sensors embedded in smartphones can enhance air
pollution exposure assessment efforts by enabling researchers to collect time-activity data in
important microenvironments with much higher accuracy and lower participant burden than
previously possible using traditional methods. Smartphones and mobile application advances are
also allowing researchers to conduct mHealth studies using custom apps that can serve as the
primary platform for conducting air pollution health studies. The Asthma Mobile Health Study
(AMHS) (https://apps.icahn.mssm.edu/asthma/) was designed based on this concept, where the
recruitment, consent, and enrollment were all completed remotely by smartphone. The
application was downloaded nearly 50,000 times over the first six months. This study is support
for conducting new research entirely through a smartphone platform, and it suggested the
association between asthma symptoms to changes in heat, pollen, and air pollution
66
.
Related to this work, the Los Angeles PRISMS Breathe Kit (Biomedical Real-Time
Health Evaluation Kit) was developed as a non-invasive, secure informatics platform for
exposure studies of pediatric asthma. It includes real time, personal air pollution, geolocation,
inhaler medication and physiological health sensor and applications to collect self-report data
using Ecological Momentary Assessment (EMA). This thesis will utilize data collected through
this informatics platform deployed in a repeated measures panel study to investigate the
relationship between acute, personal air pollution exposure and pediatric asthma exacerbation.
20
2.7 CONCLUSION
In this chapter, we discussed the key features of pediatric asthma and the epidemiology
evidence of air pollution’s contribution to the current pediatric asthma epidemic. We provided
evidence that air pollution exposure is an important trigger or risk factor for exacerbating
pediatric asthma, especially on an acute basis in children. However, evidence on how within-day
air pollution exposure change contributes to acute worsening of asthma is limited. Therefore,
further research is needed to understand the complexity of short-term air pollution change on day
level and within-day level and asthma exacerbation in children.
We also discussed the personal exposure assessment methods and advances in sensor
technology to improve personal exposure assessment. We also provided an overview of how
advances in modern data science technologies offer a new approach for improving personal
exposure assessment.

21
2.8 REFERENCES
1. Centers for Disease Control and Prevention. CDC - Asthma - Most Recent Asthma Data.
Most Recent Asthma Data.
2. Kuschner WG. The asthma epidemic [12]. N Engl J Med. 2007.
doi:10.1056/NEJMc063596
3. Murdoch JR, Lloyd CM. Chronic inflammation and asthma. Mutat Res - Fundam Mol
Mech Mutagen. 2010. doi:10.1016/j.mrfmmm.2009.09.005
4. GINA C científico de l. Global Strategy for Asthma Management and Prevention ( 2016
update ). Guía Gina Actual. 2016.
5. Zahran HS, Bailey CM, Damon SA, Garbe PL, Breysse PN. Vital signs: Asthma in
children — United States, 2001-2016. Morb Mortal Wkly Rep. 2018.
doi:10.15585/mmwr.mm6705e1
6. Akinbami L. The state of childhood asthma, United States, 1980-2005. Adv Data. 2006.
7. Lang A, Mowinckel P, Sachs-Olsen C, et al. Asthma severity in childhood, untangling
clinical phenotypes. Pediatr Allergy Immunol. 2010. doi:10.1111/j.1399-
3038.2010.01072.x
8. Sennhauser FH, Braun-Fahrländer C, Wildhaber JH. The burden of asthma in children: A
European perspective. Paediatr Respir Rev. 2005. doi:10.1016/j.prrv.2004.11.001
9. van Aalderen WMC, Meijer GG, Oosterhoff Y, Bron AO. Epidemiology and the concept
of underlying mechanisms of nocturnal asthma. Respir Med. 1993. doi:10.1016/S0954-
6111(06)80352-5
10. Fuhlbrigge AL, Adams RJ, Guilbert TW, et al. The burden of asthma in the United States:
Level and distribution are dependent on interpretation of the National Asthma Education
22
and Prevention Program guidelines. Am J Respir Crit Care Med. 2002.
doi:10.1164/rccm.2107057
11. Akinbami LJ, Moorman JE, Garbe PL, Sondik EJ. Status of childhood asthma in the
United States, 1980-2007. Pediatrics. 2009. doi:10.1542/peds.2008-2233C
12. Meijer GG, Postma DS, Wempe JB, Gerritsen J, Knol K, Van Aalderen WMC. Frequency
of nocturnal symptoms in asthmatic children attending a hospital out-patient clinic. Eur
Respir J. 1995. doi:10.1183/09031936.95.08122076
13. Hatzmann J, Heymans HSA, Ferrer-i-Carbonell A, Van Praag BMS, Grootenhuis MA.
Hidden consequences of success in pediatrics: Parental health-related quality of life-
results from the care project. Pediatrics. 2008. doi:10.1542/peds.2008-0582
14. Barnes PJ, Jonsson B, Klim JB. The costs of asthma. Eur Respir J. 1996.
doi:10.1183/09031936.96.09040636
15. National Asthma Education and Prevention Program. Expert Panel Report 3 (EPR-3):
Guidelines for the Diagnosis and Management of Asthma-Summary Report 2007. J
Allergy Clin Immunol. 2007.
16. Akinbami LJ, Moorman JE, Liu X. Asthma prevalence, health care use, and mortality:
United States, 2005-2009. Natl Health Stat Report. 2011.
17. Ferrante G, La Grutta S. Reasons for inadequate asthma control in children: An important
contribution from the “French 6 cities study.” Multidiscip Respir Med. 2012.
doi:10.1186/2049-6958-7-23
18. Bønnelykke K, Sleiman P, Nielsen K, et al. A genome-wide association study identifies
CDHR3 as a susceptibility locus for early childhood asthma with severe exacerbations.
Nat Genet. 2014. doi:10.1038/ng.2830
23
19. Borrell LN, Nguyen EA, Roth LA, et al. Childhood obesity and asthma control in the
GALA II and SAGE II studies. Am J Respir Crit Care Med. 2013.
doi:10.1164/rccm.201211-2116OC
20. Forno E, Lescher R, Strunk R, Weiss S, Fuhlbrigge A, Celedón JC. Decreased response to
inhaled steroids in overweight and obese asthmatic children. J Allergy Clin Immunol.
2011. doi:10.1016/j.jaci.2010.12.010
21. Black MH, Zhou H, Takayanagi M, Jacobsen SJ, Koebnick C. Increased asthma risk and
asthma-related health care complications associated with childhood obesity. Am J
Epidemiol. 2013. doi:10.1093/aje/kwt093
22. Pembrey L, Barreto ML, Douwes J, et al. Understanding asthma phenotypes: the World
Asthma Phenotypes (WASP) international collaboration. ERJ Open Res. 2018.
doi:10.1183/23120541.00013-2018
23. Asthma Self-Management Support | Severe Asthma Toolkit.
https://toolkit.severeasthma.org.au/living-severe-asthma/self-management-support/.
24. Hirose M, Horiguchi T. Asthma phenotypes. 2017;(June 2015):189-194.
doi:10.1002/jgf2.7
25. Puranik S, Forno E, Bush A, Celedón JC. Predicting severe asthma exacerbations in
children. Am J Respir Crit Care Med. 2017. doi:10.1164/rccm.201606-1213PP
26. Randolph C. Exercise-induced bronchospasm in children. Clin Rev Allergy Immunol.
2008. doi:10.1007/s12016-007-8035-0
27. www.mayoclinic.org. Exercise-induced asthma - Symptoms and causes - Mayo Clinic.
2014. https://www.mayoclinic.org/diseases-conditions/exercise-induced-
asthma/symptoms-causes/syc-20372300.
24
28. Saraya T, Kurai D, Ishii H, et al. Epidemiology of virus-induced asthma exacerbations:
With special reference to the role of human rhinovirus. Front Microbiol. 2014.
doi:10.3389/fmicb.2014.00226
29. Matsumoto K, Inoue H. Viral infections in asthma and COPD. Respir Investig. 2014.
doi:10.1016/j.resinv.2013.08.005
30. Vargas PA, Brenner B, Clark S, Boudreaux ED, Camargo CA. Exposure to environmental
tobacco smoke among children presenting to the emergency department with acute
asthma: A multicenter study. Pediatr Pulmonol. 2007. doi:10.1002/ppul.20637
31. Tatum AJ, Shapiro GG. The effects of outdoor air pollution and tobacco smoke on
asthma. Immunol Allergy Clin North Am. 2005. doi:10.1016/j.iac.2004.09.003
32. Gehring U, Wijga AH, Brauer M, et al. Traffic-related air pollution and the development
of asthma and allergies during the first 8 years of life. Am J Respir Crit Care Med. 2010.
doi:10.1164/rccm.200906-0858OC
33. Delfino R, Wu J, Tjoa T, Gullesserian S, Nickerson B, Gillen D. Asthma Morbidity And
Ambient Air Pollution: Effect Modification By Residential Traffic-Related Air Pollution.
In: ; 2012. doi:10.1164/ajrccm-conference.2012.185.1_meetingabstracts.a2478
34. Kanchongkittiphon W, Mendell MJ, Gaffin JM, Wang G, Phipatanakul W. Indoor
environmental exposures and exacerbation of asthma: An update to the 2000 review by
the institute of medicine. Environ Health Perspect. 2015. doi:10.1289/ehp.1307922
35. Environmental N, Foundation E. National Environmental Education Foundation
Environmental Management of Pediatric Asthma Guidelines for Health Care Providers.
https://www.niehs.nih.gov/health/materials/environmental_management_of_pediatric_ast
hma_guidelines_for_health_care_providers_508.pdf.
25
36. Guevara M. Emissions of primary particulate matter. Issues Environ Sci Technol. 2016.
doi:10.1039/9781782626589-00001
37. Orellano P, Quaranta N, Reynoso J, Balbi B, Vasquez J. Effect of outdoor air pollution on
asthma exacerbations in children and adults: Systematic review and multilevel meta-
analysis. PLoS One. 2017. doi:10.1371/journal.pone.0174050
38. Wu J, Zhong T, Zhu Y, Ge D, Lin X, Li Q. Effects of particulate matter (PM) on
childhood asthma exacerbation and control in Xiamen, China. BMC Pediatr. 2019.
doi:10.1186/s12887-019-1530-7
39. Khalili R, Bartell SM, Hu X, et al. Early-life exposure to PM 2.5 and risk of acute asthma
clinical encounters among children in Massachusetts: A case-crossover analysis. Environ
Heal A Glob Access Sci Source. 2018. doi:10.1186/s12940-018-0361-6
40. Williams AM, Phaneuf DJ, Barrett MA, Su JG. Short-term impact of PM 2.5 on
contemporaneous asthma medication use: Behavior and the value of pollution reductions.
Proc Natl Acad Sci U S A. 2019. doi:10.1073/pnas.1805647115
41. Delfino RJ, Quintana PJE, Floro J, et al. Association of FEV1 in asthmatic children with
personal and microenvironmental exposure to airborne particulate matter. Environ Health
Perspect. 2004. doi:10.1289/ehp.6815
42. Habre R, Moshier E, Castro W, et al. The effects of PM2.5 and its components from
indoor and outdoor sources on cough and wheeze symptoms in asthmatic children. J Expo
Sci Environ Epidemiol. 2014. doi:10.1038/jes.2014.21
43. Sarnat JA, Wilson WE, Strand M, Brook J, Wyzga R, Lumley T. Panel discussion review:
Session 1 — exposure assessment and related errors in air pollution epidemiologic studies.
J Expo Sci Environ Epidemiol. 2007. doi:10.1038/sj.jes.7500621
26
44. Zeger SL, Thomas D, Dominici F, et al. Exposure measurement error in time-series
studies of air pollution: Concepts and consequences. Environ Health Perspect. 2000.
doi:10.1289/ehp.00108419
45. Armstrong BG. Effect of measurement error on epidemiological studies of environmental
and occupational exposures. Occup Environ Med. 1998. doi:10.1136/oem.55.10.651
46. Goldman GT, Mulholland JA, Russell AG, et al. Impact of exposure measurement error in
air pollution epidemiology: Effect of error type in time-series studies. Environ Heal A
Glob Access Sci Source. 2011. doi:10.1186/1476-069X-10-61
47. Larkin A, Hystad P. Towards Personal Exposures: How Technology Is Changing Air
Pollution and Health Research. Curr Environ Heal reports. 2017. doi:10.1007/s40572-
017-0163-y
48. Brauer M. How much, how long, what, and where: Air pollution exposure assessment for
epidemiologic studies of respiratory disease. In: Proceedings of the American Thoracic
Society. ; 2010. doi:10.1513/pats.200908-093RM
49. Ott WR. Concepts of human exposure to air pollution. Environ Int. 1982.
doi:10.1016/0160-4120(82)90104-0
50. Setton EM, Allen R, Hystad P, Keller CP. Outdoor Air Pollution and Health – A Review
of the Contributions of Geotechnologies to Exposure Assessment. In: Geospatial Analysis
of Environmental Health. ; 2011. doi:10.1007/978-94-007-0329-2_3
51. Clougherty JE, Levy JI, Kubzansky LD, et al. Synergistic effects of traffic-related air
pollution and exposure to violence on urban asthma etiology. Environ Health Perspect.
2007. doi:10.1289/ehp.9863
52. Næss Ø, Nafstad P, Aamodt G, Claussen B, Rosland P. Relation between concentration of
27
air pollution and cause-specific mortality: Four-year exposures to nitrogen dioxide and
particulate matter pollutants in 470 neighborhoods in Oslo, Norway. Am J Epidemiol.
2007. doi:10.1093/aje/kwk016
53. Monn C. Chapter 4 Exposure assessment of air pollutants: a review on spatial
heterogeneity and indoor/outdoor/personal exposure to suspended particulate matter,
nitrogen dioxide and ozone. Dev Environ Sci. 2002. doi:10.1016/S1474-8177(02)80007-9
54. Duan N. Models for human exposure to air pollution. Environ Int. 1982.
doi:10.1016/0160-4120(82)90041-1
55. Nethery E, Leckie SE, Teschke K, Brauer M. From measures to models: An evaluation of
air pollution exposure assessment for epidemiological studies of pregnant women. Occup
Environ Med. 2008. doi:10.1136/oem.2007.035337
56. Gerharz LE, Krüger A, Klemm O. Applying indoor and outdoor modeling techniques to
estimate individual exposure to PM2.5 from personal GPS profiles and diaries: A pilot
study. Sci Total Environ. 2009. doi:10.1016/j.scitotenv.2009.06.006
57. Zou B, Wilson JG, Zhan FB, Zeng Y. Air pollution exposure assessment methods utilized
in epidemiological studies. J Environ Monit. 2009. doi:10.1039/b813889c
58. Georgopoulos PG, Lioy PJ. Conceptual and theoretical aspects of human exposure and
dose assessment. J Expo Anal Environ Epidemiol. 1994.
59. Dockery DW, Spengler JD. Personal exposure to respirable particulates and sulfates. J Air
Pollut Control Assoc. 1981. doi:10.1080/00022470.1981.10465205
60. Srivastava L. Mobile phones and the evolution of social behaviour. Behav Inf Technol.
2005. doi:10.1080/01449290512331321910
61. Sexton K. Air Pollution, the Automobile, and Public Health. In: Air Pollution, the
28
Automobile, and Public Health. ; 1988:208-238.
62. Klepeis NE. An introduction to the indirect exposure assessment approach: Modeling
human exposure using microenvironmental measurements and the recent National Human
Activity Pattern Survey. In: Environmental Health Perspectives. ; 1999.
doi:10.2307/3434429
63. Gonzales M, Ngo L, Hammond SK, Tager I. Validation of a questionnaire and
microenvironmental model for estimating past exposures to ozone. Int J Environ Health
Res. 2003. doi:10.1080/0960312031000122406
64. Purcell K. Half of adult cell phone owners have apps on their phones. 2011:1-33.
http://pewinternet.org/Reports/2011/Apps-update.aspx.
65. Swan M. Health 2050: The realization of personalized medicine through crowdsourcing,
the quantified self, and the participatory biocitizen. J Pers Med. 2012;2(3):93-118.
doi:10.3390/jpm2030093
66. Chan YFY, Wang P, Rogers L, et al. The Asthma Mobile Health Study, a large-scale
clinical observational study using ResearchKit. Nat Biotechnol. 2017.
doi:10.1038/nbt.3826

29
3 CHAPTER 3: MICROENVIRONMENTAL TIME-ACTIVITY
CLASSIFICATION USING SMARTPHONE SENSOR DATA FOR PERSONAL
AIR POLLUTION EXPOSURE ASSESSMENT
3.1 ABSTRACT
Accurate assessment of personal exposures is critical for air pollution studies. Recent
advances in mobile sensing technologies can improve microenvironmental classification at high
resolution.
Two participants (S1 and S2) used the ‘AndroSensor' application to log 21,296 and 22,340
observations on 31 parameters from Samsung S4 smartphone sensors every 10 seconds for ~5 days.
We trained 5 machine learning algorithms to classify microenvironments as indoor, outdoor and
in-transit. We used conventional and targeted cross-validation (CV) accounting for temporal
and/or spatial autocorrelation to evaluate model prediction performance in S1 (main) and S2
(independent test).
Random forest had the highest balanced accuracy (BA) of 95.1% in S1 and 57.5% in S2.
The top five features were sound level, speed, temperature, location accuracy, and relative
humidity. BA dropped after considering temporal, spatial, and spatiotemporal autocorrelation
(63.2%, 60.9%, and 51.4%, respectively). Model performance in S2 was poor under natural living
conditions (50.6% BA) but improved (70.0%BA) under scripted activities.
We built a model to predict 10-second level microenvironment from smartphone sensor data.
We demonstrated the importance of accounting for autocorrelation in high resolution data to when
assessing prediction performance and the potential to generalize models to groups sharing
somewhat similar time-activity patterns.

30
3.2 INTRODUCTION
Exposure to particulate and gaseous air pollution has been associated with acute and chronic
morbidities and mortality
1–7
. While our understanding of the association between air pollution
exposure and health outcomes has improved, challenges remain in accurately assessing personal
air pollution exposures in epidemiological studies. Many studies rely on outdoor air pollution to
estimate total personal exposures without accounting for time-activity patterns (e.g., time spent
indoors or in transit) which introduces exposure measurement error. According to the National
Human Activity Pattern Survey (NHAPS), respondents spend an average of 87% of their time in
enclosed buildings and about 6% of their time in enclosed vehicles
8
.
Since the gold standard approach of personal monitoring is not feasible to apply in large
populations, the microenvironmental model to estimate personal exposure has been developed to
address some of these challenges
9
. According to Sexton et al.,
10
total personal exposure can be
estimated as a time-weighted integration of average concentrations encountered in major
microenvironments, defined as areas where air pollution concentrations are assumed to be
homogenous
11
, according to equation (1) :
𝐸 = ∑ 𝐶 𝑗 𝑡 𝑗 𝐽 𝑗 (1)
where E is the time-weighted integrated exposure for a person over the specific time period, 𝐶 𝑗 is
the air pollution concentration in the 𝑗 th microenvironment, 𝑡 𝑗 is the fraction of time the person
spends in microenvironment j; and J is the total number of major microenvironments encountered
during the specified period.
Many previous studies relied on time-activity diaries to collect detailed or aggregate
information about time spent in major microenvironments
12
. However, self-report methods can
be burdensome to complete and are prone to recall error. Mobile health (mHealth) technologies
31
may provide smart alternatives as smartphone use is increasingly ubiquitous and people tend to
carry their smartphones daily
13
. Smartphone sensors and embedded global positioning systems
(GPS) may provide an opportunity to develop automated smartphone methods to continuously
classify time-activity in key microenvironments with low participant burden.
Data collected over time and space are not independent and identical distributed.
Autocorrelation is a measurement of similarity (correlation) between nearby observations. Two
observations that are close to each other in time are more similar and tend to have higher temporal
autocorrelation. Two observations that are close at spatial locations are similar to each other and
tend to have high spatial autocorrelation. Therefore, for any data that are not temporal and spatial
independent, we need to consider the temporal and spatial autocorrelation in our analysis.
Thus, the first aim of this analysis is to illustrate the potential of using personal smartphone
sensor data to classify one subject’s microenvironment into three categories of indoor, outdoor,
and in-transit by building a machine learning model. We also explore how well a model trained on
one person can predict another person’s time-activity patterns. Finally, we demonstrate how
conventional cross-validation (CV) methods (e.g., randomly splitting data for training and testing)
may result in overly optimistic prediction performance compared to targeted CV methods that
account for the high spatiotemporal autocorrelation inherent in sensor data
14,15
. When ignored,
this characteristic of highly space- and time-resolved data results in models that can be used for
interpolation but cannot be extrapolated beyond the locations and time periods of the training data
16,17
.
32
3.3 METHODS
3.3.1 Data Collection
We used the AndroSensor
18
smartphone application to log Samsung Galaxy S4
environmental sensor data (operating Android OS v5.0) on 34 features. We were trying to build a
machine learning model to predict a subject’s microenvironment (indoor, outdoor, and in-transit).
Of note, S4 smartphones used in this analysis contained environmental temperature and relative
humidity sensors – a unique feature that was subsequently removed from later model phones.
These represented personal temperature and relative humidity measurements (not ambient,
outdoor or battery-related), with some degree of expected noise from real-world deployment
considerations such as carry mode or sensor calibration issues.
We recruited a convenience sample of two 30-year-old female participants who commute
to work mainly using fixed routes and modes of transport. Two types of data were collected for
both subjects: scripted and natural living (i.e., unscripted). For scripted data collection, each
subject was instructed to follow a pre-specified microenvironment and route change script
(Supplement Table S1) while a researcher accompanied them and recorded the time (to the
second) of each microenvironment change. Participants were instructed to carry the phones as they
normally would in daily life to reflect realistic conditions as much as possible. For natural living
data collection, participants were instructed to carry the smartphone as they go about their usual
daily routines and manually label or annotate when their microenvironment changed in a notepad.
We defined "Indoor" as being in any indoor enclosed location such as at home, in restaurants, at
work, or in any other building types. We defined "Outdoor" as being in an outside (non-enclosed)
environment, either motionless or in motion. "In-transit" represented in-vehicle travel either in a
car, bus, or other mode of transport.
33
Participants logged their data at 10-second intervals. Subject 1 (S1) data collection
occurred on November 9, 2016, and May 23- 26, 2017 (~63 hours over 5 days), and Subject 2 (S2)
data collection occurred June 6-9, 2017 (~56 hours over 4 days), with data collection start and end
times for each day shown in Table 3.1.
Table 3.1. Dates and times of scripted and natural living (unscripted) data collection from subjects
S1 and S2.
Subject Date Scripted data Natural living
(unscripted data)
Time (start, end) Time (start, end)
S1 11/09/2016
(Wednesday)
NA 06:20:32, 19:47:22
05/23/2017
(Tuesday)
NA 14:15:39, 23:59:59
05/24/2017
(Wednesday)
10:06:09, 11:29:59 00:00:09, 10:06:09
11:29:59, 12:25:49
05/25/2017
(Thursday)
11:11:02, 12:34:52 10:35:52, 11:11:02
12:34:52, 23:59:52
05/26/2017
(Friday)
12:35:02, 13:44:12 00:00:02, 12:35:02
13:44:12, 13:46:32
S2 06/06/2017
(Tuesday)
12:48:58, 14:00:18 10:46:08, 12:48:58
14:00:18, 23:59:58
06/07/2017
(Wednesday)
11:46:58, 13:02:58 00:00:08, 11:46:58
13:02:58, 23:59:58
06/08/2017
(Thursday)
10:13:18, 11:31:28 00:00:08, 10:13:18
11:31:28, 12:44:18
18:05:18, 23:59:58
06/09/2017
(Friday)
10:57:18, 12:14:38 00:00:08, 10:57:18
12:14:38, 14:00:58
NA = not applicable.
34
3.3.2 Data Pre-Processing
Orientation (Maps API) and Altitude (Maps API) were dropped because of high missing
frequency (88.06% and 25.65%, respectively). To ensure that the model would be blind to the
actual geographic location of participants’ major microenvironments and agnostic to person-
specific contexts (e.g., personal home or work), Latitude and Longitude were dropped as well.
Acceleration vector sum was calculated by taking the square root of the sum of squares of three
components of acceleration in the x, y, and z directions. Number of satellites in use and in range
were extracted from the GPS satellites variable as the numerator and denominator, respectively.
As speed was obtained from GPS and was missing (22.6% missing frequency for S1) in a fashion
related to GPS signal loss when being indoors and stationary, we imputed missing values so as not
to allow this missingness pattern to bias our algorithm. We calculated the missing speed values as
the distance between the current coordinate and previous data point coordinate calculated using
the Haversine formula and divided by our data logging frequency (10s). For valid estimation of
distance-based ML algorithms, such as k-nearest neighbors (KNN) and logistic regression, we also
generated a standardized version of the data with all continuous features transformed to Gaussian
distributions with a mean of 0 and standard deviation of 1. Finally, we kept data records with no
missing values on the 31 features listed in Table 3.2 for a total of 21,296 observations from S1
and 22,340 observations from S2 used in the analysis.
Table 3.2. Characteristics of the 31 measured and calculated AndroSensor features for S1
(N=21,296) by sensor category.
Feature Range Unit
Location sensor: location with longitude and latitude, along with other values such
as speed etc. Measurements are provided through GPS or the Google Maps API (over
network connection).
Location-Altitude-google [54.06, 123.67] Degree ( )
35
Location-Altitude-atmospheric-
pressure
[-69.83, 49.45] Degree ( )
Location Speed* [0, 25] Meter per second (m/s)
Location Accuracy
#
[3, 1652] Meter (m)
GPS satellites: in use [0, 9] Unitless, count
GPS satellites: in range [14, 25] Unitless, count
Accelerometer sensor: physical forces applied on the mobile phone's 3-axis plane.
Accelerometer X [-19.49, 19.61] Meter per second squared
(m/s
2
)
Accelerometer Y [-14.76, 12.71] Meter per second squared
(m/s
2
)
Accelerometer Z [-18.29, 18.50] Meter per second squared
(m/s
2
)
Total Acceleration* [2.29, 23.82] Meter per second squared
(m/s
2
)
Gravity sensor: forces measured by the accelerometer that are caused by gravity. If
the phone is still, the gravity sensor should have the same values as the
accelerometer.
Gravity X [-9.81, 9.81] Meter per second squared
(m/s
2
)
Gravity Y [-9.12, 8.26] Meter per second squared
(m/s
2
)
Gravity Z [-9.80, 9.81] Meter per second squared
(m/s
2
)
Linear acceleration sensor: forces measured by the accelerometer that are not caused
by gravity. If the phone is still, the linear acceleration values should be all 0.
Linear Acceleration X [-7.17, 12.42] Meter per second squared
(m/s
2
)
Linear Acceleration Y [-15.82, 11.69] Meter per second squared
(m/s
2
)
36
Linear Acceleration Z [-17.33, 15.48] Meter per second squared
(m/s
2
)
Gyroscope sensor: measures angular rate of motion expressed in rad/s. Integrating
angular rate with respect to time results in a measured angle of travel, which can be
used to track changes in orientation.
Gyroscope X [-4.70, 6.17] Degrees per second ( /s)

Gyroscope Y [-3.83, 3.86] Degrees per second ( /s)

Gyroscope Z [-3.46, 5.18] Degrees per second ( /s)

Magnetic sensor: measures magnetic force in the 3-axis. By computing the angle of
the detected earth’s magnetic field and comparing that measurement angle to gravity
as measured by an accelerometer, it is possible to measure a device’s heading with
respect to North with a high degree of accuracy.
Magnetic Field X [-629, 328] microTeslas (uT)
Magnetic Field Y [-185, 353] microTeslas (uT)
Magnetic Field Z [-264, 418] microTeslas (uT)
Orientation sensor: shows which way the phone is pointing at. X is relative to
magnetic north pole (0 means device’s top points to the north pole), Y is relative to
the ground (0 means phone is left flat, 90 means that phone is standing vertical) and
Z shows “roll”, this is rotation around itself where 0 means phone is facing to the sky
and 180 means the screen is facing the ground.
Orientation X [-57.46, 80.00] Degree ( )
Orientation Y [-179.64, 179.95] Degree ( )
Orientation Z [0.00, 360.00] Degree ( )
Pressure sensor (Barometer): measures the atmospheric pressure in the user’s
location. The pressure sensor can be used as a normal barometer, but it also helps for
a more accurate calculation of the user’s elevation/altitude, since GPS on its own can
be off by unacceptable amounts.
Atmospheric Pressure [745.90, 755.00] Millibar (mBar)
Proximity sensor: examines whether there is another object in close proximity
(within ~3.5 inches radius) of the face of the mobile phone.
Proximity [0, 8] Centimeter (cm)
37
Sound level: measures the sound level (in decibels) of the mobile phone's
surroundings through the microphone.
Sound Level [0, 77.52] Decibel (dB)
Temperature sensor: measures temperature in immediate surroundings (personal).
Temperature [20.30, 36.70] Degrees Celsius (°C)
Relative humidity sensor measures the air’s relative humidity in immediate
surroundings in percent (%). Relative humidity is a term used to describe the amount
of water vapor in a mixture of air and water vapor.
Relative Humidity [20.20, 70.20] Percent (%)
Light Sensor: measures the intensity of light or illumination detected by smartphone
sensor.
Light [0, 784] Lux (lx)
*Features calculated based on original sensor measured features: Location speed: calculated by
first generating the distance between the current coordinate and previous data point coordinate
using Haversine formula, then divided by our data logging frequency (10s). Total acceleration:
calculated as vector sum or square root of the sum of squares of three components (x, y, z) of
acceleration.
#
Location accuracy had a wide range was because of the change of indoor and outdoor.
3.3.3 Additional Feature Extraction
We calculated the first and second derivatives of the 31 features since we hypothesized that
the degree and rate of change, respectively, of a signal over time may also be informative for
microenvironment classification. We then calculated moving window summaries of all variables
as the mean, median, standard deviation, maximum, minimum in moving windows of 1 and 5
minutes. Moving window summaries capture the change of a sensor value over time and have been
used in many previous studies
19–21
.
3.3.4 Machine Learning Algorithms Evaluated
We evaluated five multi-class classifiers in terms of their ability to predict
microenvironment in three classes of indoor, outdoor and in-transit based on 10-second data. We
applied the following popular machine learning (ML) classification algorithms for model training
38
using the R package “MLR”: k-nearest neighbor (KNN), naïve Bayesian classifier (NB), logistic
regression (LR), Classification and Regression Tree (CART) decision tree (DT) algorithm, and
random forests (RF). KNN classifies an object purely by the proximity (distance) of its neighbors.
NB is a classification technique based on Bayes’ Theorem with an assumption of independence
among predictors. LR is a statistical model that uses a logistic function to model a binary or multi-
class dependent variable. DT uses tree representation to solve the problem. Each node of the tree
represents an attribute, and each leaf node represents a class label. The RF algorithm consists a
large number of individual decision trees that operate as an ensemble: each tree has a class
prediction and the class with the most votes becomes the model’s prediction
22
.
3.3.5 Initial Model Training and Testing
Data from S1 was used for the primary model building, and data from S2 was used as an
independent testing set. We noted that the microenvironment class-ratio of our data was
unbalanced (% indoor:in-transit:outdoor was 77:17:6 in S1). Therefore, stratified sampling was
used to ensure that the training and testing sets had approximately the same proportion of samples
from each target class. The dataset was split into two subsets with 70% for training and validation
and 30% for testing without shuffling to maintain the original temporal order of the data. We used
10-fold CV to select optimal hyperparameters in the training and validation stage.
We initially trained three sets of models using the five ML algorithms described earlier to
examine whether the additional features we extracted improved model performance. We described
these sets as M1 (base models with 31 original features), M2 (M1 features plus their 1
st
and 2
nd

derivatives, for a total of 93 features), and M3 (M2 features plus their moving 1-min and 5-min
summaries for a total of 1,023 features).
39
3.3.6 Model Prediction Accuracy Assessment
In each model’s testing phase (five algorithms and three sets of models for a total of 15
models), the model was applied to the held-out samples to classify microenvironment. Balanced
accuracy (BA) is a method based on the confusion matrix commonly employed for unbalanced
data.
23
BA weights each class equally since it is calculated as the average of the correctly classified
proportion of the data per class. We selected the best microenvironment classifier based on highest
BA. We compared the computational time and BA of the five ML algorithms trained and applied
on S1 across the three sets of models (M1, M2, and M3) to determine the optimal ML algorithm
and number of input features to proceed with, respectively. Once the final model was selected, it
underwent additional testing as described below.
3.3.7 Training Data Considerations and Feature Importance Ranking
In order to investigate how much data would have been sufficient to achieve the same BA
of the final model under the conventional CV scenario, and differences in having scripted versus
unscripted data, we iteratively created smaller subsets of S1 data, re-ran the final selected ML
algorithm, and compared the change in BA with fewer data available to train the model. We first
ran the model on each of Days 1-5 individually, then took different subsets of Day 5 (includes
scripted and unscripted data) and from the scripted portion only.
To rank and understand the relative importance of each feature at predicting
microenvironment in our final model, we applied the permutation feature importance method
24,25
.
The drop-column feature importance method was also applied to understand the influence of
individual features on the final model
25
. Furthermore, to get the feature importance ranking for
each microenvironment class, we applied the “one-vs-all” algorithm
26
. For example, to calculate
40
feature importance for the indoor sub-class, we combined the outdoor and in-transit sub-classes as
one non-indoor class and re-ran the permutation feature importance.
3.3.8 Additional Model Prediction Accuracy Assessment Under Various Scenarios
All model training and hyperparameter tuning were conducted using S1 training data. Once
the final optimal model was selected, we conducted several tests to assess its prediction accuracy
on the same subject (S1) and on a different subject (S2) under various scenarios. These included
independent tests on S2 data under natural living and scripted activity settings, and tests on both
S1 and S2 data accounting for temporal (only), spatial (only), or spatiotemporal (both)
autocorrelation in the data.
3.3.9 Independent Subject (S2) Tests with Variable Degrees of Resemblance in Time-Activity Patterns
We leveraged scripted and natural living (unscripted) activity periods in the data to
examine the performance of the model trained on S1 at predicting the time-resolved
microenvironment classes of S2 as an independent subject, while varying the range of similarity
between time-activity patterns of S1 and S2. We used S2’s scripted data as an independent test on
a different person with very similar time-activity patterns. This test can potentially provide insights
into how well a model trained on one person can predict another person’s microenvironment if
they happen to share very similar daily time-activity patterns or routines (e.g., college students,
working adults, etc.). We used S2’s natural living data as an independent test based on natural
living conditions and activities.
3.3.10 Same Subject (S1) and Independent Subject (S2) Tests Considering Temporal Autocorrelation
For data with temporal dependency, care must be taken when splitting the data to prevent
the model from learning from “future” data that would not be available in a real-world scenario
27,28
. To address this, we applied a day forward-chaining (DFC) nested CV method
29
as shown in
41
Figure 3.1 with an outer loop for error estimation and an inner loop for parameter tuning which
splits the training set into a training and a validation subset to select the parameters that minimized
error. We also added an outer loop that splits the dataset into multiple different training and test
sets, and the error on each split was averaged in order to compute a robust estimate of model error.
To retain five independent days of data, we dropped data from midnight-2AM for same subject
(S1) and independent subject (S2) tests. We chose two hours as the duration of time after which
the temporal autocorrelation of sound level (selected as the most important predictor in M1 model)
approached zero based on an auto regressive integrated moving average (ARIMA) model
(Supplement Figure S2).

Figure 3.1. DFC nested CV workflow to account for temporal autocorrelation in prediction
performance evaluations (reprinted from: https://towardsdatascience.com/time-series-nested-CV-
76adba623eb9)
51
.
3.3.11 Same Subject (S1) and Independent Subject (S2) Tests Considering Spatial Autocorrelation
Spatial data, especially when collected at high resolution, is not statistically independent.
Conventional CV may fail to capture model extrapolation performance because the training and
testing data points are too close to each other in space and thus too similar. Spatial partitioning can
be used to split observations into spatially disjointed subsets to overcome this problem and account
for spatial autocorrelation when evaluating models
16
. We first constructed square spatial blocks
of a specified size defined as the median of the spatial autocorrelation ranges from the 6 most
42
important features based on the permutation method. The range is the distance where the
semivariogram reaches an asymptote beyond which we assume there is minimal spatial
autocorrelation remaining in the data
30
. We allocated blocks into three folds by setting constraints
that avoid folds with little or no representation of all 3 microenvironment classes. We used the R
packages “BlockCV” and “MLR” to conduct this spatial blocking followed by nested spatial CV
on the same subject (S1) and independent subject (S2).
3.3.12 Same Subject (S1) and Independent Subject (S2) Tests Considering Spatiotemporal
Autocorrelation
This last test considers the co-occurrence of both temporal and spatial autocorrelation in
the data. We used the R packages “BlockCV” and “MLR” to conduct this evaluation. Similar to
spatial evaluations, we first defined spatial blocks, but instead we allocated blocks into five folds
to match the number of days in the data. We then conducted a nested spatiotemporal CV as shown
in Figure 3.2. Briefly, we selected testing data from the same spatial block and day as the outer
fold number. For example, outer fold 1 used testing data from spatial block 1 and day 1, outer fold
2 used testing data from spatial block 2 and day 2, etc. We only used data from non-testing blocks
and days for training/validation for each fold. For example, training/validation data for outer fold
1 used data from spatial blocks 2, 3, 4 and 5 and days 2, 3, 4 and 5. Similarly to the temporal and
spatial evaluations, we used S2 as an independent test.

43
Figure 3.2. Nested spatiotemporal CV framework to account for spatiotemporal autocorrelation
in prediction performance evaluations, adapted and modified from
17
. The color represents
assigned spatial blocks and the shape represents different days of data collection denoted by D
(e.g., D1, D2, etc.).
3.4 RESULTS
In total, we collected 21,296 observations from S1 of which 1,082 (5%) were scripted and
22,340 observations from S2 of which 1,544 (7%) were scripted. The microenvironment
distribution in the scripted portions of the data is similar across the two subjects (% indoor:in-
transit:outdoor is 26:16:58 for S1 and 24:16:60 for S2) and more variable during natural living (%
indoor:in-transit:outdoor is 80:16:4 for S1 and 87:7:6 for S2). We compared our estimated speed
to the original logged speed and found they were highly correlated (Pearson correlation=0.76)
(Supplement Figure S1).
3.4.1 Optimal Machine Learning Model for Microenvironment Classification
The first column of Table 3.3 shows the test performance results for the five ML algorithms
trained and tested on S1 data using conventional CV. RF performed the best among the algorithms
(95.1% BA), followed by KNN (90.5% BA) and DT (82.8% BA) in M1 models. LR and NB
classifier performed the worst with 77.6%, and 73.5% BA, respectively. After including additional
extracted features in M2 and M3 models, RF still had the highest accuracy (Table 3.3). Because
M1 with the original 31 features yielded slightly higher prediction accuracy and models with more
features needed longer computational time, the RF model with 31 features (M1) was selected as
the optimal microenvironment classification model (hereafter refers to as the optimal or final
model). Its confusion matrix is shown in Supplement Table S2. The model predicts S1’s indoor,
in-transit and outdoor microenvironments with very high accuracy (99.4%, 97.0% and 89.0%,
respectively).
44
Table 3.3. Prediction performance (reported as balanced accuracy, %) of five ML classifiers based
on S1 data using 31 features (M1), 93 features (M2), and 1,023 features (M3).
M1 M2 M3
Total number of features in the model
Algorithm 31 Features
1
93 Features
2
1,023 Features
3
Random forests 95.1% 95.0% 94.6%
K-Nearest Neighbor 90.5% 90.1% 91.1%
Decision Tree 82.8% 83.5% 83.7%
Naïve Bayesian 77.6% 77.5% 74.5%
Logistic Regression 73.5% 73.5% 72.1%
1
31 features as listed in Table 2.
2
93 features including the 31 features from M1 and their 1
st
and 2
nd
derivatives.
3
1,023 features including M2 features and their rolling statistics (mean, median, standard deviation,
minimum, maximum) in 1- and 5-minute moving windows.
As for how much data would have been sufficient to achieve similar BA (~95%) of the
final RF model, all single day models achieved >90% BA, but days that included scripted data had
higher BA (94.7%, 94.3%, and 93.0% for Days 5, 4, and 3 respectively) compared to days without
scripted data (92.2% for Day 1 and 92.1% for Day 2). When running the RF model on Day 5 data
only, less training data generally resulted in lower BA, but only ~30% of Day 5 data (equivalent
to ~4.2 hours) was enough to obtain >90% BA (Supplement Table S3). However, when using
scripted data with carefully designed and more balanced representation of microenvironments and
more frequent transitions across them, the minimum amount of training data needed to achieve
BA >90% was ~30 minutes (Supplement Table S4).
Figure 3.3 shows the ten most important predictive features for the overall model and for
each microenvironment ranked by the permutation feature importance method. Sound level, speed,
temperature, location accuracy and relative humidity are the top five features for overall model
prediction. The drop-column feature importance method resulted in similar feature importance
ranking as the permutation method for the overall model (Supplement Figure S3). Partial
dependence plots for the top five ranked features are shown in Supplement Figure S4.
45
Figure 3.3. Top 10 most important features for overall and microenvironment class prediction
ranked by permutation feature importance method from the optimal, final model in S1 (random
forests algorithm with 31 features).

For each microenvironment sub-class, sound level was always the most important feature.
The rankings in indoor and in-transit have similar patterns as the overall ranking, where sound
level, speed, temperature, relative humidity, and location accuracy ranked among the top features,
but satellites in range played a more important role for predicting the in-transit microenvironment.
Sound level, gravity (X), accelerometer (X), location accuracy, and satellites in range were the top
five important features for outdoor prediction.
3.4.2 Model Prediction Accuracy in Independent Subject (S2) Tests with Variable Degrees of
Resemblance in Time-Activity Patterns
The prediction accuracy of our final model dropped from 95.1% to 50.6% when applied to
S2 data during natural living conditions with the greatest dissimilarity in time-activity patterns
(Model 1 in Table 3.4). However, BA increased to 70.0% when applied to S2 during scripted time-
activity sessions with similar indoor:in-transit:outdoor ratios (Model 2 in Table 3.4). Results from
additional scenarios with variable degrees of resemblance in time-activity patterns across S1 and
S2 are presented in Supplement Table S5.
46
Table 3.4. Model prediction accuracy in independent subject tests (S2) under natural living
(unscripted) and scripted time-activity patterns.
Model
#
Data Used
For
From Subject Natural living or
scripted data

% indoor:in-transit:outdoor
microenvironments

BA (%)
1 Training Subject 1 Natural living
(unscripted)
80:16:4
Testing Subject 2 Natural living
(unscripted)
87:7:6 50.6%
2 Training Subject 1 Scripted 26:16:58
Testing Subject 2 Scripted 24:16:60 70.0%

3.4.3 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Temporal Autocorrelation
The model performance in S1 based on DFC nested CV dropped from 95.1% (conventional)
to 63.2% when considering temporal autocorrelation, calculated as the average BA from outer
loops one, two, and three (50.6%, 66.7%, and 72.1%, respectively). Moreover, outer loops with
more training data (outer loop 3 versus 1 and 2) had higher BA (Table 3.5). In S2 independent
tests, BA dropped to 43.8% (compared to 44.7% conventional CV) (Supplement Table S6).
Table 3.5. Model prediction accuracy in same subject (S1) tests accounting for temporal
autocorrelation (DFC) compared to conventional CV (stratified 7:3 random selection).
Training/Testing
Data Selection
Method
Hyperparameter
Tuning Method
Outer
Loop #
Training
Data
Validation
Data
Testing Data BA (%)
DFC Nested CV 1 Day 1 Day 2 Day 3 50.6%
2 Day 1, 2 Day 3 Day 4 66.7%
3 Day 1, 2, 3 Day 4 Day 5 72.1%
Stratified 7:3
random selection
Conventional CV 1 70% (Day1, 2, 3) 30% (Day 1, 2, 3) 95.6%
2 70% (Day1, 2, 3, 4) 30% (Day 1, 2, 3, 4) 95.2%
3 70% (Day1, 2, 3, 4, 5) 30% (Day 1, 2, 3, 4, 5) 95.5%

47
3.4.4 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Spatial Autocorrelation
We selected the median spatial autocorrelation range of 1,006 meters as our block size
based on the top five most predictive features (sound level, speed, temperature, location accuracy
and relative humidity) and the number of total satellites in range (Figure 3.4A). Figure 3.4B shows
the corresponding spatial blocks and folds defined based on unique S1 locations by setting
constraints to avoid folds with no representation of all three microenvironments.

Figure 3.4. Spatial autocorrelation ranges for six selected features (A) and the corresponding
spatial blocks created by setting block size equal to the median spatial autocorrelation range
across these features (B).

The BA of the first, second, and third outer fold was 68.5%, 58.9%, and 55.2%,
respectively, and the overall BA in S1 after accounting for spatial autocorrelation was 60.9%
(compared to 95.1% in conventional CV). In S2 independent tests, BA was 41.8%, 33.8%, and
33.3% respectively in the three outer folds, and the overall BA was 36.3% (compared to 44.7%
conventional CV in S2).
48
3.4.5 Model Prediction Accuracy in Same Subject (S1) and Independent Subject (S2) Tests
Considering Spatiotemporal Autocorrelation
Supplement Figure S5 shows the five-fold allocation for spatiotemporal nested CV. S1
model BA was 76.0%, 20.6% and 57.5% in the first, second and fourth outer folds respectively,
with an overall BA of 51.4%. The BA of outer folds three and five could not be calculated because
the outdoor (fold three) and the indoor and outdoor (fold five) were not represented. BA in S2
dropped to 39.4% when considering spatiotemporal autocorrelation (compared to 44.7% in
conventional CV).
3.5 DISCUSSION
Using smartphone sensor data and a RF classifier, we built a prediction model that can
classify time-activity patterns in three major microenvironments with the overall goal of improving
personal air pollution exposure assessment and minimizing measurement error in health studies.
As smartphone sensors and wearables are continuously improving in accuracy and functionality at
a rapid rate, our work serves to demonstrate their vast potential in air pollution exposure and health
applications. We found that physical environment (sound level, relative humidity), geolocation
metadata (location accuracy and number of satellites) and motion related variables (speed) were
some of the most important predictive features, with slight differences by microenvironment class.
Our main findings indicate that: 1) mHealth methods have the potential to improve
microenvironment classification; 2) models trained on one person could be generalized to groups
of people that share similar daily time-activity patterns; and 3) when using highly space- and time-
resolved sensor data for building predictive models, conventional CV approaches that do not
account for the spatiotemporal autocorrelation inherent in these data streams can lead to overly
optimistic prediction performance results due to violation of the IID assumption.
49
3.5.1 Mobile Health (mHealth) Approaches to Microenvironment Classification Problems
As Larkin and Hystad
31
stated, the integration of air pollution sensors, smartphone
applications, and models has the potential to enhance personal air pollution exposure assessment
especially in large-scale population-based health studies. In 2017, ~68% of the U.S. population
used a smartphone, and ~29% of them downloaded an application related to health management
or tracking
32
. This number of smartphone users relying on sensors for health-related tracking is
increasing
33
.
There are two general mHealth approaches for microenvironment classification: GPS-
based and smartphone sensor-based. Some GPS-based classification methods depend on image
processing technologies that require prior maps or satellite imagery
34
. Others might spatially
intersect or calculate distances between GPS points and boundaries or centroids of points of
interest (e.g., building footprints) to classify microenvironment. Given sometimes inaccurate, out
of date, or oddly shaped building footprint data, and increased poorer GPS location accuracy when
indoors, these methods can have a difficult time distinguishing indoor and outdoor
microenvironment with high resolution.
Whereas, smartphone sensor-based methods can overcome some of the limitations of using
only GPS to classify microenvironment
35
. The intensity, variation or rate of change in a user’s
surrounding environmental properties (e.g., light
36
, temperature
37
, sound
38
) can potentially
indicate transitions between different microenvironments. Our approach can overcome single
sensor shortcomings
39
by using information from several sensors together, and future models can
be similarly trained on smartphone sensor data from newer models with improved performance.
50
3.5.2 Optimal Model Performance and Feature Importance
RF algorithm performed optimally in our application. RF is a well-known ensemble
classification method
40
. Its performance is generally considered to be robust to noise with less
parameter tuning needs
41
.
Sound level was the most important feature for predicting microenvironment overall and
by class. Locher et al.
42
and Dai
43
also found informative differences between indoor and outdoor
sound levels. Sound level in our data varied across the three microenvironments with the greatest
variation seen indoors, and the highest median levels outdoors possibly explaining its predictive
value (Supplement Figure S6). Speed is also widely used in indoor/outdoor classification
problems
44,45
. Speed ranked in the top two positions for indoor and in-transit prediction in our
work possible due to higher speeds usually encountered in a car or bus in-transit compared to lower
speeds when indoors
44
.
Location accuracy was also an important predictive feature in our model, similarly reported
in Lee et al.
46
. Its value (in meters, where a larger number corresponding to a wider radius and
poor accuracy) generally increases inside buildings compared to being in an open, outdoor
environment (with some exceptions when surrounded with complex structures or urban canyons
or being out of satellite range), and it correlated with GPS signal loss indoors. Figure 3.5 illustrates
two examples of changes in location accuracy when transitioning between indoor and outdoor
microenvironments. Blue dots forming a right-angle pattern in Figure 3.5A show the subject
walking outdoors along the perimeter of a building, while red dots represent a larger radius for
location accuracy (worse geolocation signal) indoors inside the building. Figure 3.5B
demonstrates the sudden increase in location accuracy (in outlined box) when the subject
transitioned from the outdoor to the indoor microenvironment. Also, Wi-Fi signals tend to have
51
high accuracy indoors, so future study may also consider the Wi-Fi connection as a potential
feature.
Figure 3.5. Illustration of location accuracy changes in space (A) and in time (B) between indoor
and outdoor microenvironments.

3.5.3 Generalizability of the Optimal Model Across People for Larger Scale Applications
For a model to be useful for larger scale exposure and epidemiology applications,
generalizability across individuals is also highly desirable. Our first goal was trying to develop a
personal microenvironment prediction model and secondly, to try to evaluate if the model can be
applied to other subjects. Given that we trained our model on S1, we found that it did not translate
well to an independent subject (S2) under the natural living scenario (Model 1 in Table 3.4). This
was the case despite intentionally withholding geographic coordinates as inputs to our model to
make it blind to one person’s commonly frequented locales (specific home and work coordinates)
and instead forcing it to rely on variations in sensor data to learn and predict microenvironment.
However, our model accuracy increased significantly (70.0%) if S1 and S2 shared more similar
time-activity patterns as shown in Model 2 in Table 3.4 and Model 12 in Supplement Table S5.
52
This scenario can apply in real life to several groups of people that share somewhat similar time-
activity patterns, such as students attending the same school or college, or adults commuting to
work on somewhat similar schedules.
In addition, several recommendations are provided to overcome generalizability or
extrapolation challenges across individuals for future model training efforts. First, diversifying the
training data in terms of the representativeness and variety of the predicted outcome being
observed in space and time is always desirable, despite our findings that ~30 minutes of data per
person were generally sufficient to obtain similar performance. Further, devising ways to minimize
the burden of collecting more diverse and less human error-prone sets of annotated training data
is crucial. For example, crowdsourcing the collection of annotation data through ecological
momentary assessment (EMA) smartphone surveys or applications (apps) can be helpful. These
can be designed to prompt users to annotate their microenvironment for a more intensive period
of “learning” time that can later be downscaled in a random and/or context-sensitive fashion (e.g.,
sudden change in 1+ sensor readings).
3.5.4 Cross Validation Strategies That Consider Spatiotemporal Autocorrelation Inherent in High
Resolution Sensor Data
Selecting the appropriate CV strategy depends on the researcher’s ultimate prediction goal
and intended future use of the model. If the goal is to be able to predict values that are very similar
or close to the observed training data (interpolation or in-sample tests), conventional CV
approaches can suffice. However, if the goal is to extrapolate or generalize the model beyond the
observed training data (out of sample tests), more sophisticated CV strategies are needed that
consider inherent spatial and temporal dependencies in the data. In personal exposure modeling
and mHealth applications, a combination of the two goals is often desired (reproducing training
53
data while also extrapolating further beyond it, within defined or expected ranges). In this case,
especially when using highly spatiotemporally resolved sensor data, relying solely on conventional
CV strategies will result in overly optimistic results. Therefore, special consideration should be
taken to determine the appropriate separation in time and space when selecting training and testing
data to reflect these desired ranges.
Specifically for time series data,
28
researchers need to pay attention to the appropriateness
of forecast-error measures, and prediction accuracy should be assessed using out-of-sample tests
rather than the goodness of fit to past data
47
. Similarly, for highly spatially resolved data, prediction
evaluations must account for spatial autocorrelation
17,48
. These issues similarly apply to a range
of other environmental exposure prediction applications, such as air pollution, heat, or noise
exposure models.
3.5.5 Limitations, Strengths and Future Directions
Many of the sensors embedded in the Samsung Galaxy S4 smartphones used in this study
are currently outdated; however, the goal of our work was to demonstrate the potential of
smartphone sensor data streams in air pollution exposure applications and to offer
recommendations for training and testing newer models. More specifically, the environmental
temperature and relative humidity sensors were subsequently removed from later Samsung
smartphones. However, a variety of wearable, personal temperature and relative humidity sensors
are currently available and can be substituted in future model building efforts. Furthermore, we
expect our model performance across individuals, locations and time periods to improve with the
ability to collect more diverse training data in the future. However, as with all models,
extrapolation beyond training data is usually limited. Our work illustrates methods to formally and
explicitly test and report on this.
54
Compared to a similar prior study, our model had significantly greater temporal and spatial
data coverage
49
. By designing scripted data collection, we managed to increase the utility of our
training data (especially important for unbalanced outcomes) despite only having two participants.
We also showed that ~30 minutes of carefully scripted and executed data collection might suffice
to achieve >90% prediction accuracy. Areas of future improvement include differentiating
transport mode (e.g., biking in-transit) or physical activity type (e.g. walking outdoors vs indoors)
within microenvironments building on our previous work
50
. Finally, while providing the model
with a priori information on most frequented locations (e.g., home and work) might improve
performance, we refrained from using these strategies to minimize overfitting and data privacy
concerns with potential future exposure and health applications (e.g., sending GPS locations across
the internet to retrieve this information).
In conclusion, we illustrated an example of using smartphone environmental sensor data to
classify microenvironments. Our work provides potential directions and recommendations for
developing future models and demonstrates the potential of mHealth approaches at accounting for
time-activity patterns and minimizing exposure measurement error in air pollution studies. As
sensors and technology develop, we expect these approaches to continually improve. We found
that models can potentially be generalizable to groups of people if they share somewhat similar
activities or routines; however, the minimum amount of data to train an individualized model was
also fairly minimal. Finally, we demonstrated the importance of accounting for spatiotemporal
autocorrelation in high resolution sensor data to properly characterize model prediction
performance.

55
3.6 REFERENCES:
1. Pope CA, Dockery DW. Health effects of fine particulate air pollution: Lines that connect.
J Air Waste Manag Assoc. 2006. doi:10.1080/10473289.2006.10464485
2. Howe CG, Eckel SP, Habre R, et al. Association of Prenatal Exposure to Ambient and
Traffic-Related Air Pollution With Newborn Thyroid Function: Findings From the
Children’s Health Study. JAMA Netw open. 2018.
doi:10.1001/jamanetworkopen.2018.2172
3. Kim JS, Alderete TL, Chen Z, et al. Longitudinal associations of in utero and early life
near-roadway air pollution with trajectories of childhood body mass index. Environ Heal
A Glob Access Sci Source. 2018. doi:10.1186/s12940-018-0409-7
4. Schultz AA, Schauer JJ, Malecki KM. Allergic disease associations with regional and
localized estimates of air pollution. Environ Res. 2017. doi:10.1016/j.envres.2017.01.039
5. Cohen AJ, Brauer M, Burnett R, et al. Estimates and 25-year trends of the global burden
of disease attributable to ambient air pollution: an analysis of data from the Global Burden
of Diseases Study 2015. Lancet. 2017. doi:10.1016/S0140-6736(17)30505-6
6. Li W, Wilker EH, Dorans KS, et al. Short-Term Exposure to Air Pollution and
Biomarkers of Oxidative Stress: The Framingham Heart Study. J Am Heart Assoc. 2016.
doi:10.1161/JAHA.115.002742
7. Heinzerling A, Hsu J, Yip F. Respiratory Health Effects of Ultrafine Particles in Children:
A Literature Review. Water Air Soil Pollut. 2016. doi:10.1007/s11270-015-2726-6
8. Klepeis NE, Nelson WC, Ott WR, et al. The National Human Activity Pattern Survey
(NHAPS): A resource for assessing exposure to environmental pollutants. J Expo Anal
Environ Epidemiol. 2001. doi:10.1038/sj.jea.7500165
56
9. Georgopoulos PG, Lioy PJ. Conceptual and theoretical aspects of human exposure and
dose assessment. J Expo Anal Environ Epidemiol. 1994.
10. Sexton K. Air Pollution, the Automobile, and Public Health. In: Air Pollution, the
Automobile, and Public Health. ; 1988:208-238.
11. Dockery DW, Spengler JD. Personal exposure to respirable particulates and sulfates. J Air
Pollut Control Assoc. 1981. doi:10.1080/00022470.1981.10465205
12. Cho GH, Rodríguez DA, Evenson KR. Identifying walking trips using GPS data. Med Sci
Sports Exerc. 2011. doi:10.1249/MSS.0b013e3181ebec3c
13. Dey AK, Wac K, Ferreira D, Tassini K, Hong JH, Ramos J. Getting closer: An empirical
investigation of the proximity of user to their smart phones. In: UbiComp’11 -
Proceedings of the 2011 ACM Conference on Ubiquitous Computing. ; 2011.
doi:10.1145/2030112.2030135
14. Bahn V, McGill BJ. Testing the predictive performance of distribution models. Oikos.
2013. doi:10.1111/j.1600-0706.2012.00299.x
15. Juel A, Groom GB, Svenning JC, Ejrnæs R. Spatial application of Random Forest models
for fine-scale coastal vegetation classification using object based analysis of aerial
orthophoto and DEM data. Int J Appl Earth Obs Geoinf. 2015.
doi:10.1016/j.jag.2015.05.008
16. Valavi R, Elith J, Lahoz-Monfort JJ, Guillera-Arroita G. blockCV: An r package for
generating spatially or environmentally separated folds for k-fold cross-validation of
species distribution models. Methods Ecol Evol. 2019. doi:10.1111/2041-210X.13107
17. Meyer H, Reudenbach C, Hengl T, Katurji M, Nauss T. Improving performance of spatio-
temporal machine learning models using forward feature selection and target-oriented
57
validation. Environ Model Softw. 2018. doi:10.1016/j.envsoft.2017.12.001
18. AndroSensor. Android Application, AndroSensor.Google Play Store,2012. Google Play
Store. https://play.google.com/store/apps/details?id=com.fivasim.androsensor&hl=en_US.
Published 2012.
19. Anagnostopoulos T, Garcia JC, Goncalves J, Ferreira D, Hosio S, Kostakos V.
Environmental exposure assessment using indoor/outdoor detection on smartphones. Pers
Ubiquitous Comput. 2017;21(4):761-773. doi:10.1007/s00779-017-1028-y
20. Wang W, Chang Q, Li Q, Shi Z, Chen W. Indoor-outdoor detection using a smart phone
sensor. Sensors (Switzerland). 2016. doi:10.3390/s16101563
21. Zhu Y, Luo H, Wang Q, et al. A fast indoor/outdoor transition detection algorithm based
on machine learning. Sensors (Switzerland). 2019. doi:10.3390/s19040786
22. James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statiscal Learning with
Applications in R.; 2013. doi:10.1016/j.peva.2007.06.006
23. García V, Mollineda RA, Sánchez JS. Index of balanced accuracy: A performance
measure for skewed class distributions. In: Lecture Notes in Computer Science (Including
Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). ;
2009. doi:10.1007/978-3-642-02172-5_57
24. Strobl C, Boulesteix AL, Kneib T, Augustin T, Zeileis A. Conditional variable importance
for random forests. BMC Bioinformatics. 2008. doi:10.1186/1471-2105-9-307
25. Parr T, Turgutlu K, Csiszar C, Howard J. Beware Default Random Forest Importances.
ExplainedAi. 2018.
26. Bishop C. Pattern Recognition and Machine Learning. Springer; 2006.
27. Wiens J, Shenoy ES. Machine Learning for Healthcare: On the Verge of a Major Shift in
58
Healthcare Epidemiology. Clin Infect Dis. 2018. doi:10.1093/cid/cix731
28. Tashman LJ. Out-of-sample tests of forecasting accuracy: An analysis and review. Int J
Forecast. 2000. doi:10.1016/S0169-2070(00)00065-0
29. Cochran C. Time Series Nested Cross-Validation. https://towardsdatascience.com/time-
series-nested-cross-validation-76adba623eb9. Published 2018.
30. Matheron G. Principles of geostatistics. Econ Geol. 1963.
doi:10.2113/gsecongeo.58.8.1246
31. Larkin A, Hystad P. Towards Personal Exposures: How Technology Is Changing Air
Pollution and Health Research. Curr Environ Heal reports. 2017. doi:10.1007/s40572-
017-0163-y
32. Purcell K. Half of adult cell phone owners have apps on their phones. 2011:1-33.
http://pewinternet.org/Reports/2011/Apps-update.aspx.
33. Swan M. Health 2050: The realization of personalized medicine through crowdsourcing,
the quantified self, and the participatory biocitizen. J Pers Med. 2012;2(3):93-118.
doi:10.3390/jpm2030093
34. Li X, Wang J, Li T. Seamless positioning and navigation by using geo-referenced images
and multi-sensor data. Sensors (Basel). 2013. doi:10.3390/s130709047
35. Quinn C, Anderson GB, Magzamen S, Henry CS, Volckens J. Dynamic classification of
personal microenvironments using a suite of wearable, low-cost sensors. J Expo Sci
Environ Epidemiol. 2020. doi:10.1038/s41370-019-0198-2
36. Zhou P, Zheng Y, Li Z, Li M, Shen G. IODetector - A generic service for indoor outdoor
detection. In: SenSys 2012 - Proceedings of the 10th ACM Conference on Embedded
Networked Sensor Systems. ; 2012. doi:10.1145/2426656.2426709
59
37. Krumm J, Hariharan R. Tempio: inside/outside classification with temperature. Second Int
Work Man- …. 2004:241-250.
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=646032%5Cnhttp://www.msr-
waypoint.com/en-us/um/people/jckrumm/Publications 2004/tempio_distribute.pdf.
38. Sung R, Jung S hoon, Han D. Sound based indoor and outdoor environment detection for
seamless positioning handover. ICT Express. 2015. doi:10.1016/j.icte.2016.02.001
39. Ali M, Elbatt T, Youssef M. SenseIO: Realistic Ubiquitous Indoor Outdoor Detection
System Using Smartphones. IEEE Sens J. 2018. doi:10.1109/JSEN.2018.2810193
40. Li Y, Gao J, Li Q, Fan W. Ensemble learning. In: Data Classification: Algorithms and
Applications. ; 2014. doi:10.1201/b17320
41. Boehmke,Bradley Greenwell B. No Title. In: Hands-On Machine Learning with R. ; 2019.
42. Locher B, Piquerez A, Habermacher M, et al. Differences between outdoor and indoor
sound levels for open, tilted, and closed windows. Int J Environ Res Public Health. 2018.
doi:10.3390/ijerph15010149
43. Dai W. Acoustic Scene Recognition with Deep Learning. 2016:1-18.
44. Saffar I, Morel MLA, Singh KD, Viho C. Machine Learning with partially labeled Data
for Indoor Outdoor Detection. 2019 16th IEEE Annu Consum Commun Netw Conf CCNC
2019. 2019. doi:10.1109/CCNC.2019.8651736
45. Saffar I, Morel MLA, Singh KD, Viho C. Semi-Supervised Deep Learning-Based
Methods for Indoor Outdoor Detection. IEEE Int Conf Commun. 2019;2019-May.
doi:10.1109/ICC.2019.8761297
46. Lee B, Lim C, Lee K. Classification of indoor-outdoor location using combined global
positioning system (GPS) and temperature data for personal exposure assessment. Environ
60
Health Prev Med. 2017. doi:10.1186/s12199-017-0637-4
47. Bergmeir C, Benítez JM. On the use of cross-validation for time series predictor
evaluation. Inf Sci (Ny). 2012. doi:10.1016/j.ins.2011.12.028
48. Pohjankukka J, Pahikkala T, Nevalainen P, Heikkonen J. Estimating the prediction
performance of spatial models via spatial k-fold cross validation. Int J Geogr Inf Sci.
2017. doi:10.1080/13658816.2017.1346255
49. Wang W, Chang Q, Li Q, Shi Z, Chen W. Indoor-outdoor detection using a smart phone
sensor. Sensors (Switzerland). 2016;16(10). doi:10.3390/s16101563
50. A. Hosseini, D. Zamanzadeh, L. Valencia, R. Habre AATB and MS. Domain Adaptation
in Children Activity Recognition. In: 2019 41st Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany. ;
2019:1725-1728.

61
3.7 SUPPLEMENT
Table S1. Microenvironment and route change script for scripted data collection.
NAME:
DATE:
Activity Time
Start
Time End Notes
1. Soto building- (cafeteria) (~3 mins) (Inside)
2. Inside/Outside Transition (exactly 3 mins each)
a. Outside (walk toward pay parking, back of Soto)
b. Inside (walk inside to gym/classrooms)
c. Outside (walk toward pay parking, back of Soto)
d. Inside (walk inside to gym/classrooms)
e. Outside (walk toward pay parking, back of Soto)
f. Inside (walk inside to gym/classrooms)
3. Take shuttle to Norris (In-Transit)
4. Walk through hospital to outdoor plaza (~3mins) (Outside)
5. Plaza (outside) (~3mins)
62
6. Food court (Norris) (~3 mins) (Inside)
7. Plaza (outside and walk around the hospital to shuttle)
8. Shuttle back to Soto (In-transit)
9. Back to Soto- Out of shuttle @ Soto2 (Outside)
10. Hang out at Sam's food truck (~3 mins) (Outside)
11. Walk back to Soto1 (Outside)
12. Inside Soto1(Inside)

63
Table S2. Confusion matrix for optimal model (random forests algorithm with 31 features) for
S1.
Predicted
Truth Indoor In-transit Outdoor
Indoor (N=4,933) 4,904 (99.4%) 14 (0.3%) 15 (0.3%)
In-Transit (N=1,047) 21 (2.0%) 1016 (97.0%) 10 (1.0%)
Outdoor (N=410) 29 (7.1%) 16 (3.9%) 365 (89.0%)

64
Table S3. Model accuracy change by including different percentage of day 5 data under
conventional “naïve” CV.
Day 5
(N=4,629)
N BA (%) Hours
90% 4,165 94.1 12.6
80% 3,702 93.9 11.2
70% 3,239 93.8 9.8
60% 2,776 93.8 8.4
50% 2,314 93.7 7.0
40% 1,850 92.9 5.6
30% 1,387 92.8 4.2
20% 924 87.2 2.8
10% 461 85.9 1.4
5% 230 77.8 0.7 (about 42 minutes)

65
Table S4. Model accuracy change by including different percentage of day 5 script data under
conventional “naïve” CV.
Day 5 Scripted
(N=254)
N BA (%) Minutes
90% 227 91.7 38
80% 202 91.1 34
70% 176 90.4 29
60% 151 89.9 25
50% 126 84.8 21
40% 100 84.3 17
30% 75 54.8 13
20% 49 37.8 8
10% 24 33.3 4
5% 11 NA 2

66
Table S5. Model prediction accuracy in same subject and independent subject tests with variable
degrees of resemblance in time-activity patterns (unscripted natural living, scripted, or both
combined).
Model
#
Data
Used For
From
Subject
Natural living
(unscripted),
scripted, or
combined data

% indoor, outdoor and in
transit
microenvironments

BA
(%)
1^ Training S1 Combined 77:17:6
Testing S1 Combined 77:17:6 95.1%
2^ Training S1 Non-Scripted 80:16:4
Testing S1 Non-Scripted 80:16:4 95.0%
3^ Training S1 Scripted 26:16:58
Testing S1 Scripted 26:16:58 92.8%
5^ Training S2 Combined 87:7:6
Testing S2 Combined 87:7:6 93.6%
6^ Training S2 Non-Scripted 87:7:6
Testing S2 Non-Scripted 87:7:6 90.1%
7^ Training S2 Scripted 24:16:60
Testing S2 Scripted 24:16:60 93.2%
8^^ Training S1+S2 Combined 82:12:6
Testing S1+S2 Combined 82:12:6 93.5%
9^^ Training S1+S2 Non-Scripted 83:12:5
Testing S1+S2 Non-Scripted 83:12:5 93.3%
10^^ Training S1+S2 Scripted 25:17:58
Testing S1+S2 Scripted 25:17:58 94.5%
11* Training S2 Non-Scripted 87:7:6
Testing S1 Non-Scripted 80:16:4 48.8%
12* Training S2 Scripted 24:16:60
Testing S1 Scripted 26:16:58 67.3%
^Single person model: training and testing data come from the same person and the ratio of indoor/in-transit/outdoor was the
same in training and testing data. Training and testing data were selected by using stratified random selection with a 7:3 ratio.
^^Multiple persons model: training and testing data came from a combined dataset with subject 1 and subject 2, and the ratio of
indoor/in-transit/outdoor was the same in training and testing data. Training and testing data were selected by using stratified
random selection with a 7:3 ratio.
*Independent person testing model: subject 2 data were used for training and subject 1 data were used for testing.

67
Table S6. Model prediction accuracy in independent subject (S2) tests considering temporal
autocorrelation (day forward chain) compared to conventional “naïve” cross-validation (stratified
7:3 random selection).
Training/Testing
Data Selection
Method
Hyperparameter
Tuning method
Outer
Loop #
Training
Data (S1)
Validation
Data (S1)
Testing
Data
(S2)
BA
Day Forward Chain Nested cross-
validation
1 Day 1 Day2 S2 38.0%
2 Day 1, 2 Day3 S2 44.9%
3 Day 1, 2,
3
Day4 S2 48.6%
Stratified 7:3 random
selection
Conventional 10-
fold cross-validation
1 70% (Day 1, 2, 3) S2 33.3%
2 70% (Day 1, 2, 3, 4) S2 47.2%
3 70% (Day 1, 2, 3, 4, 5) S2 53.4%

68
Figure S1. Distribution of calculated speed versus sensor collected speed (m/s) (Pearson
correlation=0.76).

69
Figure S2. Temporal autocorrelation of sound level for S1 by date. The longest duration of time
for the temporal autocorrelation in sound level to decrease below 0 is about 2 hours.

70
Figure S3. Top 10 important features for overall microenvironment prediction ranked by drop-
column method from the optimal, final model in S1 (random forests algorithm with 31 features).

71
Figure S4. Partial dependence plots of top five ranked features in optimal model (S1) for overall model
prediction.

72
4 CHAPTER 4: ASSOCIATION OF PERSONAL AND AMBIENT AIR
POLLUTION AND LUNG FUNCTION DECREMENTS IN
ASTHMATIC CHILDREN IN LOG ANGELES
4.1 ABSTRACT
Background:
Exposure to air pollution is associated with pediatric asthma exacerbation including reduced lung
function. Advances in mobile health (mHealth) and wearable sensor technologies are enabling
the collection of highly space- and time-resolved exposure and outcome information to
understand these acute effects.
Methods:
We conducted a panel study of 40 children aged 8-16 with moderate to severe asthma, recruited
from the UCLA Pediatric Asthma Center of Excellence in 2019. We used the Biomedical REAI-
Time Health Evaluation (BREATHE) kit to continuously monitor personal exposure to
particulate matter of aerodynamic diameter < 2.5 µm (PM2.5), relative humidity and temperature,
geolocation (GPS), and asthma outcomes including lung function, medication use and symptoms
for 14 days. Hourly ambient (PM2.5, nitrogen dioxide (NO2), ozone (O3)) and traffic-related
(nitrogen oxides (NOx) and PM2.5) air pollution exposures were also modeled based on GPS
using spatial interpolation from ambient monitoring and the RLINE line source dispersion
model, respectively. We used mixed effects models to examine the within-person association of
these exposures with daily changes in % predicted forced expiratory volume in 1 sec (FEV1) and
peak expiratory flow (PEF).
Results:
73
Mean age of all children was 12.0 years (range: 8.4-16.8), average 67.9% predicted morning
FEV1 and 69.1% morning PEF, and the mean rescue inhaler daily use was 1.4 puffs. We found
prior 24-hour exposure to O3 was significantly associated with reduced morning (Estimate [95%
CI]: -4.11 [-6.86, -1.36]), evening (Estimate [95% CI]: -2.65 [-5.19, -0.10]) and daily FEV1
(Estimate [95% CI]: -3.45 [-6.42, -0.47]) and traffic-related PM2.5 was associated with reduced
morning PEF on day lag 0 (Estimate [95% CI]: -3.97 [-7.69, -0.26]) in models adjusted for
Hispanic ethnicity and personal relative humidity.
Conclusions:
Daily air pollution exposure was associated with significant lung function decrements in children
with moderate to severe asthma. Our study demonstrates the utility of the BREATH kit at
collecting highly resolved exposure and outcome data for investigating acute health effects.
4.2 INTRODUCTION
Asthma is a chronic disorder of both adults and children, which afflicts more than 25
million Americans accounting for 7.7 percent of adults and 8.4 percent of children
1
. Asthma
prevalence has been increasing over the last few decades in all age, sex, and racial groups, but
especially in children
2
. Children with asthma are significantly burdened by asthma morbidity and
bear higher rates of emergency department visits, hospitalizations, and even deaths
3,4
. A large
proportion of the burden of asthma is attributed to treating the consequences of poor asthma-
control
5
, so asthma control becomes an essential aspect of the evaluation and management of the
disease. Asthmatic children with poorly controlled asthma reported a decreased health-related
quality of life
6
. Therefore, according to the 2016 Global Initiative for Asthma (GINA)
guidelines
7
, once the treatment is established, achieving asthma control becomes the goal.
74
The mechanisms leading to asthma exacerbation in children are complex, but
environmental exposures are largely presumed to play a significant role
8
. Many previous studies
have found air pollution from ozone, nitrogen dioxide and particulate matter (PM) may induce or
aggravate asthma
9
. Exposure to air pollution has been shown to exacerbate children’s asthma,
including emergency department visits
10
, reduced percent predicted forced expiratory volume in
1 second (FEV1)
11
, rescue medication use
12
, and cough and wheeze symptoms
13
.
However, evidence on within-person acute effects (daily to sub-daily) of air pollution
exposure on asthma has lagged behind that for chronic effects, largely because of the challenges
involved in collecting concurrent exposure and outcome information with high spatiotemporal
resolution and over extended periods of time at a personal level. Given the growing advances of
mobile health (mHealth) approaches, including wearables, sensors, smartphone applications
(apps) and informatics, researchers are increasingly able to conduct these studies to further
understand acute effects of air pollution to pediatric asthma exacerbation. In this analysis, we
aimed to investigate the association of daily air pollution exposure with within-person risk of
asthma exacerbation (reduced lung function, daily count of rescue inhaler use, and whether or
not present asthma symptoms) in a panel study of children with moderate to severe asthma using
our recently developed BREATHE kit
14
.
4.3 MATERIALS AND METHODS
4.3.1 Study Design and Population
We recruited 40 children with moderate to severe asthma from the University of Los
Angeles Pediatric Asthma Center of Excellence clinics located in Los Angeles, CA and Santa
Monica, CA starting in February and until December 2019. Eligibility criteria included English-
75
speaking children aged 8-16 years, with doctor diagnosed asthma. We followed subjects over
two weeks. Participants were pre-screened for eligibility based on their medical records and
recruited by a dedicated study coordinator within the clinic during their pre-scheduled
appointments. Each child was given a Biomedical REAL-Time Health Evaluation (BREATHE)
kit which included an Android smartphone (Samsung S4) with a custom app to display sensor
data and deliver Ecological Momentary Assessment (EMA) surveys, a smartwatch (Moto 360
Smartwatch), and a minimum of three sensors (Airbeam personal PM2.5 sensor (HabitatMap),
handheld spirometer (Vitalograph asma-1 BT), and rescue and control medication inhaler sensors
(Propeller Health)). Children and their caretakers were trained on how to properly use and charge
the BREATHE kit and its components in the clinic, including how to initiate and perform proper
spirometry maneuvers, respond to smartphone surveys, charge devices and verify data
communications connectivity. The day following recruitment, a detailed baseline questionnaire
was conducted over the phone with the child and their caregiver to collect asthma-related health
and environmental data (typical activity patterns of the child, household operation conditions,
indoor source of air pollution, etc.). Over the 14 days of monitoring, data collection and
transmission status were continuously monitored by the research coordinator in a dedicated
researcher dashboard, and participants were contacted as necessary to help troubleshoot issues or
encourage compliance with data collection. Once the monitoring period was completed,
participants mailed their kits back in pre-labelled packages and complete an interviewer-
administered exit survey over the phone asking about their experience with the BREATHE Kit
and the study. Once the kit was returned, participants were mailed an Amazon gift card in
appreciation of their efforts. The institutional review board of the University of California, Los
76
Angeles approved the study protocol. Informed consent and assent were obtained from all
children and their primary legal guardians in the clinic upon recruitment.
4.3.2 Asthma Outcomes
4.3.2.1 Spirometry
Subjects conducted self-administered spirometry at home using the Vitalograph Asma-1
BT enabled spirometer. They were instructed to collect three “good” spirometry maneuvers in
the morning and in the evening every day, with up to six attempts per session, at the same time
every day. Morning and evening spirometry schedules were pre-decided in the clinic upon
recruitment based on typical wake and sleep times, within pre-defined windows. The handheld
device provided immediate feedback (beeping tone and visual symbol on display) to indicate to
participants whether a “good” maneuver was obtained, and this data was logged by the sensor
and transmitted along with the data in real time. Overall, 1,172 maneuvers were attempted by 40
participants and 887 (76%) were reported “good”. Of 218 person-days of spirometry data, 78
(36%) days achieved a minimum of six attempts, and among those days didn’t achieve the
minimum attempts, “good” attempts achieved a median of 3.
We compared the FEV1 (L/s) measurements obtained with the Asma-1 BT handheld
spirometer on recruitment day (following instruction by the research coordinator) with those
measured by the clinic using dedicated clinical grade spirometers (Morgan rolling seal
spirometer LT) on the same day or on the nearest date if same day measurement was unavailable,
as obtained from the medical record, as an indicator of data quality from the Breathe Kit sensor
on recruitment day.
The maximum (best effort) lung function from the two reproducible maneuvers were
selected for analysis. If the maximum value was not reproducible, the next largest value was
77
selected. We calculated predicted lung function based on age, gender, and height for FEV1

(forced expiratory volume in 1 second [L]) and PEF (peak expiratory flow rate [L/s]) based on
equations from Knudson et al
15
.
4.3.3 Exposure Assessment
Personal exposure to particulate matter with aerodynamic diameter < 2.5 µm (PM2.5),
relative humidity and temperature were continuously measured using the AirBeam 1.0
(HabitatMap). A running median filter (based on a window of 10 observations) was applied to
personal relative humidity to remove outliers. Because the degree of missingness in personal
relative humidity was high, we developed a model to impute the missing personal relative
humidity on the person-day level. Daily ambient relative humidity (calculated from ambient dew
point and ambient temperature
17
), daily ambient temperature and the visit-level difference
between mean ambient and personal relative humidity were included as predictors in a mixed
effects model with a random intercept for subject and a random slope for ambient relative
humidity to account for person-level clustering in the data and the likelihood that the relationship
between ambient and personal day-level relative humidity could vary by person depending on
their typical activity patterns or household characteristics, respectively. The Pearson correlation
between daily predicted and measured personal relative humidity was 0.97.
Ambient and traffic-related air pollution exposures were also modeled based on
geolocation (GPS), using inverse-distance squared interpolation from the nearest regulatory
monitors (for PM2.5, ozone (O3), nitrogen dioxide (NO2), nitrogen oxide (NO) and nitrogen
oxides (NOx)) and the RLINE line source dispersion model (for traffic-related PM2.5, NOx, and
NO2). The RLINE dispersion model used local weather data and a comprehensive database of
roadways, annual traffic volume, and vehicle emission factors for southern California to estimate
78
ambient pollutant concentrations contributed by on-road mobile source emissions at the
participant receptor points. Briefly, an Environmental Data Web Service was built for the
BREATHE Kit by Sonoma Technology Inc. (STI), to provide real-time and archived weather, air
quality, and traffic-related air pollution data streams to support the BREATHE platform based on
user location, date and hour. Exposures were estimated retrospectively following study
completion based on archive data for every hour with GPS data and averaged to daily level. In
addition, STI webservices provided ambient meteorological data on temperature and dew point.
Since we wanted to investigate the association of these exposures with morning, evening
and day-average outcomes, two averaging intervals were used to calculate 24-hour exposure
averages that precede the outcomes as follows: For morning outcomes such as morning lung
function, 24-hour averages were calculated starting from 6AM the previous day to 6AM of the
current day (or the morning when these outcomes are assessed). For evening (eg, evening lung
function) and daily outcomes, such as, daily average lung function, 24-hour averages were
calculated starting from 6PM the previous day to 6PM of the current day. The cut points of 6AM
and 6PM were selected because most subjects did their morning and evening lung function tests
after 6AM and 6PM, respectively (Supplement Figure 1). Moreover, 24-hour averages were
calculated using 30% completeness criteria, which is more relaxed than typical air pollution
investigations utilizing modeled ambient data given the greater chance of missing data using
personal sensors and real time data transmission.
4.3.4 Covariate Information
Based on a search of previous air pollution and asthma literature, we considered the
following covariates: sex, race, Hispanic ethnicity, caretaker’s education level, household
income, personal and ambient relative humidity and temperature, subject’s person-day level
79
time-activity patterns, medication use, outdoor physical activities, exposure to smoking
(exposure to secondhand smoking in the home and in-utero exposure to maternal smoking),
home characteristics (kitchen ventilation, fuel use, presence of pets) and day of the week.
However, based on the 10% principle confounder selection rule, only Hispanic ethnicity and
personal relative humidity were selected to be included in the model as confounders.
4.3.5 Statistical Analysis
Daily averaged air pollutants were not normally distributed, so Spearman correlations
were calculated to assess correlations between different pollutants.
We tested the association between daily air pollution exposures and lung function
(%predicted FEV1 and PEF) using mixed-effects models, with a random intercept for each
subject to account for the repeated measures design. We investigated these associations for the
preceding 24 hours (lag 0) as well as lags 1 and 2 days (defined as the average of the 25
th
to 48
th

hour (lag 1) and the average of the 49
th
to 72
nd
hour (lag 2)) for daily or evening outcomes (and
similarly for morning outcomes only aligned with the earlier defined exposure window).
We expressed results as change in % predicted lung function. All effects estimates were
scaled to a standard deviation (std) increase in each pollutant (based on lag 0 distributions) to
allow standardized inter-pollutant comparisons. Statistical significance was determined based on
a p-value < 0.05. We also added control medication use in the model for sensitivity analysis. All
Analyses were conducted in SAS 9.3 (SAS Institute Inc., Cary, NC).
80
4.4 RESULTS
4.4.1 Descriptive Summaries
We recruited 40 children with mean age of 12 years old, 45% female, and 47.5% non-Hispanic
ethnicity. Descriptive statistics for the 40 subjects in the present analysis are stated in Table 4.1.
Table 4.1. Study Group Characteristics.
Characteristics Statistics
Age [years, mean (range)] 12.0 (8.4-16.8)
Sex [no. (%)]

Female 18 (45.0)
Male 22 (55.0)
Race [no. (%)]

White 15 (37.5)
Black/African American 2 (5.0)
Black/Not African American 1 (2.5)
Asian 4 (10.0)
Other 15 (37.5)
Hispanic Ethnicity [no. (%)]

No 19 (47.5)
Yes 17 (42.5)
Medications take in the last 12 months

Short-acting beta2-agonist bronchodilators [no. (%)]

No 11 (27.5)
Regularly, every day 4 (10.0)
Occasionally, as needed 25 (62.5)
Long-acting beta2-agonist bronchodilators [no. (%)]

No 40 (100.0)
Corticosteroids [no. (%)]

No 28 (70.0)
Regularly, every day 12 (30.0)
Oral steroid medication [no. (%)]

No 25 (62.5)
Yes 15 (37.5)
Have allergies [no. (%)]

No 8 (20.0)
Yes 27 (67.5)
Any biological parents ever been diagnosed with asthma [no. (%)]

81
No 25 (62.5)
Yes 10 (25.0)
Caretaker’s highest completed educational grade [no. (%)]

High school or GED 3 (7.5)
Some college or trade school 9 (22.5)
College 9 (22.5)
Graduate school 15 (37.5)
Total household income per year [no. (%)]

Prefer not to say 12 (30)
$30,000-$40,000 2 (5.0)
Over $50,000 23 (57.5)
Anyone currently smoke cigarettes, or anything other than
cigarettes (e-cigarettes, cigars, pipes, hookah's, tobacco products,
other) at the home on a regular basis [no. (%)]

No 37 (92.5)
Child ever smoked electronic cigarettes/e-cigarettes or other
electronic nicotine devices (e-hookah, e-cigars, etc.) [no. (%)]

No 34 (85.0)
Mother have any problems during the pregnancy [no. (%)]

An infection 1 (2.5)
Early labor 7 (17.5)
High blood pressure 2 (5)
High blood sugar 6 (15)
None of these 18 (45)
Mother use during pregnancy [no. (%)]

Cigarettes 0 (0)
Drugs 1 (2.5)
None of these 36 (90)
Kitchen fan over the cooking stove, range, oven or elsewhere in
the kitchen [no. (%)]

No 5 (12.5)
Yes 32 (80.0)
Child own pets [no. (%)]

No 15 (37.5)
Yes 22 (55.0)
Any sports in a typical week [no. (%)]

No 7 (17.5)
Yes 29 (72.5)
Days play outdoors for at least half an hour in a typical week [no.
(%)]

No 6 (15.0)
82
Yes 31 (77.5)
Has health insurance [no. (%)]

Yes 40 (100.0)
Type of Health Insurance [no. (%)]

HMO 18 (45.0)
PPO or POS 20 (50.0)

We followed each subject for 2 weeks, except for 3 subjects who withdrew from the study earlier
and only completed 2, 3, and 5 days of follow-up. On average, FEV1 and PEF were lower in the
morning and increased in the evening (Table 4.2). The fitted regression line between the FEV1
(L/s) measurements obtained with the Asma-1 BT handheld spirometer on recruitment day and
measurements by the clinic using dedicated clinical grade spirometers (Morgan rolling seal
spirometer LT) had high R square (R
2
=0.68) (Figure 4.1).
Table 4.2. Distributions of Person-Day Level Lung Function Outcomes.
Respiratory Related Outcomes Mean±SD
Percent-predicted FEV1 (L/s)
Morning (N=175) 67.9±17.3
Evening (N=147) 70.9±17.7
Daily Average (N=96) 68.7±15.7
Percent-predicted PEF (L/s)
Morning (N=175) 69.1±18.4
Evening (N=147) 73.8±18.3
Daily Average (N=96) 69.3±15.8

For air pollutants and meteorology, measured personal exposures had less person-days of valid
observations as compared to modeled ambient and traffic-related exposures (Table 4.3) but
wider variability in general.

83
Figure 4.1. Comparison between Collected Forced Expiratory Volume in One Second (FEV1)
and Clinic Spirometer Tested FEV1.

Table 4.3. Descriptive Statistics of 24-Hour Average (6AM-6AM) Air Pollutant and
Meteorology Measurements.
Exposure N Mean Std Dev Min/Max
Pollutant (unit)
Personal Exposure (measured)
PM2.5 (µg/m
3
) 182 6.9 9.1 0.3/64.7
Temperature (̊C) 182 27.4 2.2 22.5/33.2
Relative Humidity (%) 182 45.9 8.1 15.3/61.1
Ambient Exposure (modeled)
PM2.5 (µg/m
3
) 359 9.5 3.7 1.0/21.1
NO2 (ppb) 356 7.6 4.6 0.27/35.9
NOx (ppb) 345 9.5 6.1 0.27/36.0
NO (ppb) 345 1.8 2.4 0/17.3
O3 (ppb) 359 32.3 9.2 7.7/64.7
Temperature (̊C) 342 18.0 4.0 8.2/32.2
Relative Humidity (%) 342 69.7 19.1 8.2/95.8
Traffic-related Exposure (modeled)
PM2.5 (µg/m
3
) 338 0.7 0.7 0.1/4.8
NO2 (ppb) 338 4.9 7.1 0/42.8
NOx (ppb) 338 10.6 14.7 0.1/88.9

84
Concentrations of personal PM2.5 were highly variable between subjects, with a maximum
reaching 64.7 µg/m
3
(Supplement Figure 2). Table 4 presents Spearman correlations among all
daily average pollutants. Moderate correlations between personal and ambient (Spearman  =
0.39), but weak correlations of personal PM2.5 with traffic-related PM2.5 (  = 0.14) were found.
Ambient NO2, NOx, and NO presented moderate to high correlations with each other, and traffic-
related NO2, NOx, and PM2.5 were also highly correlated with each other. Ambient O3 appeared
to have small correlations with ambient NO2, NOx, and NO, and traffic-related NO2, NOx, and
PM2.5 (Table 4.4).
Table 4.4. Exposure (24-Hour Average 6AM-6AM) Correlation Matrix.
Personal Ambient Ambient Ambient Ambient Ambient
Traffic-
related
Traffic-
related
Traffic-
related

PM2.5 NO2 NOx NO O3 PM2.5 NO2 NOx PM2.5
Personal PM2.5 1.00 0.05 -0.03 -0.03 -0.10 0.39* 0.15 0.14 0.14
Ambient NO2 1.00 0.98* 0.67* -0.40* 0.21* 0.19* 0.20* 0.24*
Ambient NOx 1.00 0.78* -0.37* 0.18* 0.16* 0.17* 0.21*
Ambient NO 1.00 -0.25* 0.24* -0.05 -0.03 -0.01
Ambient O3 1.00 -0.03 -0.22* -0.23* -0.24*
Ambient PM2.5 1.00 0.09 0.08 0.09
Traffic-related NO2 1.00 0.99* 0.98*
Traffic-related NOx 1.00 0.99*
Traffic-related PM2.5 1.00
*P<0.05 from Wald-based tests of Spearman correlation coefficient.
4.4.2 Exposure and Health Analyses
Results of mixed effects models investigating the association between air pollution
exposures and asthma outcomes are presented below. Table 4.5 presents results of FEV1 models.
We found significant inverse associations between ambient O3 (lag 0) and morning, evening and
daily average FEV1. Lag 1 ambient O3 was also significantly inversely associated with evening
and daily FEV1, but at lag 2 days, only the association with morning FEV1 remained. For traffic-
related pollutants (PM2.5, NOx, NO2), lag 1 exposure was significantly and inversely associated
85
with morning FEV1. Overall, most lag 0 air pollutant exposures were inversely associated with
morning FEV1 although some were not significant (Figure 4.2).
The associations between % predicted PEF and air pollutants are presented in Table 4.6.
Traffic-related PM2.5 showed significant inverse association with morning PEF at lag 0
(P=0.039), and marginal significance at lag 1 (P=0.067) and lag 2 (P=0.062). For the other two
traffic-related pollutants (NOx and NO2), significant inverse associations to morning PEF were
found at lag 1. Also, most air pollutants on lag 0 show inverse associations between morning
PEF although some were not significant (Figure 4.2).
For sensitivity analysis, we also adjusted control inhaler use in our health the model on
lag 0, and the results were consistent with the main results. Exposure to ambient O3 was
significantly associated with decreased morning FEV1 (Estimate [95% CI]: -4.73 [-7.98, -1.47])
(Supplement Table 1), and exposure to traffic-related PM2.5 (Estimate [95% CI]: -4.98 [-9.21, -
0.65]) and NOx (Estimate [95% CI]: -5.80 [-10.70, -0.91]) were significant associated with
averaged daily PEF (Supplement Table 2).

86

87

88

89

90

91

4.5 DISCUSSION
Our study suggested there were significant associations between personal, ambient, and
traffic related air pollutions and asthma exacerbations in asthmatic children aged 8-16 in Los
Angeles. We found that increased exposures to ambient O3 and traffic related PM2.5 was
associated with lung function deficits. The largest magnitude of association was a 4.11% drop in
precent-predicted morning FEV1 for a standard deviation increase of 9.2 ppb in 24-hour average
(lag 0) ambient O3, and a 3.97% drop in percent-predicted morning PEF for a standard deviation
increase of 0.7 µg/m
3
24-hour average (lag 0) traffic-related PM2.5.
Our findings with ambient O3 and FEV1 are consistent with several previous studies
which reported short-term exposure to ambient O3 was associated with lung function
decrements
15–18
. Ozone is a strong oxidant which is formed in the troposphere via chemical
reactions in the presence of precursor pollutants, like volatile organic compounds, and
sunshine
19
. O3 concentrations are generally higher outdoors compared with indoors (since ozone
reacts on indoor surfaces and gets removed), therefore, human exposures to ozone mainly take
place in the outdoor environment
20
. WHO
21
states that there is more consistent evidence on the
short-term rather than the long-term effects of O3, which include increases in daily mortality and
morbidity, especially for respiratory cause
22
. Because up to 90% of inhaled ozone is absorbed in
the respiratory tract along the bronchial tree
23
, O3 responses are likely initiated and localized in
the respiratory tract lining fluid due to the low solubility and high reactivity of O3
24
. When O3
reaches the lung surface, it reacts with proteins and lipids of the lung lining fluid. The secondary
products that are formed cause the inflammation and generation of cytokines, leading to an
increase in lung permeability and edema development
25
. Consequently, it causes oxidative stress
and lung inflammation, contributing to respiratory morbidities, such as reduced lung function
25
.
93
However, acute reactions to O3 exposure, such as decreases in lung function indices, display
largely unexplained heterogeneity between subjects likely due to differences in individual
susceptibility factors and co-exposures
26
.
Our finding of traffic-related pollutants (PM2.5 on lag 0, 1, 2 day; NOx and NO2 on lag 1
day) being associated with decreased morning PEF are in line with a 2001 study in the
Netherlands
27
where the authors reported significant decreases in PEF in children following
exposures to increased levels of PM10 (lag 1 and 3 days), NO2 (lag 0 and lag 1), and NO (lag 3
day). Moreover, Li et al
28
reviewed more than 30 panel studies on the effects of air pollution on
children’s lung function and respiratory symptoms. They reported that particles and NO2 showed
more significant results with PEF, but associations for many outcomes depended on the number
of lag days similarly to our results. Previous studies were trying to understand the biological
mechanism contributing to decreased lung function as a result of increased PM2.5 exposure.
Studies demonstrated that PM2.5 (fine PM) could reach alveoli, and up to 50% of them may
remain in the lung tissue
30
. Fine PM can penetrate deep enough into the airways and induce
alveolar inflammation, which then results in mediators for respiratory disease. Also, because of
their deep deposition in the alveoli, their removal rate is very slow thus increasing the chance of
causing cell damage
31
. These observed, significant lag effects also suggest that physicians
should consider warning their pediatric asthma patients and their parents to be more vigilant
about symptom monitoring not only on particular poor air quality days, but also for the following
1-2 days to adequately prepare for symptom flares.
We also found personal air pollution exposure showed higher correlation with ambient air
pollution exposure which suggested ambient air pollution contributed more to personal exposure
in our study population. This could be due to participant’s time-activity patterns and home
94
operation characteristics leading to infiltration of outdoor air pollution indoors. It is also possible
that study participants did not spend too much time in transit or commuting during the
monitoring period, which is not unexpected for a pediatric asthma population with moderate to
severe asthma.
The BREATHE mHealth platform design reduces recall bias and outcome
misclassification compared to previous studies using daily diary cards. BREATHE mHealth
platform allows multiple sensors to be connected in a single study and enables personal-level
data collection through EMAs as well as integration of additional external data (e.g.,
environmental, clinical and medical records) to paint a complete picture of each subject. Using
the BREATHE mHealth platform design, future studies could track a participant’s movements
throughout the day to assign accumulated exposures, improve classifications of time spend
indoors and outdoors, or investigate the within-person and within-day associations between
exposures, behaviors and health outcomes in context.
There were a number of limitations in this study. First, the personal PM2.5 values were
measured from the personal sensor and were not calibrated, and low-cost optical sensors as the
one used in this study tend to overestimate PM2.5 concentrations in general largely due to relative
humidity interferences. Whether or not the overestimation can have any effect on the health
model findings could be assessed in future analyses; however, it is unlikely to be differential in
relation to the outcome. Secondly, compliance was another issue in our study where subjects
sometimes forgot to charge their devices, forgot or were not able to take them with them to
school or did not do the spirometry tests in the pre-set time. Therefore, for future studies,
mHealth approaches may be applied to provide timely and proactive reminders for study subjects
with an emphasis on encouraging and monitoring compliance without biasing data collection.
95
Third, we tried to adjusted age in our model, but age was not a founder based on the 10% rule.
However, the random intercept in our mixed effect modeling approach probably captured some
of the effects of age as a person-level characteristic. Furthermore, given the rate of growth of
lungs between our study population ages (8-16 years old), the possibility remains that
variable lung growth rate based on age may be a factor in susceptibility that may be elucidated
with age stratification in future studies with larger sample size.
In addition, our study findings were limited in terms of their generalizability. Our
participants were recruited from a pediatric asthma specialty clinic and had moderate to severe
asthma. As such, triggers and factors associated with exacerbations in this population might not
necessarily translate to children with mild asthma or well-controlled asthma. In addition,
socioeconomic characteristics of our study sample are likely not representative of the general
population, where our participants had higher parental education level (37.5% of their caretaker
had graduate level degree, higher household income level (57.5% of their caretaker reported over
$50,000 household income annually), and higher health insurance coverage (100% subjects had
health insurance). Finally, selection bias may have influenced our results if parents living in
areas with higher air pollution exposure were also more concerned about its impact on their
child’s asthma and therefore more likely to participate in our study. However, it is difficult to
ascertain the direction of this potential selection bias.
In conclusion, we found significant and negative associations between increased O3 and
traffic-related PM2.5 exposures and decreased lung function in asthmatic children. The
longitudinal nature of our study and the acute outcome measurement through the BREATHE
mHealth platform provides further support for the plausibility of the associations of exposure to
air pollutants. Understanding the effect of ambient O3 and traffic-related PM2.5 on lung function
96
is increasingly important as climate change models predict a rise in O3 concentrations in the
future
40
. Our study demonstrates the utility of the BREATH kit at collecting highly resolved
exposure and outcome data for investigating acute health effects.
4.6 REFERENCES:
1. Centers for Disease Control and Prevention. CDC - Asthma - Most Recent Asthma Data.
Most Recent Asthma Data.
2. Kuschner WG. The asthma epidemic [12]. N Engl J Med. 2007.
doi:10.1056/NEJMc063596
3. van Dellen QM, Stronks K, Bindels PJE, Öry FG, Bruil J, van Aalderen WMC. Predictors
of asthma control in children from different ethnic origins living in Amsterdam. Respir
Med. 2007. doi:10.1016/j.rmed.2006.08.002
4. Akinbami L. The state of childhood asthma, United States, 1980-2005. Adv Data. 2006.
5. Barnes PJ, Jonsson B, Klim JB. The costs of asthma. Eur Respir J. 1996.
doi:10.1183/09031936.96.09040636
6. Juniper EF, Guyatt GH, Feeny DH, Ferrie PJ, Griffith LE, Townsend M. Measuring
quality of life in children with asthma. Qual Life Res. 1996. doi:10.1007/BF00435967
7. GINA C científico de l. Global Strategy for Asthma Management and Prevention ( 2016
update ). Guía Gina Actual. 2016.
8. Etzel RA. How environmental exposures influence the development and exacerbation of
asthma. In: Pediatrics. ; 2003.
9. Spann K, Snape N, Baturcam E, Fantino E. The Impact of Early-Life Exposure to Air-
borne Environmental Insults on the Function of the Airway Epithelium in Asthma. Ann
Glob Heal. 2016. doi:10.1016/j.aogh.2016.01.007
97
10. Bouazza N, Foissac F, Urien S, et al. Fine particulate pollution and asthma exacerbations.
Arch Dis Child. 2018. doi:10.1136/archdischild-2017-312826
11. Delfino RJ, Quintana PJE, Floro J, et al. Association of FEV1 in asthmatic children with
personal and microenvironmental exposure to airborne particulate matter. Environ Health
Perspect. 2004. doi:10.1289/ehp.6815
12. Williams AM, Phaneuf DJ, Barrett MA, Su JG. Short-term impact of PM 2.5 on
contemporaneous asthma medication use: Behavior and the value of pollution reductions.
Proc Natl Acad Sci U S A. 2019. doi:10.1073/pnas.1805647115
13. Habre R, Moshier E, Castro W, et al. The effects of PM2.5 and its components from
indoor and outdoor sources on cough and wheeze symptoms in asthmatic children. J Expo
Sci Environ Epidemiol. 2014. doi:10.1038/jes.2014.21
14. Bui AAT, Hosseini A, Rocchio R, et al. Biomedical REAl-Time Health Evaluation
(BREATHE): toward an mHealth informatics platform. JAMIA Open. 2020.
doi:10.1093/jamiaopen/ooaa011
15. Knudson RJ, Lebowitz MD, Holberg CJ, Burrows B. Changes in the normal maximal
expiratory flow-volume curve with growth and aging. Am Rev Respir Dis. 1983.
doi:10.1164/arrd.1983.127.6.725
16. Alduchov OA, Eskridge RE. Improved Magnus form approximation of saturation vapor
pressure. J Appl Meteorol. 1996. doi:10.1175/1520-
0450(1996)035<0601:IMFAOS>2.0.CO;2
17. Samoli E, Dimakopoulou K, Evangelopoulos D, et al. Is daily exposure to ozone
associated with respiratory morbidity and lung function in a representative sample of
schoolchildren? Results from a panel study in Greece. J Expo Sci Environ Epidemiol.
98
2017. doi:10.1038/jes.2016.32
18. Angelis N, Spyratos D, Domvri K, et al. Effect of Ambient Ozone Exposure Assessed by
Individual Monitors on Nasal Function and Exhaled NO Among School Children in the
Area of Thessaloniki, Greece. J Occup Environ Med. 2017.
doi:10.1097/JOM.0000000000001011
19. Kopp M V., Bohnet W, Frischer T, et al. Effects of ambient ozone on lung function in
children over a two-summer period. Eur Respir J. 2000.
doi:10.1183/09031936.00.16589300
20. Chang YK, Wu CC, Lee LT, Lin RS, Yu YH, Chen YC. The short-term effects of air
pollution on adolescent lung function in Taiwan. Chemosphere. 2012.
doi:10.1016/j.chemosphere.2011.11.048
21. Tager IB, Balmes J, Lurmann F, Ngo L, Alcorn S, Künzli N. Chronic exposure to ambient
ozone and lung function in young adults. Epidemiology. 2005.
doi:10.1097/01.ede.0000183166.68809.b0
22. Geyh AS, Xue J, Özkaynak H, Spengler JD. The Harvard Southern California Chronic
Ozone Exposure Study: Assessing ozone exposure of grade-school-age children in two
southern California communities. Environ Health Perspect. 2000.
doi:10.1289/ehp.00108265
23. World Health Organization. Review of Evidence on Health Aspects of Air Pollution –
REVIHAAP Project: Technical Report.; 2013. doi:10.1007/BF00379640
24. Bell ML, Zanobetti A, Dominici F. Who is more affected by ozone pollution? A
systematic review and meta-analysis. Am J Epidemiol. 2014. doi:10.1093/aje/kwu115
25. Chen C, Arjomandi M, Tager IB, Holland N, Balmes JR. Effects of antioxidant enzyme
99
polymorphisms on ozone-induced lung function changes. Eur Respir J. 2007.
doi:10.1183/09031936.00160806
26. Kelly FJ. Dietary antioxidants and environmental stress. Proc Nutr Soc. 2004.
doi:10.1079/pns2004388
27. Mudway IS, Kelly FJ. Ozone and the lung: A sensitive issue. Mol Aspects Med. 2000.
doi:10.1016/S0098-2997(00)00003-0
28. Que LG, Stiles J V., Sundy JS, Foster WM. Pulmonary function, bronchial reactivity, and
epithelial permeability are response phenotypes to ozone and develop differentially in
healthy humans. J Appl Physiol. 2011. doi:10.1152/japplphysiol.00337.2011
29. Steerenberg PA, Nierkens S, Fischer PH, et al. Traffic-related air pollution affects peak
expiratory flow, exhaled nitric oxide, and inflammatory nasal markers. Arch Environ
Health. 2001. doi:10.1080/00039890109604069
30. Li S, Williams G, Jalaludin B, Baker P. Panel studies of air pollution on children’s lung
function and respiratory symptoms: A literature review. J Asthma. 2012.
doi:10.3109/02770903.2012.724129
31. Brown JS, Gordon T, Price O, Asgharian B. Thoracic and respirable particle definitions
for human health risk assessment. Part Fibre Toxicol. 2013. doi:10.1186/1743-8977-10-
12
32. Valavanidis A, Fiotakis K, Vlachogianni T. Airborne particulate matter and human health:
Toxicological assessment and importance of size and composition of particles for
oxidative damage and carcinogenic mechanisms. J Environ Sci Heal - Part C Environ
Carcinog Ecotoxicol Rev. 2008. doi:10.1080/10590500802494538
33. See SW, Balasubramanian R. Chemical characteristics of fine particles emitted from
100
different gas cooking methods. Atmos Environ. 2008. doi:10.1016/j.atmosenv.2008.09.011
34. Sheffield PE, Knowlton K, Carr JL, Kinney PL. Modeling of regional climate change
effects on ground-level ozone and childhood asthma. Am J Prev Med. 2011.
4.7 SUPPLEMENT

102

103
Supplement Figure 1. Morning and Evening Lung Function Spirometry Tests Distributions by
Time.
A

B

104
Supplement Figure 2. Personal PM2.5 Daily Average Distribution across All Subjects.
Personal

Ambient

Traffic

105
5 CHAPTER 5. ASSOCIATION OF PERSONAL AND AMBIENT AIR
POLLUTION AND ASTHMA EXACERBATION IN ASTHMATIC CHILDREN
IN LOS ANGELES
5.1 ABSTRACT
Background:
Exposure to air pollution is associated with pediatric asthma exacerbation including
increased rescue medication usage and worsened symptoms. Advances in mobile health
(mHealth) and wearable sensor technologies are enabling the collection of highly space- and
time-resolved exposure and outcome information to understand these acute effects.
Methods:
We conducted a panel study of 40 children aged 8-16 with moderate to severe asthma,
recruited from the UCLA Pediatric Asthma Center of Excellence in 2019. We used the
Biomedical REAI-Time Health Evaluation (BREATHE) kit to continuously monitor personal
exposure to particulate matter of aerodynamic diameter < 2.5 µm (PM2.5), relative humidity and
temperature, geolocation (GPS), and asthma outcomes including lung function, medication use
and symptoms for 14 days. Hourly ambient (PM2.5, nitrogen dioxide (NO2), ozone (O3)) and
traffic-related (nitrogen oxides (NOx) and PM2.5) air pollution exposures were also modeled
based on GPS using spatial interpolation from ambient monitoring and the RLINE line source
dispersion model, respectively. We used mixed effects models to examine the within-person
association of these exposures with daily count of rescue inhaler use and whether or not
presented asthma symptoms (e.g. chest tightness, wheeze, trouble breathing, cough, etc.).
Results:
106
Mean age of all children was 12.0 years (range: 8.4-16.8), and the mean rescue inhaler
daily use was 1.4 puffs. Among all person-days of data, 26.4% had chest tightness, 17.5% had
wheeze, 32.3% had trouble breathing, and 42.9% had cough. We found ambient O3 (rate ratio
[95% CI]: 1.52 [1.02, 2.27]), NOx (rate ratio [95% CI]: 1.61 [1.23, 2.11]), and NO (rate ratio
[95% CI]: 1.80 [1.37, 2.35]) were significantly associated with increased rate ratio of daily count
of rescue inhaler use. Increased Traffic-related PM2.5 on day lag 0 was significantly associated
with increased odds of present trouble breathing symptoms (odds ratio [95% CI]: 1.83 [1.03,
3.24]).
Conclusions:
Daily air pollution exposure was associated with acute risk of asthma exacerbation in
children with moderate to severe asthma. Our study demonstrates the utility of the BREATH kit
at collecting highly resolved exposure and outcome data for investigating acute health effects.
5.2 INTRODUCTION
Asthma is a chronic disorder of both adults and children, which afflicts more than 25
million Americans accounting for 7.7 percent of adults and 8.4 percent of children
1
. Asthma
prevalence has been increasing over the last few decades in all age, sex, and racial groups, but
especially in children
2
. Children with asthma are significantly burdened by asthma morbidity and
bear higher rates of emergency department visits, hospitalizations, and even deaths
3,4
. A large
proportion of the burden of asthma is attributed to treating the consequences of poor asthma-
control
5
, so asthma control becomes an essential aspect of the evaluation and management of the
disease. Asthmatic children with poorly controlled asthma reported a decreased health-related
quality of life
6
. Therefore, according to the 2016 Global Initiative for Asthma (GINA)
guidelines
7
, once the treatment is established, achieving asthma control becomes the goal.
107
The mechanisms leading to asthma exacerbation in children are complex, but
environmental exposures are largely presumed to play a significant role
8
. Many previous studies
have found air pollution from ozone, nitrogen dioxide and particulate matter (PM) may induce or
aggravate asthma
9
. Exposure to air pollution has been shown to exacerbate children’s asthma,
including emergency department visits
10
, rescue medication use
12
, and cough and wheeze
symptoms
13
.
However, evidence on within-person acute effects (daily to sub-daily) of air pollution
exposure on asthma has lagged behind that for chronic effects, largely because of the challenges
involved in collecting concurrent exposure and outcome information with high spatiotemporal
resolution and over extended periods of time at a personal level. Given the growing advances of
mobile health (mHealth) approaches, including wearables, sensors, smartphone applications
(apps) and informatics, researchers are increasingly able to conduct these studies to further
understand acute effects of air pollution to pediatric asthma exacerbation. In this analysis, we
aimed to investigate the association of daily air pollution exposure with within-person risk of
asthma exacerbation (reduced lung function, daily count of rescue inhaler use, and whether or
not present asthma symptoms) in a panel study of children with moderate to severe asthma using
our recently developed BREATHE kit
14
.
5.3 MATERIALS AND METHODS
5.3.1 Study Design and Population
We recruited 40 children with moderate to severe asthma from the University of Los
Angeles Pediatric Asthma Center of Excellence clinics located in Los Angeles, CA and Santa
Monica, CA starting in February and until December 2019. Eligibility criteria included English-
108
speaking children aged 8-16 years, with doctor diagnosed asthma. We followed subjects over
two weeks. Participants were pre-screened for eligibility based on their medical records and
recruited by a dedicated study coordinator within the clinic during their pre-scheduled
appointments. Each child was given a Biomedical REAL-Time Health Evaluation (BREATHE)
kit which included an Android smartphone (Samsung S4) with a custom app to display sensor
data and deliver Ecological Momentary Assessment (EMA) surveys, a smartwatch (Moto 360
Smartwatch), and a minimum of three sensors (Airbeam personal PM2.5 sensor (HabitatMap),
handheld spirometer (Vitalograph asma-1 BT), and rescue and control medication inhaler sensors
(Propeller Health)). Children and their caretakers were trained on how to properly use and charge
the BREATHE kit and its components in the clinic, including how to initiate and perform proper
spirometry maneuvers, respond to smartphone surveys, charge devices and verify data
communications connectivity. The day following recruitment, a detailed baseline questionnaire
was conducted over the phone with the child and their caregiver to collect asthma-related health
and environmental data (typical activity patterns of the child, household operation conditions,
indoor source of air pollution, etc.). Over the 14 days of monitoring, data collection and
transmission status were continuously monitored by the research coordinator in a dedicated
researcher dashboard, and participants were contacted as necessary to help troubleshoot issues or
encourage compliance with data collection. Once the monitoring period was completed,
participants mailed their kits back in pre-labelled packages and complete an interviewer-
administered exit survey over the phone asking about their experience with the BREATHE Kit
and the study. Once the kit was returned, participants were mailed an Amazon gift card in
appreciation of their efforts. The institutional review board of the University of California, Los
109
Angeles approved the study protocol. Informed consent and assent were obtained from all
children and their primary legal guardians in the clinic upon recruitment.
5.3.2 Asthma Outcomes
5.3.2.1 Inhaler Medication
Propeller Health Bluetooth-enabled sensors were provided to participants based on
compatibility of these sensors with participants’ rescue and controller medications. We aimed to
provide as many sensors to each child as they typically use (rescue or control); however, sensor
incompatibility and inability of participants to properly set up the sensors they were given by
themselves with verbal instruction over the phone (if they did not bring their complete
medication with them to the clinic) sometimes led to our inability to deploy sensors for all
medication types per participant. Inhaler sensors logged every puff of medications dispensed by
type and transmitted this data with geographic coordinates and a timestamp to the BREATHE
Kit servers. Medication use was then summarized on a person-day level for rescue (count,
modeled as outcome) and control inhaler use (binary for any use, adjusted for as potential
confounder).
5.3.2.2 Asthma Symptoms
Asthma symptoms were collected via ecological momentary assessment (EMA) surveys
deployed on a custom app with the BREATHE Kit using three types of surveys to minimize
recall bias, maximize validity, and capture participants’ symptoms in natural settings. The
following questions based on the Asthma Control Test (ACT)
14
were asked but rephrased to refer
to either the past hour, or the previous night/day/school time.
• Morning surveys: EMAs were sent out at the time of wake up, and those questions were trying to
collect asthma symptoms happened last night. The collected questions were “During last night,
110
did you wake up because of your asthma?” (response choices: Yes/No) and “During last night,
how many times did you use your inhaler during last night?” (response choices: Never/One
time/Two times/Three times/Four or more times. We dichotomized them into Never/One or more
times).
• Random surveys: EMAs were sent out within a predesignated window of time to study behaviors
throughout the day and this type of survey can be triggered by other sensors, for example,
AirBeam Personal PM 2.5 sensor, etc. This survey was trying to collect asthma symptoms
happened in the last hour. The collected questions were “In the past hour, did your chest feel tight
because of asthma?”, “In the past hour, did you feel wheezy because of asthma?”, “In the past
hour, did you feel trouble breathing because of your asthma?”, “In the past hour, did you cough
because of your asthma?”, “In the past hour, how much of a problem was your asthma when you
ran, exercise or play sports?”, “In the past hour, did you feel scared that you might have trouble
breathing because of your asthma?”, and “In the past hour, have you avoided strenuous activities,
or had slow down or stop exercising because of your asthma?”. The response choices for all
above questions except for the question “In the past hour, how much of a problem was your
asthma when you ran, exercise or play sports?” were Not at all/A little/Quite a bit/Very much so,
and we dichotomized them into Not at all and A little/Quite a bit/Very much so. For the “In the
past hour, how much of a problem was your asthma when you ran, exercise or play sports?”, one
more response choice of “I didn’t run, exercise or play sports” was given, and this choice was
treated as missing in the analysis.
• Last random survey of the day: EMAs were sent out at the end of the day, and questions were
trying to collect day level asthma information. The collected questions were “How much of the
time did your asthma keep you from getting as much done at school or at home today?”. The
response choices for the above questions were Not at all/A little/Quite a bit/Very much so, and
we dichotomized them into Not at all and A little/Quite a bit/Very much so.
111
5.3.3 Exposure Assessment
Personal exposure to particulate matter with aerodynamic diameter < 2.5 µm (PM2.5),
relative humidity and temperature were continuously measured using the AirBeam 1.0
(HabitatMap). A running median filter (based on a window of 10 observations) was applied to
personal relative humidity to remove outliers. Because the degree of missingness in personal
relative humidity was high, we developed a model to impute the missing personal relative
humidity on the person-day level. Daily ambient relative humidity (calculated from ambient dew
point and ambient temperature
15
), daily ambient temperature and the visit-level difference
between mean ambient and personal relative humidity were included as predictors in a mixed
effects model with a random intercept for subject and a random slope for ambient relative
humidity to account for person-level clustering in the data and the likelihood that the relationship
between ambient and personal day-level relative humidity could vary by person depending on
their typical activity patterns or household characteristics, respectively. The Pearson correlation
between daily predicted and measured personal relative humidity was 0.97.
Ambient and traffic-related air pollution exposures were also modeled based on
geolocation (GPS), using inverse-distance squared interpolation from the nearest regulatory
monitors (for PM2.5, ozone (O3), nitrogen dioxide (NO2), nitrogen oxide (NO) and nitrogen
oxides (NOx)) and the RLINE line source dispersion model (for traffic-related PM2.5, NOx, and
NO2). The RLINE dispersion model used local weather data and a comprehensive database of
roadways, annual traffic volume, and vehicle emission factors for southern California to estimate
ambient pollutant concentrations contributed by on-road mobile source emissions at the
participant receptor points. Briefly, an Environmental Data Web Service was built for the
BREATHE Kit by Sonoma Technology Inc. (STI), to provide real-time and archived weather, air
quality, and traffic-related air pollution data streams to support the BREATHE platform based on
112
user location, date and hour. Exposures were estimated retrospectively following study
completion based on archive data. In addition, STI webservices provided ambient meteorological
data on temperature and dew point.
Since we wanted to investigate the association of these exposures with morning, evening
and day-average outcomes, two averaging intervals were used to calculate 24-hour exposure
averages that precede the outcomes as follows: For morning outcomes such as morning lung
function, 24-hour averages were calculated starting from 6AM the previous day to 6AM of the
current day (or the morning when these outcomes are assessed). For evening (e.g., evening lung
function) and daily outcomes, such as, daily average lung function, 24-hour averages were
calculated starting from 6PM the previous day to 6PM of the current day. The cut points of 6AM
and 6PM were selected because most subjects did their morning and evening lung function tests
after 6AM and 6PM, respectively (Supplement Figure 1). Moreover, 24-hour averages were
calculated using 30% completeness criteria, which is more relaxed than typical air pollution
investigations utilizing modeled ambient data given the greater chance of missing data using
personal sensors and real time data transmission.
5.3.4 Covariate Information
Based on a search of previous air pollution and asthma literature, we considered the
following covariates: sex, race, Hispanic ethnicity, caretaker’s education level, household
income, personal and ambient relative humidity and temperature, subject’s person-day level
time-activity patterns, medication use, outdoor physical activities, exposure to smoking
(exposure to secondhand smoking in the home and in-utero exposure to maternal smoking),
home characteristics (kitchen ventilation, fuel use, presence of pets) and day of the week.
113
However, based on the 10% principle confounder selection rule, only Hispanic ethnicity and
personal relative humidity were selected to be included in the model as confounders.
5.3.5 Statistical Analysis
Daily averaged air pollutants were not normally distributed, so Spearman correlation was
utilized in order to assess the correlations between different pollutants.
We tested the association between daily air pollution exposures and daily count of rescue
inhaler medication use, and whether or not presented asthma-related symptoms (e.g. cough,
wheeze, trouble breathing, etc.) using mixed-effects models, with a random intercept for each
subject to account for the repeated measures design. We investigated these associations for the
preceding 24 hours (lag 0) as well as lags 1 and 2 days (defined as the average of the 25
th
to 48
th

hour (lag 1) and the average of the 49
th
to 72
nd
hour (lag 2)) for daily or evening outcomes (and
similarly for morning outcomes only aligned with the earlier defined exposure window).
We expressed results as change in rate ratio for rescue inhaler use and odds of
experiencing a specific symptom within a day. All effects estimates were scaled to a standard
deviation (std) increase in each pollutant (based on lag 0 distributions) to allow standardized
inter-pollutant comparisons. We also tested two-pollutant regression models for rescue inhaler
use to assess between-pollutant confounding. The aim here was to assess the extent to which
associations with one pollutant were confounded independent of another pollutant.
Statistical significance was determined based on a p-value < 0.05. We also added control
medication use in the model for sensitivity analysis. All Analyses were conducted in SAS 9.3
(SAS Institute Inc., Cary, NC).
114
5.4 RESULTS
5.4.1 Descriptive Summaries
We recruited 40 children with mean age of 12 years old, 45% female, and 47.5% non-
Hispanic ethnicity. Descriptive statistics for the 40 subjects in the present analysis are stated in
Table 5.1. We followed each subject for 2 weeks, except for 3 subjects who withdrew from the
study earlier and only completed 2, 3, and 5 days of follow-up.
Table 5.1. Study Group Characteristics.
Characteristics Statistics
Age [years, mean (range)] 12.0 (8.4-16.8)
Sex [no. (%)]

Female 18 (45.0)
Male 22 (55.0)
Race [no. (%)]

White 15 (37.5)
Black/African American 2 (5.0)
Black/Not African American 1 (2.5)
Asian 4 (10.0)
Other 15 (37.5)
Hispanic Ethnicity [no. (%)]

No 19 (47.5)
Yes 17 (42.5)
Medications take in the last 12 months

Short-acting beta2-agonist bronchodilators [no. (%)]

No 11 (27.5)
Regularly, every day 4 (10.0)
Occasionally, as needed 25 (62.5)
Long-acting beta2-agonist bronchodilators [no. (%)]

No 40 (100.0)
Corticosteroids [no. (%)]

No 28 (70.0)
Regularly, every day 12 (30.0)
Oral steroid medication [no. (%)]

No 25 (62.5)
Yes 15 (37.5)
Have allergies [no. (%)]

115
No 8 (20.0)
Yes 27 (67.5)
Any biological parents ever been diagnosed with asthma [no. (%)]

No 25 (62.5)
Yes 10 (25.0)
Caretaker’s highest completed educational grade [no. (%)]

High school or GED 3 (7.5)
Some college or trade school 9 (22.5)
College 9 (22.5)
Graduate school 15 (37.5)
Total household income per year [no. (%)]

Prefer not to say 12 (30)
$30,000-$40,000 2 (5.0)
Over $50,000 23 (57.5)
Anyone currently smoke cigarettes, or anything other than
cigarettes (e-cigarettes, cigars, pipes, hookah's, tobacco products,
other) at the home on a regular basis [no. (%)]

No 37 (92.5)
Child ever smoked electronic cigarettes/e-cigarettes or other
electronic nicotine devices (e-hookah, e-cigars, etc.) [no. (%)]

No 34 (85.0)
Mother have any problems during the pregnancy [no. (%)]

An infection 1 (2.5)
Early labor 7 (17.5)
High blood pressure 2 (5)
High blood sugar 6 (15)
None of these 18 (45)
Mother use during pregnancy [no. (%)]

Cigarettes 0 (0)
Drugs 1 (2.5)
None of these 36 (90)
Kitchen fan over the cooking stove, range, oven or elsewhere in
the kitchen [no. (%)]

No 5 (12.5)
Yes 32 (80.0)
Child own pets [no. (%)]

No 15 (37.5)
Yes 22 (55.0)
Any sports in a typical week [no. (%)]

No 7 (17.5)
Yes 29 (72.5)
116
Days play outdoors for at least half an hour in a typical week [no.
(%)]

No 6 (15.0)
Yes 31 (77.5)
Has health insurance [no. (%)]

Yes 40 (100.0)
Type of Health Insurance [no. (%)]

HMO 18 (45.0)
PPO or POS 20 (50.0)

For asthma related symptoms, in general, subjects answered more random survey
questions (for example, in the past hour, did you have trouble breathing because of your
asthma?) compared to the morning survey questions (for example, how was your asthma during
last night?) and evening survey questions (for example, Did you have asthma attack today?). On
average, subjects used 1.4 puffs rescue inhaler per day, and there was a big range of daily rescue
inhaler use among all subjects (range: 0-24 puffs per day) (Table 5.2 and Supplement Figure
2).
Table 5.2. Distributions of Person-Day Level Health Outcomes.
Asthma Symptoms Related Outcomes N (%)
Did you wake up last night because of your
asthma?

No 123 (93.2)
Yes 9 (6.8)
How many times did you use your inhaler during
the night?

Never 111 (84.1)
One or more times 21 (15.9)
How much of the time did your asthma keep you
from getting as much done at school or at home
today?

Not at all 94 (86.2)
A little/Quite a bit/Very much so 15 (13.8)
Did your chest feel tight because of asthma today?
Not at all 138 (63.6)
A little/Quite a bit/Very much so 79 (26.4)
117
Did you feel wheezy because of your asthma
today?

Not at all 179 (82.5)
A little/Quite a bit/Very much so 38 (17.5)
Did you have trouble breathing because of your
asthma today?

Not at all 147 (67.7)
A little/Quite a bit/Very much so 70 (32.3)
Did you cough because of your asthma today?
Not at all 124 (57.1)
A little/Quite a bit/Very much so 93 (42.9)
How much of a problem was your asthma when
you ran, exercise or play sports today?

Not at all 79 (71.8)
A little/Quite a bit/Very much so 31 (28.2)
In the past hour, did you feel scared that you might
have trouble breathing because of your asthma?

Not at all 152 (81.7)
A little/Quite a bit/Very much so 34 (18.3)
In the past hour, have you avoided strenuous
activities, or had to slow down or stop exercising
because of your asthma?

Not at all 155 (83.3)
A little/Quite a bit/Very much so 31 (16.7)
Inhaler Medication Use Sensor Measurements Mean±SD
Number of rescue inhaler puffs per day (N=324) 1.4±3.5
Number of control inhaler puffs per day (N=312) 1.5±1.9

For air pollutants and meteorology, measured personal exposures had less person-days of
valid observations as compared to modeled ambient and traffic-related exposures (Table 5.3) but
wider variability in general.
Table 5.3. Descriptive Statistics of 24-Hour Average (6AM-6AM) Air Pollutant and
Meteorology Measurements.
Exposure N Mean Std Dev Min/Max
Pollutant (unit)
Personal Exposure (measured)
PM2.5 (µg/m
3
) 182 6.9 9.1 0.3/64.7
Temperature (̊C) 182 27.4 2.2 22.5/33.2
Relative Humidity (%) 182 45.9 8.1 15.3/61.1
Ambient Exposure (modeled)
PM2.5 (µg/m
3
) 359 9.5 3.7 1.0/21.1
NO2 (ppb) 356 7.6 4.6 0.27/35.9
118
NOx (ppb) 345 9.5 6.1 0.27/36.0
NO (ppb) 345 1.8 2.4 0/17.3
O3 (ppb) 359 32.3 9.2 7.7/64.7
Temperature (̊C) 342 18.0 4.0 8.2/32.2
Relative Humidity (%) 342 69.7 19.1 8.2/95.8
Traffic-related Exposure (modeled)
PM2.5 (µg/m
3
) 338 0.7 0.7 0.1/4.8
NO2 (ppb) 338 4.9 7.1 0/42.8
NOx (ppb) 338 10.6 14.7 0.1/88.9

Concentrations of personal PM2.5 were highly variable between subjects, with a maximum
reaching 64.7 µg/m
3
(Supplement Figure 3). Table 4 presents Spearman correlations among all
daily average pollutants. Moderate correlations between personal and ambient (Spearman  =
0.39), but weak correlations of personal PM2.5 with traffic-related PM2.5 (  = 0.14) were found.
Ambient NO2, NOx, and NO presented moderate to high correlations with each other, and traffic-
related NO2, NOx, and PM2.5 were also highly correlated with each other. Ambient O3 appeared
to have small correlations with ambient NO2, NOx, and NO, and traffic-related NO2, NOx, and
PM2.5 (Table 5.4).
Table 5.4. Exposure (24-Hour Average 6AM-6AM) Correlation Matrix.
Personal Ambient Ambient Ambient Ambient Ambient
Traffic-
related
Traffic-
related
Traffic-
related

PM2.5 NO2 NOx NO O3 PM2.5 NO2 NOx PM2.5
Personal PM2.5 1.00 0.05 -0.03 -0.03 -0.10 0.39* 0.15 0.14 0.14
Ambient NO2 1.00 0.98* 0.67* -0.40* 0.21* 0.19* 0.20* 0.24*
Ambient NOx 1.00 0.78* -0.37* 0.18* 0.16* 0.17* 0.21*
Ambient NO 1.00 -0.25* 0.24* -0.05 -0.03 -0.01
Ambient O3 1.00 -0.03 -0.22* -0.23* -0.24*
Ambient PM2.5 1.00 0.09 0.08 0.09
Traffic-related NO2 1.00 0.99* 0.98*
Traffic-related NOx 1.00 0.99*
Traffic-related PM2.5 1.00
*P<0.05 from Wald-based tests of Spearman correlation coefficient.
119
5.4.2 Exposure and Health Analyses
Results of mixed effects models investigating the association between air pollution
exposures and asthma outcomes are presented below. Table 5.5 shows the associations between
daily count of rescue inhaler use and air pollutants. Most significant associations were found on
lag 0 day. Ambient air pollutants (PM2.5, O3, NOx, NO, NO2) on lag 0 day were all positively
associated with daily rescue inhaler use; however, this association was not significant for
ambient PM2.5 and NO2. On the other hand, traffic-related air pollutants (PM2.5, NOx, NO2) on
lag 0 were all significantly negatively associated with daily rescue inhaler use. Although
positive, personal PM2.5 on lag 0 day was not significantly associated with daily count of rescue
inhaler use.
Table 5.5. Associations of Rate Ratio (per standard deviation change) Between Air Pollutant
Exposures and Daily Count of Rescue Inhaler Use.
Exposure Estimate (95% CI) P-value N
Personal PM2.5
Lag 0 1.09 (0.80, 1.50) 0.580 86
Lag 1 1.34 (0.77, 2.35) 0.306 81
Lag 2 1.16 (0.91, 1.47) 0.238 77
Traffic PM2.5
Lag 0 0.48 (0.26, 0.88) 0.021 166
Lag 1 0.68 (0.30, 1.53) 0.352 154
Lag 2 0.56 (0.23, 1.39) 0.217 143
Traffic NOx
Lag 0 0.33 (0.13, 0.84) 0.023 166
Lag 1 0.47 (0.13, 1.72) 0.261 154
Lag 2 0.48 (0.12, 1.92) 0.307 143
Traffic NO2
Lag 0 0.34 (0.14, 0.84) 0.022 166
Lag 1 0.50 (0.15, 1.71) 0.276 154
Lag 2 0.74 (0.22, 2.45) 0.624 143
Ambient PM2.5
Lag 0 1.45 (0.90, 2.33) 0.127 167
Lag 1 0.78 (0.42, 1.47) 0.448 155
Lag 2 2.56 (0.89, 4.71) 0.187 144
Ambient O3
Lag 0 1.52 (1.02, 2.27) 0.046 167
Lag 1 0.99 (0.59, 1.68) 0.982 155
120
Lag 2 0.80 (0.48, 1.32) 0.384 144
Ambient NOx
Lag 0 1.61 (1.23, 2.11) 0.001 166
Lag 1 0.79 (0.59, 1.07) 0.134 155
Lag 2 0.42 (0.26, 1.09) 0.145 144
Ambient NO
Lag 0 1.80 (1.37, 2.35) 0.000 166
Lag 1 1.01 (0.85, 1.21) 0.902 155
Lag 2 0.45 (0.17, 1.33) 0.208 144
Ambient NO2
Lag 0 1.20 (0.85, 1.68) 0.302 166
Lag 1 0.46 (0.29, 1.12) 0.198 155
Lag 2 0.45 (0.28, 1.11) 0.186 144
*Effect estimates were scaled to a standard deviation change in pollutant concentrations as follows: personal PM
2.5
:
9.1 µg/m
3
; Traffic-related PM 2.5: 0.7 µg/m
3
; Traffic-related NO x: 14.7 ppb; Traffic-related NO 2: 7.1 ppb; Ambient
PM 2.5: 3.7 µg/m
3
; Ambient O 3: 9.2 bbp; Ambient NO x: 6.1 ppb; Ambient NO: 2.4 ppb; Ambient NO 2: 4.6 ppb;

Figure 5.1 shows single-pollutant compared with two-pollutant models for daily count of
rescue inhaler use including subjects with traffic-related PM2.5, NOx, NO2 data. Positive
associations for daily count of rescue inhaler use and lag 0 traffic-related PM2.5, NOx, NO2
remained when ambient O3 regressed together in the same model, but the associations become
non-significant.
Table 5.8 shows the associations between asthma symptoms and different air pollutants
on lag 0. One standard deviation increases of 24-hour average lag 0 traffic-related PM2.5 is
significantly associated with 83% (95% CI: 3%, 224%) higher odds of experiencing trouble
breathing because of asthma, and all the other pollutants show positive associations except for
ambient PM2.5. Most pollutants presented positive associations with day level symptoms of
wheezing, cough, exertional symptoms, or fear of dyspnea but none of the associations was
significant.

121
Figure 5.1. Results of single- and two-pollutant models (rate ratio and 95% CI per standard
deviation increase in traffic-related pollutant exposure) for daily count of rescue inhaler use in
relation to traffic-related PM2.5, NO, and NOx exposure in the last 24 hours (lag 0).

Table 5.8. Associations of Odds Ratio (per standard deviation change) Between Air Pollutant
Exposures and Asthma Symptoms on lag 0 day.
Did you wake up last night because of your asthma?
Exposure Estimate (95% CI) P-value N
Personal PM2.5 0.71 (0.11, 4.64) 0.724 52
Traffic-related PM2.5 0.52 (0.15, 1.85) 0.315 85
Traffic-related NOx 0.52 (0.16, 1.70) 0.283 85
Traffic-related NO2 0.38 (0.09, 1.56) 0.183 85
Ambient PM2.5 1.73 (0.60, 5.00) 0.315 87
Ambient O3 1.19 (0.43, 3.31) 0.737 87
Ambient NOx 0.51 (0.06, 4.03) 0.525 87
Ambient NO 0.15 (0.00, 4.73) 0.285 87
Ambient NO2 0.79 (0.14, 4.62) 0.794 87
How many times did you use your inhaler during the night?
Personal PM2.5 0.39 (0.05, 3.08) 0.379 52
Traffic-related PM2.5 0.95 (0.37, 2.42) 0.914 85
Traffic-related NOx 0.79 (0.33, 1.92) 0.609 85
Traffic-related NO2 0.61 (0.22, 1.68) 0.343 85
Ambient PM2.5 0.63 (0.23, 1.72) 0.374 87
Ambient O3 0.97 (0.47, 1.99) 0.927 87
122
Ambient NOx 0.77 (0.27, 2.22) 0.633 87
Ambient NO 1.36 (0.48, 3.87) 0.571 87
Ambient NO2 0.52 (0.15, 1.77) 0.301 87
How much of the time did your asthma keep you from getting as much done at school
or at home today?
Personal PM2.5 0.99 (0.42, 2.35) 0.986 52
Traffic-related PM2.5 1.24 (0.50, 3.04) 0.642 81
Traffic-related NOx 1.11 (0.53, 2.33) 0.791 81
Traffic-related NO2 1.09 (0.52, 2.27) 0.826 81
Ambient PM2.5 0.70 (0.25, 2.00) 0.509 83
Ambient O3 1.13 (0.45, 2.81) 0.798 83
Ambient NOx 1.15 (0.38, 3.43) 0.808 83
Ambient NO 1.48 (0.53, 4.14) 0.458 83
Ambient NO2 0.95 (0.30, 2.96) 0.929 83
Did your chest feel tight because of asthma today?
Personal PM2.5 0.85 (0.51, 1.42) 0.549 92
Traffic-related PM2.5 0.94 (0.61, 1.45) 0.796 151
Traffic-related NOx 0.97 (0.65, 1.43) 0.874 151
Traffic-related NO2 0.95 (0.65, 1.38) 0.794 151
Ambient PM2.5 1.01 (0.57, 1.79) 0.969 154
Ambient O3 0.81 (0.53, 1.25) 0.349 154
Ambient NOx 1.24 (0.67, 2.31) 0.496 153
Ambient NO 1.13 (0.55, 2.33) 0.746 153
Ambient NO2 1.29 (0.72, 2.30) 0.399 154
Did you feel wheezy because of your asthma today?
Personal PM2.5 0.93 (0.46, 1.88) 0.840 92
Traffic-related PM2.5 1.37 (0.74, 2.57) 0.320 151
Traffic-related NOx 1.17 (0.60, 2.28) 0.653 151
Traffic-related NO2 1.16 (0.61, 2.20) 0.661 151
Ambient PM2.5 1.08 (0.45, 2.57) 0.862 154
Ambient O3 0.61 (0.32, 1.14) 0.122 154
Ambient NOx 1.35 (0.63, 2.88) 0.437 153
Ambient NO 1.33 (0.57, 3.11) 0.513 153
Ambient NO2 1.33 (0.64, 2.75) 0.444 154
Did you have trouble breathing because of your asthma today?
Personal PM2.5 0.88 (0.44, 1.75) 0.720 92
Traffic-related PM2.5 0.85 (0.53, 1.36) 0.490 151
Traffic-related NOx 0.73 (0.42, 1.27) 0.261 151
Traffic-related NO2 0.73 (0.43, 1.25) 0.252 151
Ambient PM2.5 1.00 (0.55, 1.81) 0.999 154
Ambient O3 0.96 (0.61, 1.52) 0.875 154
Ambient NOx 1.10 (0.58, 2.11) 0.763 153
Ambient NO 0.87 (0.40, 1.88) 0.730 153
123
Ambient NO2 1.22 (0.66, 2.24) 0.529 154
Did you cough because of your asthma today?
Personal PM2.5 1.17 (0.68, 2.01) 0.581 92
Traffic-related PM2.5 1.45 (0.84, 2.48) 0.181 151
Traffic-related NOx 1.26 (0.74, 2.16) 0.402 151
Traffic-related NO2 1.34 (0.80, 2.27) 0.271 151
Ambient PM2.5 0.87 (0.47, 1.61) 0.654 154
Ambient O3 0.66 (0.40, 1.09) 0.106 154
Ambient NOx 1.44 (0.75, 2.76) 0.273 153
Ambient NO 1.46 (0.71, 3.03) 0.309 153
Ambient NO2 1.37 (0.74, 2.55) 0.318 154
How much of a problem was your asthma when you ran, exercise or play sports
today?
Personal PM2.5 0.66 (0.04, 10.20) 0.767 49
Traffic-related PM2.5 1.36 (0.76, 2.45) 0.302 83
Traffic-related NOx 1.34 (0.70, 2.58) 0.383 83
Traffic-related NO2 1.23 (0.68, 2.24) 0.499 83
Ambient PM2.5 1.11 (0.52, 2.39) 0.782 84
Ambient O3 1.23 (0.62, 2.43) 0.558 84
Ambient NOx 1.15 (0.52, 2.54) 0.731 84
Ambient NO 1.15 (0.50, 2.67) 0.746 84
Ambient NO2 1.16 (0.53, 2.53) 0.718 84
Did you feel scared that you might have trouble breathing because of your asthma
today?
Personal PM2.5 1.36 (0.70, 2.66) 0.365 85
Traffic-related PM2.5 1.83 (1.03, 3.24) 0.042 138
Traffic-related NOx 1.38 (0.88, 2.15) 0.163 138
Traffic-related NO2 1.31 (0.86, 2.00) 0.206 138
Ambient PM2.5 0.76 (0.33, 1.72) 0.507 140
Ambient O3 1.02 (0.56, 1.84) 0.961 140
Ambient NOx 1.74 (0.77, 3.92) 0.184 140
Ambient NO 1.35 (0.55, 3.32) 0.512 140
Ambient NO2 1.88 (0.85, 4.19) 0.124 140
Have you avoided strenuous activities, or had to slow down or stop exercising because
of your asthma today?
Personal PM2.5 0.71 (0.18, 2.84) 0.635 85
Traffic-related PM2.5 1.12 (0.66, 1.88) 0.680 138
Traffic-related NOx 0.98 (0.56, 1.71) 0.938 138
Traffic-related NO2 0.96 (0.55, 1.67) 0.880 138
Ambient PM2.5 0.67 (0.33, 1.36) 0.271 140
Ambient O3 1.00 (0.60, 1.67) 0.986 140
Ambient NOx 0.78 (0.38, 1.58) 0.485 140
Ambient NO 0.86 (0.37, 2.01) 0.733 140
124
Ambient NO2 0.76 (0.39, 1.49) 0.424 140
*Effect estimates were scaled to a standard deviation change in pollutant concentrations as follows: personal PM 2.5:
9.1 µg/m
3
; Traffic-related PM 2.5: 0.7 µg/m
3
; Traffic-related NO x: 14.7 ppb; Traffic-related NO 2: 7.1 ppb; Ambient
PM 2.5: 3.7 µg/m
3
; Ambient O 3: 9.2 bbp; Ambient NO x: 6.1 ppb; Ambient NO: 2.4 ppb; Ambient NO 2: 4.6 ppb;
5.5 DISCUSSION
Our study suggested there were significant associations between personal, ambient, and
traffic related air pollutions and asthma exacerbations in asthmatic children aged 8-16 in Los
Angeles. We found that exposures to increased ambient air pollutants of O3, NOx, NO was
significantly associated with increased rate ratio of daily count of rescue inhaler use. For daily
rescue inhaler use, the largest magnitude of association was found with ambient NO exposure. If
a child’s exposure to ambient NO were to increase by one standard deviation (2.4 ppb), his/her
rate ratio for daily rescue inhaler use would be expected to increase by a factor of 1.80. Increased
traffic-related PM2.5 exposure on day lag 0 (same day) was also significantly associated with
increased odds of symptoms of trouble breathing (odds ratio [95% CI]: 1.83 [1.03, 3.24]).
Several previous studies reported exposure to ambient air pollution of particle, O3, and
NO2 were significantly associated with increased daily rescue inhaler use
16,17
. Our study
presented significant positive association between increased ambient O3 (rate ratio [95% CI]:
1.52 [1.02, 2.27]), ambient NOx (rate ratio [95% CI]: 1.61 [1.23, 2.11]), ambient NO (rate ratio
[95% CI]: 1.80 [1.37, 2.35]) and daily rescue inhaler use. These findings are consistent with
existing evidence that short-term exposure to ambient O3 and NO2 can cause airway
inflammation, reduced pulmonary function, and exacerbation in individuals with asthma
18
. The
current U.S. EPA standard for 8-hour maximum O3 concentration is 70 bbp
19
, yet we
demonstrated significant associations between O3 much lower than this standard (mean 24-hour
average O3 exposure was 32.3 bbp). Moreover, these results suggest that the current EPA ozone
standard may not sufficiently protect individuals with asthma from experiencing rescue inhaler
125
use. Exposure to ambient PM2.5 was not significantly associated with daily rescue inhaler use
(rate ratio [95% CI]: 1.45 [0.90, 2.33]) in this study which is contradictory to previous studies.
Possible reasons could be the average PM2.5 exposure in this study may have been too low to
have an effect on asthma. The Air Quality Index for PM2.5 defines a “Good” index value as 0–12
μg/m
3
, and in our study, the average PM2.5 exposure was 9.5 μg/m
3
, which falls within this range.
All traffic-related pollutants showed significant negative associations with rescue inhaler use
events, but after adjusting for ambient O3 in the same model, these associations became non-
significant.
For traffic-related PM2.5, NO and NOx, we found increased exposure to those pollutants
was associated with decreased rate of rescue inhaler use. However, after we adjusted ozone in
the same model, these associations became non-significant, but the point estimate was still
negative. The possible explanation could be residual confounding due to other co-occurring
exposures, behaviors or time-activity patterns.
Our study found significant associations between traffic-related PM2.5 and being scared
of having trouble breathing at a daily level. But we did not see significant associations of
increased traffic-related PM2.5 and daily trouble breathing symptoms. The possible explanations
could be increased perception of risk when parents or children were aware of poorer air quality,
or that stress and anxiety levels of children anticipating experiencing an asthma attack may play
a role as effect modifiers. Several studies reported the associations between increased air
pollution levels and asthma symptoms, and limited research has studied the symptom of
trouble/difficulty breathing. One study in Washington, United States
20
, studied the effects of
exposure to different sizes of particles and several asthma symptoms in children, including
trouble breathing. From their results, the effect of pollutants on trouble breath was positive but
126
not significant. Our study results also show positive but not significant associations between air
pollutant exposures and daily asthma symptoms of cough and having asthma problems when
running, exercising, or playing sports. Cough is a commonly studies symptom of asthma
exacerbation in many previous studies
21–23
, but associations were not consistent across different
studies. One possible reason for the lack of effects between air pollution and asthma symptoms
in our study may be the small sample size.
The BREATHE mHealth platform design reduces recall bias and outcome
misclassification compared to previous studies using daily diary cards. BREATHE mHealth
platform allows multiple sensors to be connected in a single study and enables personal-level
data collection through EMAs as well as integration of additional external data (e.g.,
environmental, clinical and medical records) to paint a complete picture of each subject. Using
the BREATHE mHealth platform design, future studies could track a participant’s movements
throughout the day to assign accumulated exposures, improve classifications of time spend
indoors and outdoors, or investigate the within-person and within-day associations between
exposures, behaviors and health outcomes in context.
There were a number of limitations in this study. First, the personal PM2.5 values were
measured from the personal sensor and were not calibrated, and low-cost optical sensors as the
one used in this study tend to overestimate PM2.5 concentrations in general largely due to relative
humidity interferences. Whether or not the overestimation can have any effect on the health
model findings could be assessed in future analyses; however, it is unlikely to be differential in
relation to the outcome. Secondly, compliance was another issue in our study where subjects
sometimes forgot to charge their devices, forgot or were not able to take them with them to
school or did not do the spirometry tests in the pre-set time. Therefore, for future studies,
127
mHealth approaches may be applied to provide timely and proactive reminders for study subjects
with an emphasis on encouraging and monitoring compliance without biasing data collection.
Third, we tried to adjusted age in our model, but age was not a founder based on the 10% rule.
However, the random intercept in our mixed effect modeling approach probably captured some
of the effects of age as a person-level characteristic. Furthermore, given the rate of growth of
lungs between our study population ages (8-16 years old), the possibility remains that
variable lung growth rate based on age may be a factor in susceptibility that may be elucidated
with age stratification in future studies with larger sample size.
In addition, our study findings were limited in terms of their generalizability. Our
participants were recruited from a pediatric asthma specialty clinic and had moderate to severe
asthma. As such, triggers and factors associated with exacerbations in this population might not
necessarily translate to children with mild asthma or well-controlled asthma. In addition,
socioeconomic characteristics of our study sample are likely not representative of the general
population, where our participants had higher parental education level (37.5% of their caretaker
had graduate level degree, higher household income level (57.5% of their caretaker reported over
$50,000 household income annually), and higher health insurance coverage (100% subjects had
health insurance). Finally, selection bias may have influenced our results if parents living in
areas with higher air pollution exposure were also more concerned about its impact on their
child’s asthma and therefore more likely to participate in our study. However, it is difficult to
ascertain the direction of this potential selection bias.
In conclusion, we found significant and positive associations between O3, NOx, and NO
exposures and daily rescue inhaler use, and traffic-related PM2.5 and trouble breathing. The
longitudinal nature of our study and the acute outcome measurement through the BREATHE
128
mHealth platform provides further support for the plausibility of the associations of exposure to
air pollutants. Our study demonstrates the utility of the BREATH kit at collecting highly
resolved exposure and outcome data for investigating acute health effects.

129
5.6 REFERENCES:
1. Centers for Disease Control and Prevention. CDC - Asthma - Most Recent Asthma Data.
Most Recent Asthma Data.
2. Kuschner WG. The asthma epidemic [12]. N Engl J Med. 2007.
doi:10.1056/NEJMc063596
3. van Dellen QM, Stronks K, Bindels PJE, Öry FG, Bruil J, van Aalderen WMC. Predictors
of asthma control in children from different ethnic origins living in Amsterdam. Respir
Med. 2007. doi:10.1016/j.rmed.2006.08.002
4. Akinbami L. The state of childhood asthma, United States, 1980-2005. Adv Data. 2006.
5. Barnes PJ, Jonsson B, Klim JB. The costs of asthma. Eur Respir J. 1996.
doi:10.1183/09031936.96.09040636
6. Juniper EF, Guyatt GH, Feeny DH, Ferrie PJ, Griffith LE, Townsend M. Measuring
quality of life in children with asthma. Qual Life Res. 1996. doi:10.1007/BF00435967
7. GINA C científico de l. Global Strategy for Asthma Management and Prevention ( 2016
update ). Guía Gina Actual. 2016.
8. Etzel RA. How environmental exposures influence the development and exacerbation of
asthma. In: Pediatrics. ; 2003.
9. Spann K, Snape N, Baturcam E, Fantino E. The Impact of Early-Life Exposure to Air-
borne Environmental Insults on the Function of the Airway Epithelium in Asthma. Ann
Glob Heal. 2016. doi:10.1016/j.aogh.2016.01.007
10. Bouazza N, Foissac F, Urien S, et al. Fine particulate pollution and asthma exacerbations.
Arch Dis Child. 2018. doi:10.1136/archdischild-2017-312826
11. Williams AM, Phaneuf DJ, Barrett MA, Su JG. Short-term impact of PM 2.5 on
130
contemporaneous asthma medication use: Behavior and the value of pollution reductions.
Proc Natl Acad Sci U S A. 2019. doi:10.1073/pnas.1805647115
12. Habre R, Moshier E, Castro W, et al. The effects of PM2.5 and its components from
indoor and outdoor sources on cough and wheeze symptoms in asthmatic children. J Expo
Sci Environ Epidemiol. 2014. doi:10.1038/jes.2014.21
13. Bui AAT, Hosseini A, Rocchio R, et al. Biomedical REAl-Time Health Evaluation
(BREATHE): toward an mHealth informatics platform. JAMIA Open. 2020.
doi:10.1093/jamiaopen/ooaa011
14. duRivage N, Ross M, Mayne SL, et al. Asthma Control Test. Clin Pediatr (Phila). 2017.
doi:10.1177/0009922816675013
15. Alduchov OA, Eskridge RE. Improved Magnus form approximation of saturation vapor
pressure. J Appl Meteorol. 1996. doi:10.1175/1520-
0450(1996)035<0601:IMFAOS>2.0.CO;2
16. Schildcrout JS, Sheppard L, Lumley T, Slaughter JC, Koenig JQ, Shapiro GG. Ambient
air pollution and asthma exacerbations in children: An eight-city analysis. Am J
Epidemiol. 2006. doi:10.1093/aje/kwj225
17. Pepper JR, Barrett MA, Su JG, et al. Geospatial-temporal analysis of the impact of ozone
on asthma rescue inhaler use. Environ Int. 2020. doi:10.1016/j.envint.2019.105331
18. Guarnieri M, Balmes JR. Outdoor air pollution and asthma. Lancet. 2014.
doi:10.1016/S0140-6736(14)60617-6
19. USEPA. National Ambient Air Quality Standards.; 2014.
20. Mar TF, Larson T V., Stier RA, Claiborn C, Koenig JQ. An analysis of the association
between respiratory symptoms in subjects with asthma and daily air pollution in Spokane,
131
Washington. Inhal Toxicol. 2004. doi:10.1080/08958370490506646
21. Esposito S, Galeone C, Lelii M, et al. Impact of air pollution on respiratory diseases in
children with recurrent wheezing or asthma. BMC Pulm Med. 2014. doi:10.1186/1471-
2466-14-130
22. Timonen KL, Pekkanen J. Air pollution and respiratory health among children with
asthmatic or cough symptoms. Am J Respir Crit Care Med. 1997.
doi:10.1164/ajrccm.156.2.9608044
23. Jalaludin BB, O’Toole BI, Leeder SR. Acute effects of urban ambient air pollution on
respiratory symptoms, asthma medication use, and doctor visits for asthma in a cohort of
Australian children. Environ Res. 2004. doi:10.1016/S0013-9351(03)00038-0

132
5.7 SUPPLEMENT
Supplement Figure 1. Morning and Evening Lung Function Spirometry Tests Distributions.
A

B

133
Supplement Figure 2. Rescue Inhaler Use Daily Count Distribution across All Subjects.

134
Supplement Figure 3. PM2.5 Related Exposures Daily Average Distribution across All Subjects.
Personal

Ambient

Traffic

135
6 CHAPTER 6: CONCLUSIONS AND FUTURE DIRECTIONS

6.1 CONCLUSIONS AND FUTURE DIRECTIONS
Pediatric asthma in the United States presents a significant clinical and social burden on
children and their families
1
. Achieving clinically adequate control of pediatric asthma is critical
to avoiding the deleterious effects of asthma exacerbation
2
. While the factors involved in asthma
control are varied and complex, environmental exposures are known to be an important
contributing trigger
3
.
While there are a multitude of such environmental triggers responsible for poor long-term
control of asthma in children, significant gaps in knowledge remain on how these exposures
contribute to acute worsening of disease on an hourly or daily basis. Furthermore, many of the
previous epidemiological studies relied on ambient air pollution measurements or estimates as a
surrogate of true personal exposure, potentially introducing exposure measurement error which
may weaken statistical power to detect effects and introduce bias or distort the estimation of true
risk
4
. The microenvironmental approach has been widely used in the epidemiological literature
as an approach to assess personal exposure than personal monitoring. However, applying the
microenvironmental modeling approach requires accurate time-activity data.
This dissertation filled in the above research gaps by first developed a novel mHealth
approach to improve time-activity classification by using smartphone sensor data and machine
learning techniques. Then, based on the work in the Los Angeles Pediatric Research Using
Integrated Sensor Monitoring System (LA PRISMS) Center, where we developed an informatics
platform for wearable sensor-based environmental health, we estimated the day-level acute
effects of exposure to personal, ambient and traffic-related air pollutants on asthma
136
exacerbations (lung function decrements, use of rescue medication, increased odds of symptoms)
in children with moderate to severe asthma.
The random forest model we developed to predict the microenvironment can achieve
validation balanced accuracy (BA) of 95.1% in the same subject and 57.5% in a different subject.
However, if two subjects shared similar daily activity patterns, the BA increased to 70%. BA
also dropped after considering temporal, spatial, and spatiotemporal autocorrelation (63.2%,
60.9%, and 51.4%, respectively). From this work, we demonstrated the ability of using
smartphone sensor data to predict subject’s microenvironment, and the importance of accounting
for autocorrelation in high resolution data to when assessing prediction performance, and the
potential to generalize models to groups sharing somewhat similar time-activity patterns. For
future studies, generalize the current model to larger scale exposure and epidemiology
applications could be another direction.
We also presented exposure to ambient O3 and traffic-related PM2.5 were significantly
associated with acute day-level asthma exacerbations (lung function, rescue inhaler uses and
asthma symptoms of trouble breathing). From these two works, we demonstrated the utility of
the mHealth at collecting highly resolved exposure and outcome data for investigating acute
health effects. We also showed the possibility by using mHealth to develop more tailored and
personalized approaches to investigating acute asthma triggers and their association with
subsequent exacerbations. For future studies, estimate the associations between increased air
pollution and within-day level acute effects of asthma exacerbation by using the case-crossover
design could be another direction.
In summary, our work by utilizing sensor-based mobile health approaches supports the
importance of the three-domain framework which were illustrated by Larkin and Hystad
5
. Air
137
pollution sensors, smartphone applications, and data science are three important domains, and
combines all three domains into a properly integrated framework can reduce outcome and
exposure misclassification and contribute to advances in exposure assessment and risk evaluation
in epidemiology studies.

138
6.2 REFERENCES
1. Zahran HS, Bailey CM, Damon SA, Garbe PL, Breysse PN. Vital signs: Asthma in
children — United States, 2001-2016. Morb Mortal Wkly Rep. 2018.
doi:10.15585/mmwr.mm6705e1
2. Barnes PJ, Jonsson B, Klim JB. The costs of asthma. Eur Respir J. 1996.
doi:10.1183/09031936.96.09040636
3. Etzel RA. How environmental exposures influence the development and exacerbation of
asthma. In: Pediatrics. ; 2003.
4. Zou B, Wilson JG, Zhan FB, Zeng Y. Air pollution exposure assessment methods utilized
in epidemiological studies. J Environ Monit. 2009. doi:10.1039/b813889c
5. Larkin A, Hystad P. Towards Personal Exposures: How Technology Is Changing Air
Pollution and Health Research. Curr Environ Heal reports. 2017. doi:10.1007/s40572-
017-0163-y

Abstract (if available)

Abstract The American Public Health Association recognizes that asthma is one of the leading chronic diseases among children in the United States. Children with asthma are significantly burdened by asthma morbidity, including emergency department visits, hospitalizations, and even deaths. A large proportion of the burden of asthma is attributed to treating the consequences of poor asthma control, so asthma control becomes an important aspect of the evaluation and management of the disease. The mechanisms leading to asthma exacerbation in children are complex, but environmental exposures are largely presumed to be important. Environmental asthma triggers include indoor and outdoor allergens, such as dust mites, animal allergens, molds and pollens, and indoor and outdoor pollutants, including environmental tobacco smoke (or secondhand smoke), chemical, combustion by-products, and ozone and particulate matter. While there are a multitude of known environmental triggers responsible for poor long-term control of asthma in children, there is limited knowledge in how these exposures contribute to acute worsening of disease on an hourly or daily basis. ❧ Furthermore, many of the above epidemiological studies relied on ambient air pollution measurements or estimates (for example, measurements at nearest central site, or models of outdoor residential air pollution) as a surrogate of true personal exposure, potentially introducing exposure measurement error which may weaken statistical power to detect effects and introduce bias or distort the estimation of true risk. ❧ Therefore, our goal in this work is to develop and apply sensor-based mobile Health (mHealth) approaches to improve time-activity classification and investigate acute triggers of asthma exacerbation in children. We first develop a novel mHealth approach to improve time-activity classification by using smartphone sensor data and machine learning techniques. Then, based on our work in the Los Angeles Pediatric Research Using Integrated Sensor Monitoring System (LA PRISMS) Center, where we developed an informatics platform for wearable sensor-based environmental health studies and deployed it in a panel study of pediatric asthma, we aim to estimate the day-level and within-day acute effects of air pollution exposures on asthma exacerbations (lung function decrements, use of rescue medication, increased odds of symptoms) in children with moderate to severe asthma.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Prenatal environmental exposures and fetal growth in the MADRES cohort

PDF

Airway inflammation and respiratory health in the Southern California children's health study

PDF

Prenatal air pollution exposure, newborn DNA methylation, and childhood respiratory health

PDF

Air pollution, mitochondrial function, and growth in children

PDF

Air pollution and childhood obesity

PDF

Air pollution and breast cancer survival in California teachers: using address histories and individual-level data

PDF

Personal exposure to particulate matter PM2.5 sources during pregnancy and birthweight

PDF

Prenatal exposure to organophosphate esters and early neurodevelopment

PDF

A cohort study of air-pollution and childhood obesity incidence

PDF

Assessing the impact of air pollution on adverse birth outcomes in a low resource setting

PDF

Spatial analysis of PM₂.₅ air pollution in association with hospital admissions in California

PDF

Disparities in exposure to traffic-related pollution sources by self-identified and ancestral Hispanic descent in participants of the USC Children’s Health Study

PDF

Ambient air pollution and lung function in children

PDF

Early life air pollution exposure, gestational diabetes mellitus, and autism spectrum disorder

PDF

Spatial modeling of non-tailpipe emissions and its association with children's lung function

PDF

Early life risk factors for childhood asthma

PDF

Assessment of the mortality burden associated with ambient air pollution in rural and urban areas of India

PDF

A pilot study of a global approach to assessing air pollution exposure in port communities: passive air monitoring of nitrogen dioxide concentrations

PDF

Associations between ambient air pollution and hypertensive disorders of pregnancy

PDF

Native American ancestry among Hispanic Whites is associated with higher risk of childhood obesity: a longitudinal analysis of Children’s Health Study data

Asset Metadata

Creator Hao, Hua (author)

Core Title Sensor-based mobile health approaches for personal air pollution and pediatric asthma studies

School Keck School of Medicine

Degree Doctor of Philosophy

Degree Program Epidemiology

Publication Date 11/29/2020

Defense Date 10/23/2020

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Air pollution,asthma,Epidemiology,mHealth,OAI-PMH Harvest,sensor

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Habre, Rima (committee chair), Bastain, Theresa (committee member), Eckel, Sandrah (committee member), Loyola, Laura (committee member)

Creator Email haohua19880408@gmail.com,hhao@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-395746

Unique identifier UC11667823

Identifier etd-HaoHua-9159.pdf (filename),usctheses-c89-395746 (legacy record id)

Legacy Identifier etd-HaoHua-9159.pdf

Dmrecord 395746

Document Type Dissertation

Rights Hao, Hua

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA