University of Southern California Dissertations and Theses

Health economics and outcomes research for informed decision making in rapidly evolving therapeutic areas

(USC Thesis Other)

Health economics and outcomes research for informed decision making in rapidly evolving therapeutic areas

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content
HEALTH ECONOMICS AND OUTCOMES RESEARCH
FOR INFORMED DECISION MAKING
IN RAPIDLY EVOLVING THERAPEUTIC AREAS

by

Sang Kyu Cho

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA

In Partial Fulfillment of the Requirements for the Degree
DOCTOR OF PHILOSOPHY
(HEALTH ECONOMICS)

May 2020

ii

Acknowledgements
My sincere thanks to my advisor and dissertation chair, Dr. Jeffrey McCombs for he has
taught me how to think like an economist. His consistent and generous support has been
vital in my transition from a pharmacist to an independent researcher. Under his guidance,
I have learned how to transform ideas into tangible research work. I am also very grateful
to have Drs. Joel Hay and Soeren Mattke on my dissertation committee. I have learned
tremendously from the three semesters of pharmaceutical economics class with Dr. Hay.
I have always appreciated and his patience, passion, and meticulousness that continue
to inspire me to become a better researcher. Close mentorship and guidance from Dr.
Mattke have expanded my research horizon to different therapeutic areas and methods
and prepared me to succeed beyond the classroom and graduate school. I would also
like to express my appreciation for Drs. Steven Fox and Sze-chuan Suen for their
guidance with my dissertation proposal, and all my friends in the PMEP program for going
through this journey together.

iii

Table of Contents

Acknowledgements………………………………………..………………………………….…ii
List of Tables…………………………………………………………………………………….iv
List of Figures………………………………………………………………………...……….…v
Abstract……………………………………………………………………………………..……vi
Introduction ..................................................................................................................... 1
References ................................................................................................................... 5
Chapter 1 – Complications and hospital costs during hematopoietic stem cell
transplantation for non-Hodgkin lymphoma in the United States
Abstract ........................................................................................................................ 8
Background .................................................................................................................. 9
Materials and Methods ............................................................................................... 10
Results ....................................................................................................................... 12
Discussion.................................................................................................................. 14
References ................................................................................................................. 23
Chapter 2 – Cost-effectiveness Analysis of Regorafenib and TAS-102 in Refractory
Metastatic Colorectal Cancer in the United States
Abstract ...................................................................................................................... 30
Background ................................................................................................................ 30
Materials and Methods ............................................................................................... 32
Results ....................................................................................................................... 38
Discussion.................................................................................................................. 39
References ................................................................................................................. 43
Chapter 3 – Development of a Model to Predict Healing of Chronic Wounds within
Twelve Week
Abstract ...................................................................................................................... 47
Background ................................................................................................................ 49
Materials and Methods ............................................................................................... 51
Results ....................................................................................................................... 57
Discussion.................................................................................................................. 58
References ................................................................................................................. 67
Supplemental Appendices ............................................................................................. 73

iv

List of Tables
Chapter 1
Table 1. Baseline Characteristics……………………….……………...……....….. 19
Table 2. Complication Rates and Unadjusted Mean Hospital Costs…….....…...20
Table 3. Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Autologous HSCT………...……………....…..21
Table 4. Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Allogeneic HSCT……………...…...…….……22

Chapter 2
Table 1. Total Cost by Treatment…………………………………….…………...….40
Table 2. Incremental Cost-Effectiveness Ratios……………………………….….. 40
Table 3. Result of Scenario Analysis………………………………………….……..42

Chapter 3
Table 1. Comparison of Healed versus Not-healed Wounds………….......………62
Table 2. Model Comparison…………………………………………….…………….63
Table 3. Estimated Odds Ratios from Logistic Model……………………..………..65
Table 4. Relative Variable Importance……………………………………...………. 66

v

List of Figures

Chapter 2
Figure 1. Markov Cycle………………………...……………………………….……. 43
Figure 2. Tornado Diagram…………………………………………..…………….…44
Figure 3. Cost-Effectiveness Acceptability Curves ……………………………. …..45

Chapter 3
Figure 1. Comparison of Area under the Receiver Operator Curves………..……67

vi

Abstract
The field of health economics and outcomes research (HEOR) continues to evolve as
patients, health care providers, and payers seek to make a judicious choice between cost,
quality, and access. Taken together, these three papers illustrate how HEOR studies can
be used to assist decision making of the stakeholders. First paper reports the cost and
safety of hematopoietic stem cell transplantation, which serve as a standard of care
treatment for relapsed and refractory non-Hodgkin’s lymphoma, and assist
comprehensive evaluation of new CAR-T therapy in the absence of head-to-head clinical
trials. Second paper used a three state Markov model to examine the cost-effectiveness
of TAS-102 and regorafenib in metastatic, refractory colorectal cancer in the United
States in light of recent clinical evidence and underscores the importance of managing
treatment-related toxicities and maintaining health-related quality of life. Third paper
develops and validates severity-adjusted prediction models for chronic wound healing
that can be used to implement value-based reimbursement by estimating the differences
between actual and predicted healing rates across wound care centers.

1

Introduction

The field of health economics and outcomes research (HEOR) has grown rapidly over the
past two decades, serving as a guide to the judicious allocation of health care resources.
The need for timely HEOR studies, such as clinical and economic evaluation of new
treatment relative to standard of care, has become indispensable in recent years to inform
discussion about cost, quality, and access of medical care and services. The three papers
that comprise this dissertation aim to demonstrate the application of HEOR studies in two
rapidly evolving therapeutic areas, oncology and wound care, to assist informed decision
making for clinicians, payers, policy makers, and ultimately, patients.
First paper provides new parameter estimates needed to evaluate the essential
need for new therapies to treat patients with hematologic malignancies requiring stem cell
transplantation. Specifically, this paper examines safety and costs of initial hospitalization
for hematopoietic stem cell transplantation (HSCT) in patients with non-Hodgkin’s
lymphoma (NHL). The treatment landscape for NHL has been revolutionized in recent
years with the introduction of improved supportive care and pre-conditioning regimen for
HSCT as well as chimeric antigen receptor T-cell therapy (CAR-T). In 2017, a CAR-T
therapy received an approval from the Food and Drug Administration (FDA) to treat non-
Hodgkin’s lymphoma for patients who have failed more than two other kinds of treatment.
As of 2017, this therapy is priced at approximately $373,000, targeting patients who may
otherwise be eligible for HSCT. [1] The cost-effectiveness of a CAR-T therapy will depend
critically of the cost and outcomes achieved using existing HSCT therapies.
2

Against this background, the cost of initial transplant hospitalization for HSCT has
become an important area of research for payers, policy makers, and clinicians who are
undertaking the evaluation of new CAR-T therapy. However, currently, little is known
about the cost and outcomes associated with HSCT in patients with non-Hodgkin’s
lymphoma. Previous studies have documented that the cost of initial transplant
hospitalization accounts for ~75% of total health care costs for first 100-days following
HSCT, but with considerable variation in their cost estimates over transplant type,
preconditioning regimen, stem cell source, and in-hospital complications. [2,3] Therefore,
using a nationally representative hospital discharge database, this first paper aims to
describe the hospital costs associated with HSCT and to identify their predictors.
Second paper evaluates the cost-effectiveness of two drugs used for the treatment
of refractory metastatic colorectal cancer (mCRC) from a US payer’s perspective.
Regorafenib and TAS-102 (trifluridine/tipiracil) have shown to extend survival by 6 and 8
weeks, respectively, when compared to best supportive care alone in mCRC patients. [4]
Several value frameworks from American Society of Clinical Oncology (ASCO), European
Society for Medical Oncology (ESMO), and National Comprehensive Cancer Network
(NCCN) rank two drugs similarly and the treatment selection often depends on the
patient’s preference and their ability to tolerate different toxicities. [5,6] TAS-102 is
commonly associated with hematologic toxicity such as neutropenia at treatment initiation
while regorafenib is associated with lasting, painful dermal toxicity called hand-foot
syndrome.
Two earlier cost-effectiveness analysis studies have compared regorafenib and
TAS-102 from the perspectives of foreign payers. Their results favored TAS-102, based
3

on efficacy and safety data demonstrated during each drug’s phase 3 trials. [7,8] However,
these studies failed to adequately explore the uncertainty that arises during indirect
comparison.
Recent real-world studies have reported that two drugs have comparable efficacy.
Two European health technology agencies, the National Institute of Health and Care
Excellence (NICE) and the Institute for Quality and Efficiency in Health Care in Germany
(IQWiG), have analyzed the two drug’s clinical trials data and report new findings. [9-11]
In light of new evidence, this paper evaluates the cost-effectiveness of regorafenib and
TAS-102 in mCRC patients from a US payer’s perspective.
Third paper introduces a prediction model for chronic wound healing that can can
be used to support the implementation of value-based reimbursement models. Chronic
wounds, usually defined as wounds that take more than 4-12 weeks to heal, are a growing
but little recognized public health challenge resulting in $28.1 and $96.8 billion in
estimated Medicare expenditure in 2014. [12] From advanced dressings to growth factors
and biologics, substantial progress has been made in product development and treatment
approaches for chronic wounds. However, the uptake of these innovations and availability
of real-world evidence to support their use remain limited. [13]
For other chronic conditions such as heart failure and diabetes, payment reforms
based on value-based reimbursement models have been proposed because the
established fee-for-service payment systems are increasingly regarded as
disincentivizing high-quality care. [14] These proposed reforms aim to link reimbursement
to quality metrics that measure patient health outcomes, clinician’s adherence to clinical
practice guideline, and patient satisfaction. A critical requirement for the introduction of
4

such value-based reimbursement models is a scientifically sound and actionable quality
indicator that will trigger rewards or penalties. An obvious choice in chronic wound care
is the wound healing rate, as closing wounds is both the intended outcomes of treatment
and readily observable. Therefore, this paper introduces a quality metric for chronic
wound care by developing a severity adjusted prediction model for 12-week healing rate
that accounts for heterogeneous etiology of chronic wounds and the multitude of other
factors that influences healing rates.

5

References

1. Evidence Report: Chimeric Antigen Receptor T-Cell Therapy for B-Cell Cancers:
Effectiveness and Value. Institute for Clinical and Economic Review. February
2018. Available from https://icer-review.org/wp-
content/uploads/2017/07/ICER_CAR_T_Evidence_Report_021518.pdf
[Accessed March 2019]
2. Broder M, Quock T, Change E, Reddy S, Agarwal-Hashimi R, Arai S, et al. The
Cost of Hematopoietic Stem-Cell Transplantation in the United States. Am Health
Drug Benefits. 2017 Oct; 10(7): 366–374.
3. Majhail NS, Mau LW, Denzen EM, Arneson TJ. Costs of Autologous and
Allogeneic Hematopoietic Cell Transplantation in the United States: A Study
Using a Large National Private Claims Database. Bone Marrow Transplant. 2013
Feb; 48(2):294-300.
4. Mayer RJ, Van Cutsem E, Falcone A et al. Randomized Trial of TAS-102 for
Refractory Metastatic Colorectal Cancer. N Engl J Med 2015; 372(20):1909-19.
5. Grothey A, Van Cutsem E, Sobrero A et al. Regorafenib Monotherapy for
Previously Treated Metastatic Colorectal Cancer (CORRECT): an International,
Multicentre, Randomized, Placebo-controlled Phase 3 Trial. Lancet 2013;
381(9863):303-12.
6. NCCN Clinical Practice Guideline in Oncology: Colon Cancer NCCN Evidence
Blocks. Version 2.2018. Available from
https://www.nccn.org/professionals/physician_gls/pdf/colon_blocks.pdf
[Accessed July 2018]
6

7. Becker DJ, Lin D, Lee S, et al. Exploration of the ASCO and ESMO Value
Frameworks for Antineoplastic Drugs. J Oncol Pract. 2017 Jul;13(7):e653-e665
8. Bullement A, Underhill S, Fougeray R et al. Cost-Effectiveness of
Trifluridine/tipiracil for Previously Treated Metastatic Colorectal Cancer in
England and Wales. Clin Colorectal Cancer 2017 S1533-0028(16)30287.
9. Kimura M, Usami E, Iwai M et al. Comparison of cost-effectiveness of
regorafenib and trifluridine/tipiracil combination tablet for treating advanced and
recurrent colorectal cancer. Mol Clin Oncol. 2016 Nov;5(5):635-640.
10. [A14-09] Appendum to Comission A13-37 (regorafenib). Institute for Quality and
Efficiency in Health Care. Available from https://www.iqwig.de/en/projects-
results/projects/drug-assessment/a14-09-addendum-to-commission-a13-37-
regorafenib.6022.html [Accessed January 2017]
11. Single Technology Appraisal. Committee Paper [ID 876], Trifluridine with Tipiracil
Hydrochloride for Treating Metastatic Colorectal Cancer After Standard Therapy.
National Institute for Health and Care Excellence. Available from
https://www.nice.org.uk/guidance/ta405/documents/committee-papers [Accessed
December 2017]
12. Nussbaum SR, Carter MJ, Fife CE, DaVanzo J, Haught R, Nusgart M, Cartwright
D. An Economic Evaluation of the Impact, Cost, and Medicare Policy Implications
of Chronic Nonhealing Wounds. Value Health. 2018 Jan;21(1):27-32.
13. Murphy PS, Evans GR. Advances in wound healing: a review of current wound
healing products. Plast Surg Int. 2012; 2012:190436.
7

14. Snyder RJ, Fife C, Moore Z. Components and Quality Measures of DIME
(Devitalized Tissue, Infection/Inflammation, Moisture Balance, and Edge
Preparation) in Wound Care. Send to Adv Skin Wound Care. 2016
May;29(5):205-15.

8

Chapter 1 – Complications and hospital costs during hematopoietic stem
cell transplantation for non-Hodgkin lymphoma in the United States
Sang Kyu Cho, Jeffrey McCombs, Nathan Punwani and Jenny Lam.
Leukemia and Lymphoma. 2019 Mar 8:1-7.

Abstract
While the cost of initial hospitalization accounts for 75% of total healthcare spending
during the first 100 days following hematopoietic stem cell transplantation (HSCT), there
is a lack of studies evaluating the considerable variation in its cost estimates. Using the
National Inpatient Sample (NIS) database from 2012–2014, we identified 1832 adult non-
Hodgkin lymphoma (NHL) patients who received autologous or allogeneic HSCT.
Complications occurred in >70% of patients, and the presence of one or more
complications was associated with an increase in mean hospital costs of 46% in
autologous HSCT and 81% in allogeneic HSCT. The most common complications (~40%)
were mucositis, febrile neutropenia, and infection. Acute organ failure, acute graft-versus-
host disease, and death were less frequent (~10%) but had a greater impact on increasing
hospital costs and length of stays. Despite recent advances in supportive care and pre-
conditioning regimens, complications remain common and costly during HSCT.
9

Background
Hematopoietic stem cell transplantation (HSCT) is a standard of care treatment for many
benign and malignant hematologic disorders. [1] Autologous HSCT is recommended as
a consolidative therapy for patients with either relapsed non-Hodgkin lymphoma (NHL) or
first remission for select lymphomas. [2,3] Allogeneic HSCT is typically reserved for
patients who are not candidates for autologous HSCT in cases of lymphoma with bone
marrow involvement, stem cell mobilization failures that preclude autologous HSCT, or
relapse after autologous HSCT. [4] Allogeneic transplant requires a matching donor and
is rife with severe adverse events such as graft-versus-host disease. [5]
With recent advances in supportive care and preconditioning regimens, HSCT has
become a routine treatment for various hematologic disorders, and its clinical and
economic value is worthy of continued evaluation and discussion. [6] The cost of the initial
transplant hospitalization is particularly important because it accounts for ~75% of total
healthcare costs during the first 100 days following HSCT. [7,8]
Despite its substantial contribution to the total cost, there is a lack of studies that
investigate considerable variation in cost estimates for initial transplant hospitalization,
which ranges from $36,000–$80,000 for autologous HSCT and $96,000–$204,000 for
allogeneic HSCT (2012 US$). [1] Preconditioning regimen, transplant type, stem cell
source, and complications were identified as predictors for the cost of initial transplant
hospitalization, but the comparison of different cost estimates across studies is
constrained by differences in accounting methods, intensity of pre-conditioning regimens,
patient populations, and length of follow-up. [1,9–11]
10

Using the 2000–2001 National Inpatient Sample (NIS), Jones et al. study was the
first to report population-based estimates of complication rates and their impact on
hospital costs and length of stays (LOS). [12] Significant complications were documented
for more than half of the admissions and were associated with higher hospital costs, LOS,
and inpatient mortality. The use of HSCT has since been steadily increasing, and the NIS
database has been redesigned to provide more precise population-based estimates. [13]
Therefore, our study aims to expand on previous research and provides updates on
complication rates and hospital costs during the initial transplant hospitalization for
autologous and allogeneic HSCT in NHL.

Materials and Methods
Data
Our study used 2012–2014 NIS database sponsored by the Agency for Healthcare
Research and Quality (AHRQ). NIS is the largest publicly available, all-payer inpatient
healthcare database covering ~20% of the annual 35 million hospital discharges in the
United States. [14] The database includes variables on patient and hospital
characteristics, medical diagnoses and procedural codes, LOS, and estimated total
charges. The NIS database excludes federal institutions, rehabilitation centers, and long-
term acute care facilities.

11

Outcome Coding and Patient Selection
International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9)
codes were used to identify discharge records for adult patients (age>=18) who had a
hospitalization for an autologous or allogeneic HSCT during the 3-year period between
January 1st, 2012 and December 31st, 2014. From this population, patients with a
principal diagnosis code for NHL were selected. All subtypes of NHL patients were
therefore included in our study.
Observations with a procedure code 41.00 (bone marrow transplant, not otherwise
specified) were removed because this procedure code does not specify transplant type.
Patients with cord blood stem cell transplant (41.06) were also excluded as they are using
an alternative graft source that typically allows for greater HLA-mismatch than traditional
allogeneic transplantation. [15] Using ICD-9 codes, Charlson Comorbidity Index (CCI),
treatment characteristics, and complications were identified. [16] ICD-9 codes used to
identify clinical outcomes are presented in the Supplemental Appendix.

Cost Calculation
The NIS database includes total charges for each hospitalization. These charges were
converted to costs using the AHRQ cost-to-charge ratios (CCR), which are based on
hospital accounting reports collected by the Center for Medicare and Medicaid Services
(CMS). Costs reflect the actual expenses incurred in the production of hospital services,
such as wages, supplies, and utility costs; charges represent the amount a hospital billed
12

for the case. [17] All costs were inflated to 2014 US dollars using the Consumer Price
Index (CPI) for medical care. [18]

Statistical Analysis
In addition to descriptive statistics, separate ordinary least square (OLS) models for
autologous HSCT and allogeneic HSCT were used to assess the impact of complications
on hospital cost and LOS. These multivariate regression models adjusted for relevant
variables like the characteristic of patients, hospitals and procedures. [19–21] Treatment
characteristics and the presence of complications were analyzed as dichotomous
variables, while the patient and hospital characteristics were recorded as categorical
variables. All analyses were performed using STATA 14 (College Station, TX), and robust
standard errors were used for two-sided hypothesis testing. Statistical significance was
defined as a p-value of less than 0.05.

Results
We identified 1895 initial transplant hospitalization records for patients who received
HSCT with a principal diagnosis for NHL. A small number of observations were excluded
because of age restrictions (n = 32), cord blood stem cell transplant (n = 25), or
unspecified source of their bone marrow transplant (n = 6). Baseline characteristics are
described in Table 1.
During the three-year study period, there were 1426 autologous and 406
allogeneic HSCT hospitalizations. The number of discharge records for patients
13

undergoing autologous and allogeneic HSCT for NHL were consistent with the number of
patients receiving autologous and allogeneic HSCT for NHL reported by the US
Transplant data and the Center for International Blood & Marrow Transplant Research
(CIBMTR) after adjusting for the 20% random sampling in the NIS database. [22]Mean
hospital costs for each complication are reported in Table 2. Approximately 21% of
autologous HSCT and 27% of allogeneic HSCT patients did not have any documented
complications during their transplant hospitalization and had substantially lower hospital
costs. During autologous HSCT, mucositis and febrile neutropenia were the most
common adverse events. Severe sepsis, acute respiratory failure, intubation or
mechanical ventilation, total parenteral nutrition (TPN), and death more than doubled the
hospital costs. During allogeneic HSCT, each complication led to higher mean hospital
costs. The most frequently encountered complications were mucositis, febrile
neutropenia, and bacteremia with acute graft-versus-host disease (aGVHD) being the
sixth common adverse event. 1% and 7% of autologous and allogeneic HSCT transplant
hospitalizations ended in death and their mean hospital costs were $121,514 and
$314,513, respectively.
Table 3 shows the impact of each complication on hospital cost and LOS during
autologous HSCT. Most adverse events were associated with significantly higher hospital
costs and longer LOS. Clostridium difficile and bacteremia prolonged LOS but not hospital
costs. During allogeneic HSCT (Table 4), aGVHD, acute renal failure, TPN, and inpatient
death were upward drivers of both hospital costs and LOS. Febrile neutropenia prolonged
LOS but did not increase hospital costs. The utilization of allogeneic grafts from bone
marrow was associated with higher costs and LOS; this is likely due to delayed time to
14

engraftment since bone marrow grafts have a lower hematopoietic stem cell content
compared to G-CSF mobilized peripheral blood concentrates. [23]

Discussion
Our study provides updates on population-based estimates on the complication
rates and their impact on the hospital costs during initial transplant hospitalization for
NHL. The mean hospital cost for autologous HSCT increased from $58,673 in 2003 to
$69,140 in 2014, a relatively small increase considering the 45% increase in the
consumer price index for medical care during the same period. [11,24]
With respect to the complications posed by autologous HSCT, the rates of febrile
neutropenia, infection, intubation, and TPN remained comparable to previous estimates
by Jones et al., but stomatitis was more frequently documented in our study (49% vs.
40%). [11] Nonetheless, inpatient mortality during autologous HSCT declined from 4.4%
to 1%. Most adverse events previously documented by Jones et al. still cause significant
morbidity and mortality during HSCT, and lessons from the previous study still resonate;
strategies to minimize intermediate outcomes such as stomatitis and infection can
substantially reduce hospital costs by preventing subsequent, expensive downstream
events such as organ failure, intubation, TPN, and death.
The mean LOS for autologous HSCT in our study was also not significantly
different from the previous study by Jones et al. We hypothesize that the LOS remained
stable due to waiting for hematopoietic cell count recovery after preparative
chemotherapy and stem cell transfusion. The median time for platelet engraftment after
autologous HSCT is about 3 weeks. [25] This inherent biological constraint may be a
15

barrier to further reduction in LOS and the hospital cost during transplant hospitalization,
but outpatient HSCT and early discharge may be an option for select patients in good
health with caregivers at home. [26]
Like previous studies, patient’s age, gender, and CCI didn’t show statistically
significant association with hospital costs nor LOS in our models when relevant variables
were controlled. [1,27] It is possible that the associated myeloablative and reduced-
intensity chemotherapy regimens for either transplantation modality selected out elderly
patients and patients with advanced co-morbidities. These patient characteristics may
have had an impact on hospital cost and LOS by influencing rates of adverse events.
Hospitals in the Midwest and the South had similar LOS but lower hospital costs
than hospitals in the Northeast. Government-owned hospitals had higher hospital costs
and longer LOS than private, nonprofit hospitals, perhaps suggesting an unobserved
heterogeneity in the patient population or treatment patterns across different hospital
settings.
Our study has its limitations. It only captured the hospital cost for the actual
transplantation procedure, but not the entire episode of care that may begin before the
hospitalization and continue afterward. This fact likely accounts for the lower complication
rates observed in allogeneic HSCT compared to autologous HSCT. One of the most
ubiquitous allogenic HSCT complications is aGVHD, and past clinical trials investigating
allogeneic HSCT in NHL have reported the incidence rates of at least 40% for grade 2–4
aGVHD during the first 100 days following HSCT. [28] However, only 14% of the
allogeneic HSCT patients in our study had documented aGVHD. It is highly plausible that
there were more cases of aGVHD among the allogeneic HSCT patients in our study, but
16

the event may have occurred after hospital discharge upon successful stem cell
engraftment. On the other hand, adverse events during autologous HSCT, like sepsis and
mucositis, were more likely captured in our study because they occur early in the inpatient
course when the patient is immunocompromised and waiting for stem cell engraftment.
In addition, because NIS provides cross-sectional data and documents only major
services and procedures, patients can only be followed for each individual hospitalization,
not through their entire episode of care. Therefore, the cumulative cost estimates in our
study reflect only the costs for the initial transplant hospitalization, exclusive of the costs
of outpatient care, subsequent hospitalizations, and post-discharge medication use. A
recent study based on a claims database analysis that captured downstream costs for
the first 100-days following HSCT reports that the average total health care costs were
$164,049 in patients undergoing autologous HSCT with a myeloablative regimen (2014
US$). [7] For allogeneic HSCT, the first 100-day costs were $401,566 for the
myeloablative regimen and $300,871 for the reduced-intensity regimen. The higher cost
of allogeneic HSCT was due to higher costs for the initial transplant hospitalization,
subsequent hospitalizations, outpatient care, and medications.
The use of ICD-9 codes to capture clinical and economic outcomes represents
another common limitation for retrospective database studies like ours. ICD-9 codes may
not accurately reflect treatment characteristics or adverse events that transpired because
these codes are primarily used for reimbursement and the complexity and severity of
clinical cases may be inappropriately up-coded to maximize reimbursement. [29] In fact,
HSCT inherently entails neutropenia, anemia, thrombocytopenia, and deficiencies in such
blood cell lines are often not comprehensively coded.
17

There is also considerable overlap between sepsis, infection, bacteremia, and
neutropenic fever, and these conditions may not be fully captured in the records. This
may also explain why the percentage of patients without complication was paradoxically
less among allogeneic than in autologous transplant patients. Furthermore, several
studies have documented life-threatening complications such as sinusoidal obstruction
syndrome during HSCT, but there is no single ICD-9 code available to capture such
events. [30]
Lastly, our study is unable to characterize the type NHL that was being treated by
autologous or allogeneic HSCT. NHL is a heterogeneous disease that includes indolent,
slow-growing forms like follicular and mantle cell lymphomas and aggressive subtypes
like Burkitt’s and diffuse large B cell lymphomas. Indolent lymphomas are often managed
with active surveillance and may not warrant treatment until patients become symptomatic.
[31] As a result, it remains unclear how heavily pretreated our NHL population was. It is
conceivable that many of the allogeneic HSCTs were performed in patients who had
relapsed after first-line autologous transplantation.
Lymphomatous bone marrow involvement was also not elucidated in the study
population; any such involvement would have precluded patients from initial consideration
for autologous HSCT, steering them into the allogeneic transplantation route. Allogeneic
HSCT is not just a costly substitute for autologous transplantation. It may actually be the
only alternative treatment to autologous transplantation when autologous transplantation
is contraindicated or not feasible. The cost of allogeneic HSCT may then be higher, but
the patient population receiving allogeneic HSCT is also different. Thus, autologous
HSCT cannot simply be interpreted as the more cost-effective treatment.
18

Such shortcomings withstanding, our study documents the rates of common
complications and their impact on the hospital cost and LOS for autologous and
allogeneic HSCT in NHL. The cost of autologous HSCT may continue to decline in
inflation-adjusted terms as the modality is increasingly used in the outpatient setting, as
has been the case for multiple myeloma. [32]
Moreover, novel cell-based treatments like chimeric antigen receptor (CAR) T-cell
therapy have shown promising survival rates and it will be interesting to see what the
impact of newer treatments will have on the economic and clinical outcomes of NHL. The
development of a fiscal composite for the index hospitalization and complications
observed with stem cell transplantation in NHL may provide a broader framework for
evaluating the clinical and economic value of cellular-based therapeutics and other
innovations in comparison to HSCT. [33]

19

Table 1. Baseline Characteristics

Autologous Allogeneic

(n=1426) (n=406)
Patient Characteristics
Length of stay, day, mean
(Standard deviation)
20.9
(9.2)
26.5
(14.3)
Hospital cost, 2014$
Mean
Standard deviation

$ 69,140
(43,086)

$ 118,378
(94,882)
Median
10
th
and 90
th
percentiles
$ 60,202
$ 31,392 - $ 113,061
$ 94,871
$ 43,983 - $ 222,537
Age, mean 57.0 52.5
Age

<45 208 15% 82 20%
45/54 265 19% 125 31%
55/64 538 38% 141 35%
>65= 415 29% 58 14%
Gender

Men 936 66% 255 63%
Women 490 34% 151 37%
Race

White 1018 71% 306 75%
Black 94 7% 23 6%
Hispanic 101 7% 26 6%
Asian 44 3% 8 2%
Others/Missing 169 12% 43 11%
Charlson Comorbidity Index
CCI = 2 980 69% 290 71%
CCI >=3 446 31% 116 29%
Treatment Characteristics
Stem Cell Source
Peripheral blood (PB) 1364 96% 374 92%
Bone Marrow (BM) 62 4% 32 8%
Conditioning
No total body irradiation 1398 98% 347 85%
Total body irradiation (TBI) 28 2% 59 15%
Hospital Characteristics
Region
Northeast 318 22% 119 29%
Midwest 403 28% 96 24%
South 409 29% 122 30%
West 296 21% 69 17%
Bed size

Large 1111 78% 312 77%
Medium 80 6% 20 5%
Small 235 16% 74 18%
Ownership

Private/non-profit 1030 72% 286 70%
Private/for-profit 84 6% 23 6%
Government 312 22% 97 24%

20

Table 2. Complication Rates and Unadjusted Mean Hospital Costs

Autologous Allogeneic

(n=1426) (n=406)
Event % Cost (SD) % Cost
No adverse event 21% $ 50,610 (26292) 27% $ 74,281 (56415)
Any adverse event 79% $ 73,952 (45249) 73% $ 134,562 (100884)
Febrile neutropenia 40% $ 72,694 (33206) 31% $ 116,154 (63591)
Bacteremia 19% $ 84,286 (62787) 18% $ 178,470 (145432)
C. Difficile 8% $ 78,152 (50747) 8% $ 112,116 (57776)
Pneumonia 6% $ 97,681 (59985) 11% $ 198,245 (113690)
Fungal infection 6% $ 81,049 (41462) 5% $ 202,206 (138340)
Severe sepsis/shock 4% $ 124,374 (98807) 5% $ 300,546 (195792)
GVHD N/A N/A 14% $ 205,342 (145483)
Intubation/ mech. vent 3% $ 134,589 (105648) 8% $ 289,352 (164441)
Acute respiratory failure 4% $ 125,313 (96776) 9% $ 267,184 (168447)
Acute renal failure 9% $ 100,898 (62985) 17% $ 203,659 (151097)
Acute heart failure 3% $ 70,918 (43764) 4% $ 167,209 (129149)
Mucositis 49% $ 72,879 (37404) 43% $ 124,498 (68454)
TPN 9% $ 108,492 (79196) 16% $ 187,818 (137307)
Death 1% $ 121,514 (74672) 7% $ 314,513 (173124)
GVHD: graft-versus-host disease; TPN: total parenteral nutrition
* N/A = Not applicable.

21

Table 3. Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Autologous HSCT
Hospital cost ($) Length of stay
Variable Coefficient P-value Coefficient P-value
PB vs BM -3779 0.309 -0.660 0.326
TBI -7752 0.052 -3.374 0.003**
Febrile neutropenia 4322 0.020* 0.935 0.010**
Bacteremia 3589 0.171 2.943 <0.001**
C. Difficile 5482 0.166 1.659 0.018**
Pneumonia 13429 0.022* 4.576 0.002**
Fungal infection 8166 0.056 1.480 0.095
Severe sepsis/shock 19233 0.281 6.393 0.232
Intubation/mech.
ventilation
46897 0.002** 6.488 0.003**
Acute respiratory failure 15384 0.306 2.346 0.419
Acute renal failure 15750 0.004** 3.355 0.009**
Acute heart failure -5103 0.290 0.642 0.543
Mucositis 6617 0.001** 1.416 0.003**
TPN 29462 <0.001** 6.196 0.002**
Death -25611 0.234 -9.748 0.025**
_constant 61154 <0.001** 19.025 <0.001**
PB: stem cell sourced from peripheral blood; BM: stem cell sourced from bone marrow; TBI:
total body irradiation
* Statistically significant at p-value < 0.05
** Statistically significant at p-value < 0.01
† This table shows an incremental increase in hospital cost and length of stay for each
independent variable after controlling for other variables. Adjusted R-squared are 0.2946 and
0.2545 for cost and length of stay models, respectively.
† † The full table with patient and hospital characteristics is listed in the supplemental
appendix.

22

Table 4. Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Allogeneic HSCT
Hospital cost ($) Length of stay
Variable Coefficient P-value Coefficient P-value
PB vs BM 40687 0.016** 7.087 0.003**
TBI -7093 0.505 1.868 0.260
Febrile neutropenia 8676 0.190 3.185 0.004**
Bacteremia -5402 0.580 1.398 0.348
C. Difficile -7040 0.503 -0.162 0.920
Pneumonia 8068 0.650 4.446 0.091
Fungal infection 15479 0.500 2.123 0.557
Severe sepsis/shock 39534 0.246 5.826 0.312
GVHD 42990 0.002** 10.126 <0.001**
Intubation/mech.
Ventilation
23129 0.345 0.914 0.843
Acute respiratory failure 33953 0.114 -0.460 0.911
Acute renal failure 34836 0.002** 7.315 <0.001**
Acute heart failure -3263 0.902 -1.017 0.733
Mucositis -4259 0.548 0.802 0.432
TPN 33372 0.016** 4.193 0.017**
Death 95835 0.004** 12.342 0.038*
_constant 74919 <0.001** 18.279 <0.001**
PB: stem cell sourced from peripheral blood; BM: stem cell sourced from bone marrow; TBI:
total body irradiation
* Statistically significant at p-value < 0.05
** Statistically significant at p-value < 0.01
†This table shows an incremental increase in hospital cost and length of stay for each
independent variable after controlling for other variables. Adjusted R-squared are 0.5131
and 0.4531 for cost and length of stay models, respectively.
††The full table with patient and hospital characteristics is listed in the supplemental
appendix.

23

References

1. Khera N, Zeliadt SB, Lee SJ. Economics of Hematopoietic Cell Transplantation.
Blood. 2012 Aug 23; 120(8): 1545-51
2. Reddy N, Oluwole O, Greer JP, Engelhardt BG, Jagasia MH, Savani BN.
Outcomes of Autologous and Allogeneic Stem Cell Transplantation for Non-
Hodgkin’s Lymphoma. Exp Hematol. 2014 Jan;42(1):39-45.
3. Kuruvilla J. The Role of Autologous and Allogeneic Stem Cell Transplantation in
the Management of Indolent B-Cell Lymphoma. Blood. 2016 April
28;127(17):2093-100
4. Epperia N, Hamadani M. Hematopoietic Cell Transplantation for Diffuse Large B-
Cell and Follicular Lymphoma: Current Controversies and Advances. Hematol
Oncol Cell Ther. 2017 Dec; 10(4):277-284
5. High-Dose Chemotherapy and Stem Cell Transplant for Non-Hodgkin’s
Lymphoma. American Cancer Society. Available from
https://www.cancer.org/cancer/non-hodgkin-lymphoma/treating/bone-marrow-
stem-cell.html [Accessed January 2018]
6. Niederwieser D, Baldomero H, Szer J, Gratwohl M, Aljurf M, Atsuta Y, et al.
Hematopoietic stem cell transplantation activity worldwide in 2012 and a SWOT
analysis of the Worldwide Network for Blood and Marrow Transplantation Group
including the global survey. Bone Marrow Transplant. 2016 Jun;51(6):778-85.
7. Broder M, Quock T, Change E, Reddy S, Agarwal-Hashimi R, Arai S, et al. The
Cost of Hematopoietic Stem-Cell Transplantation in the United States. Am Health
Drug Benefits. 2017 Oct; 10(7): 366–374.
24

8. Majhail NS, Mau LW, Denzen EM, Arneson TJ. Costs of Autologous and
Allogeneic Hematopoietic Cell Transplantation in the United States: A Study
Using a Large National Private Claims Database. Bone Marrow Transplant. 2013
Feb; 48(2):294-300.
9. Saito AM, Cutler C, Zahrieh D, Soiffer R, Ho V, Alyea E, et al. Costs of allogeneic
hematopoietic cell transplantation with high-dose regimens. Biol Blood Marrow
Transplant. 2008 Feb; 14(2): 197–207.
10. Svahn BM, Remberger M, Alvin O, Karlsson H, Ringden O. Increased costs after
allogeneic haematopoietic SCT are associated with major complications and re-
transplantation. Bone Marrow Transplant. 2012 May;47(5):706-15
11. Majhail NS, Mothukuri JM, Brunstein CG, Weisdorf DJ. Costs of hematopoietic
cell transplantation: comparison of umbilical cord blood and matched related
donor transplantation and the impact of posttransplant complications. Biol Blood
Marrow Transplant. 2009 May;15(5):564-73
12. Jones JA, Qazilbash MH, Shih YC, Cantor SB, Cooksley CD, Elting LS. In-
Hospital Complications of Autologous Hematopoietic Stem Cell Transplantation
for Lymphoid Malignancies: Clinical and Economic Outcomes from the National
Inpatient Sample. Cancer 2008 Mar; 112(5):1096-105
13. Healthcare Cost and Utilization Project (HCUP). The National Inpatient Sample
Overview. Available from https://www.hcup-us.ahrq.gov/nisoverview.jsp
[Accessed April 2018]
25

14. Healthcare Cost and Utilization Project (HCUP). HCUP Nationwide Inpatient
Sample (NIS). Available from https://www.hcup-
us.ahrq.gov/nisoverview.jsp#multi [Accessed January 2018]
15. Milano F, Gooley T, Wood B, Woolfrey A, Flowers M, Doney K, et al. Cord-Blood
Transplantation in Patients with Minimal Residual Disease. N Engl J Med. 2016
Sep 8;375(10):944-53
16. Vicki Stagg, 2006. "CHARLSON: Stata module to calculate Charlson index of
comorbidity," Statistical Software
17. Cost-to-Charge Ratio Files. Healthcare Cost and Utilization Project (HCUP).
September 2018. Agency for Healthcare Research and Quality, Rockville, MD.
Available from www.hcup-us.ahrq.gov/db/state/costtocharge.jsp.
18. Bureau of Labor Statistics. United States Department of Labor. Available from
https://www.bls.gov/data/ [Accessed January 2018]
19. Fisher ES, Bynum JP, Skinner JS. Slowing the growth of health care costs--
lessons from regional variation. N Engl J Med. 2009 Feb 26;360(9):849-52.
20. Sloan FA, Picone GA, Taylor DH, Chou SY. Hospital Ownership and Cost and
Quality of Care: Is There a Dime's Worth of Difference? J Health Econ. 2001
Jan;20(1):1-21.
21. Tompkin CP, Altman SH, Eliat E. The precarious pricing system for hospital
services. Health Aff (Millwood). 2006 Jan-Feb;25(1):45-56.
22. U.S. Transplant Data by Disease Report. Health Resources and Services
Administration. U.S. Department of Health and Human Services.
26

https://bloodcell.transplant.hrsa.gov/research/transplant_data/us_tx_data/data_b
y_disease/national.aspx [Accessed August 2018]
23. Consumer Price Index in Medical Care. Bureau of Labor Statistics. United States
Department of Labor. Available from https://data.bls.gov/cgi-bin/surveymost?cu
[Accessed April 2018]
24. Stiff PJ, Micallef I, Nademanee AP, Stadtmaur EA, Maziarz RT, Bolwell BJ, et al.
Transplanted CD34(+) cell dose is associated with long-term platelet count
recovery following autologous peripheral blood stem cell transplant in patients
with non-Hodgkin lymphoma or multiple myeloma. Biol Blood Marrow Transplant.
2011 Aug;17(8): 1146-53
25. Graff T, Singavi A, Schmidt W, Eastwood D, Drobyski W, Horowitz M, et al.
Safety of Outpatient Autologous Hematopoietic Cell Transplantation for Multiple
Myeloma and Lymphoma. Bone Marrow Transplant. 2015 Jul; 50(7): 947–953.
26. Lee SJ, Klar N, Weeks JC, Antin JH. Predicting costs of stem-cell
transplantation. J Clin Oncol. 2000;18(1):64-71.
27. Glass B, Hasenkamp J, Wulf G, Dreger P, Pfreundschuh M, Gramatzki M, Silling
G et al. Rituximab after lymphoma-directed conditioning and allogeneic stem-cell
transplantation for relapsed and refractory aggressive non-Hodgkin lymphoma
(DSHNHL R3): an open-label, randomised, phase 2 trial. Lancet Oncol. 2014
Jun;15(7):757-66.
28. Iezzoni L. Risk Adjustment for Measuring Health Care Outcomes. 3rd ed.
Chicago, IL. Health Administration Press; 2003.
27

29. Cao Z, Villa KF, Lipkin CB, Robinson SB, Nejadnik B, Dvorak CC. Burden of
illness associated with sinusoidal obstruction syndrome/veno-occlusive disease
in patients with hematopoietic stem cell transplantation. J Med Econ. 2017
Aug;20(8):871-883.
30. Cheah CY, Seymour JF, Wang ML. Mantle Cell Lymphoma. J Clin Oncol. 2016
Apr 10;34(11):1256-69.
31. Gertz MA, Ansell SM, Dingli D, Dispenzieri A, Buadi FK, Elliott MA, et al.
Autologous stem cell transplant in 716 patients with multiple myeloma: low
treatment-related mortality, feasibility of outpatient transplant, and effect of a
multidisciplinary quality initiative. Mayo Clin Proc. 2008 Oct;83(10): 1131-8.
32. Neelapu SS, Locke FL, Bartlett NL, Lekakis LJ, Miklos DB, Jacobson CA, et al.
Axicabtagene Ciloleucel CAR T-Cell Therapy in Refractory Large B-Cell
Lymphoma. N Engl J Med. 2017 Dec 28;377(26):2531-2544.
33. Schuster SJ, Svoboda J, Chong EA, Nasta SD, Mato AR, Anak O, et al. Chimeric
Antigen Receptor T Cells in Refractory B-Cell Lymphomas. N Engl J Med. 2017
Dec 28;377(26):2545-2554.
28

Chapter 2 – Cost-effectiveness Analysis of Regorafenib and TAS-102 in
Refractory Metastatic Colorectal Cancer in the United States
Sang Kyu Cho, Joel Hay, and Afraseh Barzi.
Clinical Colorectal Cancer. 2018 Dec;17(4):e751-e761.

Abstract

Background: Regorafenib and TAS-102 are standard treatment options in refractory
metastatic colorectal cancer based on improvement in overall survival by 6 and 8 weeks,
respectively, when compared with best supportive care alone (BSC). Given the small
incremental clinical benefit, we evaluated their cost-effectiveness from a United States
payer’s perspective.
Materials and Methods: A Markov model was constructed to compare costs and
effectiveness of regorafenib, TAS-102, and BSC. Model inputs for clinical efficacy and
adverse events were from the CORRECT trial (Regorafenib monotherapy for previously
treated metastatic colorectal cancer: an international, multicentre, randomised, placebo-
controlled, phase 3 trial) for regorafenib and the RECOURSE trial (Randomized, Double
Blind, Phase 3 Study of TAS-102 plus BSC versus Placebo plus BSC in Patients with
Metastatic Colorectal Cancer Refractory to Standard Chemotherapies) for TAS-102. The
incremental cost-effectiveness ratios (ICERs) were reported to compare treatment
outcomes. Model robustness was checked with univariate and probabilistic sensitivity
analyses as well as a scenario analysis using the CONCUR trial data for regorafenib.
29

Results: In our base case, regorafenib and TAS-102 had the ICERs of $395,223 per
quality-adjusted life year (QALY) and $399,740 per QALY versus BSC, respectively.
Compared with regorafenib, TAS-102 provided an additional 0.041 QALY at the cost of
$16,608 or $406,104 per QALY, but the differences were not robust in sensitivity analyses.
The most influential parameters on the ICERs were efficacy and health state utility
parameters as well as the cost of treating neutropenia. In probabilistic sensitivity analysis
using cost-effectiveness acceptability curves, BSC was more cost-effective than both
regorafenib and TAS-102 in 50% of repetitions at the willingness-to-pay threshold of
$330,000 per QALY.
Conclusion: Neither TAS-102 nor regorafenib are cost-effective at standard willingness-
to-pay thresholds (i.e., $150,000 per QALY) relative to BSC. There is no clear evidence
that either treatment has better relative value.

30

Background
In the United States, 135,430 new cases of colorectal cancers were diagnosed in 2017
with estimated 50,260 deaths. [1] Stage at diagnosis determines treatment options and
is a predictor of survival. Although the 5-year survival rate for localized colorectal cancer
is about 90%, it declines to 13.9% during the metastatic stage (mCRC). [1] In mCRC,
fluorouracil and leucovorin combined with oxaliplatin or irinotecan significantly improve
survival, and further improvement can be achieved using monoclonal antibodies such as
bevacizumab, cetuximab, or panitumumab. [2]
For patients who have disease progression after these agents and still maintain a
good performance status, regorafenib or TAS-102 (trifluridine and tipiracil) can be used
to prolong overall survival. In the RECOURSE trial (Randomized, Double Blind, Phase 3
Study of TAS-102 plus Best Supportive Care [BSC] versus Placebo plus BSC in Patients
with Metastatic Colorectal Cancer Refractory to Standard Chemotherapies), TAS-102
demonstrated median overall survival of 7.1 months versus 5.3 months for best
supportive care alone (BSC). [3] In the CORRECT trial (Regorafenib monotherapy for
previously treated metastatic colorectal cancer: an international, multicentre, randomised,
placebo-controlled, phase 3 trial), regorafenib had median overall survival of 6.4 months
versus 5.0 months for BSC. [4] The National Comprehensive Cancer Network (NCCN)
recognizes both regorafenib and TAS-102 as third- and fourth-line treatment options in
appropriate patient populations. [5]
Diverse value frameworks have assessed the use of regorafenib and TAS-102 in
mCRC to inform patients, physicians, and payers of the drugs’ clinical benefit, toxicity,
and cost. Under NCCN Evidence Blocks, which uses a five point scale, both drugs have
31

the same scores in terms of efficacy, consistency of evidence, and affordability, but TAS-
102 scores higher for safety and quality of evidence as a third-line therapy for mCRC. [6]
Similarly, Becker et al. used the value frameworks proposed by American Society of
Clinical Oncology and European Society for Medical Oncology to evaluate the drugs, and
TAS-102 was favored over regorafenib in most medication benefit score metrics. [7]
To date, two cost-effectiveness analysis studies have compared the two drugs.
Bullement et al. assessed the cost-effectiveness of regorafenib versus TAS-102 in
England and Wales and found that TAS-102 was dominant over regorafenib, providing
higher quality-adjusted life years at a lower cost[8] Kimura et al. conducted a similar cost-
effectiveness analysis study for Japan and also found that TAS-102 was more cost-
effective than regorafenib. [9]
While both studies relied on efficacy data (i.e., progression free and overall survival)
from each drug’s phase III clinical trials, in which TAS-102 appeared to have greater
median overall survival than regorafenib by approximately 2 weeks, the real-world
evidence shows comparable effectiveness between the drugs. [10,11] In addition, the
previous CEA studies used health state utility scores from the CORRECT trial, which was
later criticized by the Institute for Quality and Efficiency in Health Care in Germany (IQWiG)
for inadequate data analysis. [12] These new findings are worthy of further discussion
because even a small change in efficacy or HRQoL parameters can produce directionally
different results when the incremental quality-adjusted life year (QALY) is expected to be
small between comparators. [13]
32

Given these limitations, our study evaluated the cost-effectiveness of regorafenib
and TAS-102 from a United States (US) payer’s perspective and explored the impact of
uncertainty in influential parameters on their relative cost-effectiveness.

Materials and Methods
Model Structure
Our study used a Markov model with three health states. Patients started in the stable
disease state and moved to the disease progression state over time. Death could occur
in both states. Regorafenib and TAS-102 were administered until disease progression.
All patients received BSC during the disease progression state. Figure 1 shows the model
structure.
Each Markov cycle represented thirty days. Both regorafenib and TAS-102 have
28-day dosing schedule but dose interruptions and delays frequently occurred in both the
RECOURSE and CORRECT trials owing to adverse events. QALYs and costs associated
with treatments were recorded for each cycle. Our study used a time horizon of five years
to follow outcomes until more than 99% of patients had died in all treatment arms. An
annual discount rate of 5% was applied to both costs and QALYs. [14]

Clinical Efficacy
Clinical efficacy data were obtained from the CORRECT trial for regorafenib and from the
RECOURSE trial for TAS-102 and BSC. [3,4] Using the DEALE method (Declining
33

Exponential Approximation of Life Expectancy), median overall survival and progression-
free survival were converted into rates and respective transition probabilities. [15]

Drugs
In the CORRECT trial, the mean daily dose of regorafenib was 147 mg. In the
RECOURSE trial, the mean dose of TAS-102 was 155 mg per square meter per week,
and patients’ body surface area averaged 1.78 m
2
. The calculated TAS-102 dose during
the RECOURSE trial was approximately 55 mg per dose. In our model, each dose was
160 mg for regorafenib and 60 mg for TAS-102 after rounding up to the nearest available
unit of strength. Drug costs were obtained from the US Veterans Affairs Federal Supply
Schedule. [16] Any unused medications were assumed to be wasted.

Adverse Events
Grade 3 or 4 adverse events with incidence rates greater than 1% were included in the
model. Using Medicare severity diagnosis-related groups (MS-DRG), mean hospital costs
of each adverse event were obtained from the Agency for Healthcare Research and
Quality HCUPnet. [17] The costs of managing adverse events were applied once in the
first cycle.
For routine care, patients were assumed to have biweekly doctor visits as well as
complete blood count with differentials and comprehensive metabolic panels at each visit.
Their medical costs were obtained from the Center for Medicare and Medicaid Services
34

using Current Procedural Terminology codes. MSDRGs used in our model are listed in
Supplemental Tables 1 and 2 (in the online version). All costs were inflated to 2017 dollars
using Consumer Price Index for Medical Care. [18]

Health State Utility
In the CORRECT trial, EQ-5D scores were used to measure health state utility. In the
stable disease state, mean EQ-5D scores were 0.73 in the regorafenib group and 0.74 in
the BSC group. During the progression state, both groups had mean EQ-5D score of 0.59.
These values were collected during the trial, and they accounted for disutility from
adverse events.
The RECOURSE trial did not report health state utility. Therefore, the TAS-102
group was assumed to have the same health state utility as patients receiving regorafenib.
Implications of this assumption are addressed in the sensitivity analysis and discussion.

Sensitivity Analyses
In univariate sensitivity analysis, the parameters with the greatest influence on the
incremental net monetary benefit (NMB) at a willingness-to-pay of $150,000 were
presented using a tornado diagram. When the incremental QALY between two drugs is
expected to be small, the incremental NMB provides a more tractable way for comparison
than the ICER. Both NMB and ICER provide the same answer in terms of relative cost-
effectiveness between comparators at a given threshold. [19]
35

Neutropenia was the most prevalent and expensive adverse event in our model.
Although it is reasonable to assume hospitalization for the treatment of neutropenia,
neutropenia could also be treated at an outpatient setting, and its incidence could be
reduced by prophylactic administration of a broad-spectrum antibiotic and a granulocyte
colony stimulating factor (G-CSF). [20,21] To explore its impact on the incremental NMB,
the cost of neutropenia was varied by increasing and decreasing its value by 10%, 20%,
30%, and 40%. The other parameters in the model were varied by 10%.
In probabilistic sensitivity analysis, each parameter was randomly sampled from
assumed distributions, and their simultaneous changes were recorded in 10,000 Monte
Carlo simulations. Dose for TAS-102 was randomly sampled from the body surface area
distribution of patients in the RECOURSE trial. Distributions and standard errors of all
parameters varied in the probabilistic sensitivity analysis are listed in Supplemental
Tables 3 and 4.
Efficacy parameters were particularly influential in our study owing to the small
incremental QALYs between comparators. Although our base case assumed slightly
higher median overall survival of TAS-102 than regorafenib based on each drug’s
respective phase III clinical trial data, there is no clear clinical consensus or evidence to
support the use of one drug over the other. Therefore, we conducted a scenario analysis
to consider the opposite case in which regorafenib has slightly higher median overall
survival than TAS-102. This scenario analysis used efficacy data from the CONCUR trial,
which showed 8.8 and 3.2 months for median overall survival and progression-free
survival, respectively. [22] All analyses were performed using Microsoft Excel and Visual
Basic for Applications.
36

Results
Base Case Results
The base case results are presented in Tables 1 and 2. Compared with BSC, TAS-102
had the ICER of $399,740 per QALY, and regorafenib had the ICER of $395,223 per
QALY. Neither TAS102 nor regorafenib was cost-effective versus BSC at the typical US
willingness-to-pay threshold of $150,000 per QALY. [23] Compared with regorafenib,
TAS-102 provided an incremental QALY of 0.041 at the cost of $16,608 or the ICER of
$406,104 per QALY.

Sensitivity Analyses
The tornado diagram in Figure 2 shows the most influential parameters on the NMB
between TAS-102 and regorafenib. The incremental NMB was predominantly influenced
by health state utility measured by EQ-5D scores, drug’s efficacy, and the cost of treating
neutropenia. The cost of hospitalization for the other adverse events had negligible impact
on determining relative cost-effectiveness.
Using the cost-effectiveness acceptability curves, the result of the probabilistic
sensitivity analysis is presented in Figure 3. At a willingness-to-pay below $330,000 per
QALY, BSC was more cost-effective than both regorafenib and TAS-102. However, TAS-
102 became the most cost-effective treatment with increasing willingness-to-pay
threshold.
37

In the scenario analysis, regorafenib was dominant over TAS-102, producing
higher QALY at a lower cost as shown in Table 3. The results of the sensitivity analyses
show that, as uncertainty remains around the influential model parameters, there is no
clear evidence that either regorafenib or TAS-102 has better relative value.

Discussion
Colorectal cancer remains a significant cause of cancer-related mortality in the US. Most
of the patients with mCRC succumb to the disease. In the advanced setting, regorafenib
and TAS-102 are the two US Food and Drug Administration-approved agents. Both
agents result in comparable improvement in progression free and overall survival
compared with BSC. However, this marginal improvement comes at a significant cost.
Given the significance of financial burden in patients with colorectal cancer,
consideration of cost in treatment decision is important and worthy of evaluation and
discussion. Cost-effective analysis allows a comprehensive assessment of benefits,
toxicities, and costs of care under best current estimates of model parameters and
provides assessment of model uncertainty.
Regorafenib and TAS-102 have similar ICERs versus BSC, and the choice
between the two drugs should be determined considering each patient’s preferences,
toxicity profile, goals of care, and quality of life. The RECOURSE trial reported that the
clinical benefit associated with TAS-102 was maintained irrespective of prior treatment
with regorafenib. Sueda et al. reported that patients who had crossover treatment
38

between regorafenib and TAS-102 had greater overall survival than patients who received
only one of the two drugs. [11]
Our study is not without limitations. We used the DEALE method to fit survival data,
but other distributional assumptions such as Weibull may be used if better clinical trial
data become available. However, it is unlikely that such assumption would change this
study’s conclusion because of the comparable effectiveness between regorafenib and
TAS-102. Our study also assumed hospitalization for all grade 3 and 4 adverse events.
In practice, cost and incidence of adverse events could vary considerably based on the
patient’s health status and the use of prophylactic treatment. If the cost of treating
neutropenia is reduced by 40%, TAS-102 will have the ICER of $313,355 per QALY
versus regorafenib compared with the ICER of $406,104 per QALY in the base case. Also,
our study did not include the cost of G-CSF, but 9.4% of patients in the RECOURSE trial
utilized G-CSF. Its cost ranges between $4,381 and $6,653, depending on which agent
is used. [24] Additionally, estimation of the cost of toxicities is another important limitation
in our study. Although economic and clinical impact of most toxicities were previously
studied, there was paucity of robust evidence on toxicities like hand-foot syndrome, and
this may have an impact on our findings. Nevertheless, efforts to mitigate the toxicities of
regorafenib is a focus of ongoing clinical trials. [25]
Maintenance of health state utility is an important aspect of cancer treatment.
Although our study used health state utility scores from the CORRECT trial, it is
controversial whether they are in fact similar across the three treatment groups owing to
different toxicity profiles. Our model applied the same health state utility scores for both
regorafenib and TAS-102 because, in its appraisal of TAS-102, the National Institute for
39

Health and Care Excellence discussed health state utility scores between regorafenib and
TAS-102, and stated that there wasn’t enough evidence to show whether regorafenib is
more toxic than TAS-102 or that their health state utility scores are different. [26] However,
this health state utility scores reported by the CORRECT trial were later criticized by
IQWiG for inadequately analyzing the data on patient-reported outcomes on symptoms
and quality of life. In its assessment, IQWiG stated that regorafenib had additional non-
quantifiable negative effects on health state utility relative to best supportive care alone.
[12]
In conclusion, there is no clear evidence that either TAS-102 or regorafenib has
better relative value. Because even a small increase in overall survival or health state
utility score could make regorafenib or TAS-102 drastically more cost-effective versus
BSC, the future avenue of research should include identification of the patients who may
have better response to the treatments. Our study also underscores the importance of
managing adverse events with cancer treatment as it may have substantial impact on
patients’ health state utility as well as the total cost of care.

40

Figure 1. Markov Cycle

Table 1. Total Cost by Treatment
Regorafenib TAS-102 BSC
Drug $ 17,608 $ 21,330 $ -
Adverse Events $ 7,133 $ 19,794 $ 2,295
Routine Care $ 1,915 $ 2,140 $ 1,585
Total Costs $ 26,657 $ 43,264 $ 3,879
Total QALYs 0.397 0.437 0.339

Table 2. Incremental Cost-Effectiveness Ratios
Comparison (ICER)
TAS-102 versus Regorafenib $ 406,104 per QALY
Regorafenib versus BSC $ 395,223 per QALY
TAS-102 versus BSC $ 399,740 per QALY

41

Figure 2. Tornado Diagram

42

Figure 3. Cost-Effectiveness Acceptability Curves

Table 3. Result of Scenario Analysis
Total Cost ($) Total QALY
Regorafenib (CONCUR) $ 36,171 0.513
ICER, Regorafenib vs TAS-102 Regorafenib Dominant
ICER, Regorafenib vs BSC $ 185,332

43

References
1. Siegel R, Miller K, Fedewa S et al. Colorectal Cancer Statistics, 2017. CA Cancer
J Clin 2017; 67:177-193.
2. Metastatic Colorectal Cancer, Union of International Cancer Control, 2014
Review of Cancer Medicine on the WHO List of Essential Medicines. Available
from
http://www.who.int/selection_medicines/committees/expert/20/applications/Metas
taticColorectal.pdf [Accessed November 2017]
3. Mayer RJ, Van Cutsem E, Falcone A et al. Randomized Trial of TAS-102 for
Refractory Metastatic Colorectal Cancer. N Engl J Med 2015; 372(20):1909-19.
4. Grothey A, Van Cutsem E, Sobrero A et al. Regorafenib Monotherapy for
Previously Treated Metastatic Colorectal Cancer (CORRECT): an International,
Multicentre, Randomized, Placebo-controlled Phase 3 Trial. Lancet 2013;
381(9863):303-12.
5. The National Comprehensive Cancer Network Guidelines for Patients; Colon
Cancer Version 1.2017. Available from
https://www.nccn.org/patients/guidelines/colon/files/assets/common/downloads/fil
es/colon.pdf [Accessed November 2017]
6. NCCN Clinical Practice Guideline in Oncology: Colon Cancer NCCN Evidence
Blocks. Version 2.2018. Available from
https://www.nccn.org/professionals/physician_gls/pdf/colon_blocks.pdf
[Accessed July 2018]
44

7. Becker DJ, Lin D, Lee S, et al. Exploration of the ASCO and ESMO Value
Frameworks for Antineoplastic Drugs. J Oncol Pract. 2017 Jul;13(7):e653-e665
8. Bullement A, Underhill S, Fougeray R et al. Cost-Effectiveness of
Trifluridine/tipiracil for Previously Treated Metastatic Colorectal Cancer in
England and Wales. Clin Colorectal Cancer 2017 S1533-0028(16)30287
9. Kimura M, Usami E, Iwai M et al. Comparison of cost-effectiveness of
regorafenib and trifluridine/tipiracil combination tablet for treating advanced and
recurrent colorectal cancer. Mol Clin Oncol. 2016 Nov;5(5):635-640.
10. Moriwaki T, Fukoka S, Taniguchi H et al. Propensity Score Analysis of
Regorafenib Versus Trifluridine/Tipiracil in Patients with Metastatic Colorectal
Cancer Refractory to Standard Chemotherapy (REGOTAS): A Japanese Society
for Cancer of the Colon and Rectum Multicenter Observational Study.
Oncologist. 2018 Jan;23(1):7-15. doi: 10.1634/theoncologist.2017-0275. Epub
2017 Sep 11
11. Sueda T, Sakai D, KudoT et al. Efficacy and Safety of Regorafenib and TAS-102
in Patients with Metastatic Colorectal Cancer Refractory to Standard Therapies.
Anticancer Res. 2016 Aug 36(8);4299-306.
12. [A14-09] Appendum to Comission A13-37 (regorafenib). Institute for Quality and
Efficiency in Health Care. Available from https://www.iqwig.de/en/projects-
results/projects/drug-assessment/a14-09-addendum-to-commission-a13-37-
regorafenib.6022.html [Accessed January 2017]
13. Briggs AH. A Bayesian approach to stochastic cost-effectiveness analysis.
Health Econ. 1999 May;8(3):257-61.
45

14. Smith DH. Gravelle H. The practice of discounting in economic evaluations of
healthcare interventions. Int J Technol Assess Health Care. 2001
Spring;17(2):236-43.
15. Beck JR, Kassirer JP, Pauker SG. A Convenient Approximation of Life
Expectancy (the “DEALE”). I. Validation of the Method. Am J Med 1982;
73(6):883-8.
16. US Department of Veterans Affairs National Acquisition Center Pharmaceutical
Catalogue Search. Available from https://www.va.gov/nac/Pharma/List [Accessed
November 2017]
17. Agency for Healthcare Research and Quality HCUPnet Healthcare Cost and
Utilization Project. Available from https://hcupnet.ahrq.gov/ [Accessed November
2017]
18. US Bureau of Labor Statistics Consumer Price Index – Medical Care in U.S. City
Average, All Urban Consumers, Not Seasonally Adjusted. Available from
https://data.bls.gov/cgi-bin/surveymost [Accessed November 2017]
19. Drummond M, Sculpher M, Claxton K et al. Methods for the Economic Evaluation
of Health Care Programmes. Fourth Edition. Oxford, United Kingdom. Oxford
University Press 2015.
20. Liou SY, Stephens JM, Carpiuc KT et al. Economic Burden of Hematological
Adverse Effects in Cancer Patients: a Systematic Review. Clin Drug Investig
2007; 27(6):381-96
46

21. Bucaneve G, Micozzi A, Menichetti F et al. Levofloxacin to Prevent Bacterial
Infection in Patients with Cancer and Neutropenia. N Engl J Med 2005;
353(10):977-87
22. Neuman PJ, Cohen JT, Weinstein MC. Updating Cost-Effectiveness—Curious
Resilience of the $50,000 per QALY Threshold. N Engl J Med 2014; 371(9):796-
797
23. James E, Trautman H, and Szabo E et al. Comparative Cost-Efficiency Analysis
of Granulocyte Colony-Stimulating Factors for Use in Chemotherapy Patients in
the United States. Presented at the 59
th
American Society of Hematology Annual
Meeting and Exposition. Atlanta, GA. December 9-12, 2017.
24. Bekaii-Saab T, Ou F, Ciombor K et al. Regorafenib Dose Optimization Study
(ReDOS): A Phase II Randomized Study of Lower Starting Dose Regorafenib
Compared to Standard Dose Regorafenib in Patients with Refractory Metastatic
Colorectal Cancer. Presented at Gastrointestinal Cancers Symposium, San
Francisco, CA, January 18-20, 2018
25. Single Technology Appraisal. Committee Paper [ID 876], Trifluridine with Tipiracil
Hydrochloride for Treating Metastatic Colorectal Cancer After Standard Therapy.
National Institute for Health and Care Excellence. Available from
https://www.nice.org.uk/guidance/ta405/documents/committee-papers [Accessed
December 2017]

47

Chapter 3 – Development of a Model to Predict Healing of Chronic Wounds
within Twelve Week
Sang Kyu Cho, Soeren Mattke, Hanna Gordon, Mary Sheridan, and William Ennis
Advances in Wound Care (January 2019). Mary Ann Liebert, Inc., publishers.

Abstract
Objective: Chronic wounds represent a highly prevalent but little recognized condition
with substantial implications for patients and payers. While better wound care products
and treatment modalities are known to improve healing rates, they are inconsistently used
in real-world practice. Predicting healing rates of chronic wounds and comparing to actual
rates could be used to detect and reward better quality of care. We developed a prediction
model for chronic wound healing.
Approach: We analyzed electronic medical records for 620,356 chronic wounds of various
etiologies in 261,398 patients from 532 wound care clinics in the United States. Patient-
level and wound-level parameters influencing wound healing were identified from prior
research and clinician input. Logistic regression and classification tree models to predict
the probability of wound healing within 12 weeks were developed using a random sample
of 70% of the wounds and validated in the remaining data.
Results: 365,659 (58.9%) wounds were healed by week 12. The logistic and classification
tree models predicted healing with an AUC of 0.712 and 0.717, respectively. Wound-level
characteristics, such as location, area, depth, and etiology, were more powerful predictors
than patient demographics and comorbidities.
48

Innovation: The probability of wound healing can be predicted with reasonable accuracy
in real-world data from electronic medical records.
Conclusion: The resulting severity adjustment model can become the basis for
applications like quality measure development, research into clinical practice and
performance-based payment.

49

Introduction
Chronic wounds, usually defined as wounds that do not heal within an expected time
frame, typically 4-12 weeks, are a growing but little recognized public health challenge.
[1,2] Their prevalence is similar to that of heart failure, affecting 6.5 million people or 2%
of the population in the United States. [3] In 2014, estimated Medicare expenditure related
to managing chronic wounds and associated complications ranged between $28 and
$96.8 billion. [4] Patients with chronic wounds often suffer loss of productivity,
psychological distress, decreased quality of life, and reduced life expectancy. [5,6]
From advanced dressings to growth factors and biologics, substantial progress
has been made in product development and treatment approaches for chronic wounds
that do not adequately heal with standard treatment. [7,8] These advanced wound
therapies employ various mechanisms to aid wound healing, such as providing a
structural scaffold, repairing restorative cellular mechanisms, increasing delivery of
oxygen and nutrients, preventing infection, and removing necrotic tissues or exudates. [9]
However, the uptake of these innovations and use of evidence-based treatment practices
remain limited. [1,2,10]
For other chronic conditions, payment reform has been proposed as a policy
intervention to improve quality of care because the established fee-for-service payment
system is increasingly regarded as failing to reward high-quality care. [4] The Centers for
Medicare and Medicaid Services (CMS) has charted an ambitious course to reorient its
payment models to value and has introduced such models for several chronic conditions.
[11] Several so-called Alternative Payment Models, which link reimbursement to overall
resource use, patient outcomes and adherence to clinical practice guidelines, are tested
50

for conditions like heart failure, diabetes, hypertension, and chronic obstructive pulmonary
disease. [12]
A critical requirement for the introduction of such value-based payment models is
scientifically sound and actionable quality indicators that will trigger rewards or penalties.
An obvious choice in chronic wound care is the wound healing rate, as closing wounds is
both the intended outcome of treatment and readily observable. But given the
heterogeneous etiology of chronic wounds, differences in case mix across wound care
centers, and the multitude of factors that can influence healing rates, proper severity
adjustment is needed to construct a valid quality measure.
Several severity adjustment models have been developed previously, but many
suffer from limitations such as small sample size from a single center, applicability to only
one wound type, such as diabetic foot ulcer, and low prediction accuracy. [13-18] More
recent studies leveraged electronic medical records and machine learning techniques to
improve prediction accuracy, but their usability as a quality measure is handicapped by
data-driven variable selection and incorporation of information on initial healing observed
during follow-up visits. [19, 20]
The objective of our study is to develop and validate a unified severity adjustment
model for a broad range of chronic wounds that can form the basis for a quality measure
as it is entirely based on routinely collected data from intake visits at wound care clinics.

51

Materials and Methods
Dataset
The data for our analyses came from the electronic medical records of a large network of
wound care centers in 46 states and cover the timeframe from January 1st, 2014 to
September 25th, 2018. A detailed description of the staffing and skill mix of these centers
has been published elsewhere. [21] In short, these centers are run by a wound care
management company and staffed by combination of employed and contracted
physicians, supported by specialized nurses and program managers. All clinicians are
required to attend a 1-week specialty wound care training course and are provided
evidence based algorithmic clinical practice guidelines. The clinicians subsequently
receive ongoing education, monthly conference calls and have access to regional and
local medical directors for case management support. The centers are hospital-based
and most have access to specialty consultants and advanced treatment modalities, like
hyperbaric oxygen. Standardized treatment protocols promote adherence to evidence-
based care.
The data contain detailed patient level information such as age, sex, smoking
status, body mass index, comorbid conditions, and wound level measurements such as
length, width, and depth as well as categorical descriptors such as wound etiology,
location, and appearance. Each wound is assessed at intake and every subsequent visit,
and treatment modalities are documented. At the end of each visit, the outcome for each
wound is documented as healed or not healed. The data for our analyses came from the
electronic medical records of Healogics’ outpatient wound care clinics in 46 states and
cover the timeframe from January 1
st
, 2014 to September 25
th
, 2018. The data contain
52

detailed patient level information such as age, sex, smoking status, body mass index, and
comorbid conditions, and wound level measurements such as length, width, and depth
as well as categorical descriptors such as wound etiology, stage, location, and
appearance. Each wound is assessed at intake and every subsequent visit, and
treatment modalities are documented. At the end of each visit, the outcome for each
wound is documented as healed or not healed.
Our starting sample consisted of electronic medical records for 356,649 patients
with 914,878 unique wounds. We excluded 202,842 (22%) wounds that were caused by
radiation, acute wounding events such as surgery and trauma, as well as 71,414 (8%)
wounds in patients, who were only seen in the clinic for initial consultation. We removed
20,266 (2%) wounds with implausible dimensions, such as surface areas greater than
100 cm2 for arterial ulcer and 150 cm2 for other wound types as well as wounds from
patients with missing age and sex.
The final dataset included 620,356 unique wounds from 261,398 patients from 532
wound care clinics. We only used information that was collected during the initial intake
for the prediction model and tracked wound outcomes regardless of therapeutic
interventions. We selected a training dataset containing a randomly drawn subset of 70%
(n=434,249) of the wounds to develop the model, leaving a validation set with the
remaining 30% (n=186,107) of the wounds for validation of model parameters.
We split the dataset into a training set containing randomly drawn 70% of the
wounds and a validation set containing the remaining 30% of the wounds. The study was
reviewed and approved under an expedited process by the Institutional Review Board of
the University of Southern California (UP-18-00477).
53

Definition of Outcome
Our outcome is the status of the wound within the first twelve weeks of treatment initiation,
coded dichotomously into healed versus not healed, as commonly used in clinical trials
of wound treatment. [22,23] While patients can remain in treatment for longer than twelve
weeks, wound healing in that period was no longer considered for the analysis. Following
Ennis et al., we applied a modified intent-to-treat approach in that wounds in patients lost
to follow up during the twelve weeks were considered not healed after excluding wounds
that were followed for less than 7 days. [21]
The determination of healing is made by the treating clinician during each visit
based on the following criteria: (1) wound has zero wound measurements, i.e., it is
completely covered with a full layer of epithelium and no longer has exudate, (2) wound
has received a flap procedure and presents post procedure with complete take, (3) wound
has received a graft procedure and presents post procedure with complete success, (4)
wound margins have been approximated and sutured to facilitate closure and wound has
zero measurements.

Selection of Predictor Variables
We combined a review of the published literature on prediction models for wound healing
and clinical input to select potential predictor variables. We grouped predictor variables
into three categories. The first category represented demographic characteristics, such
as age, sex and smoking status, the second patient level clinical characteristics, such as
54

comorbid conditions and the number of wounds, which were identified based on ICD
codes or clinical notes, and the third, wound level characteristics, such as area, location,
and etiology.
Following Shaw et al., wound area was calculated from measured width and length
assuming an elliptical shape. [24] Physical depth of a wound was estimated by the treating
clinicians, and the “anatomical” depth (i.e., degree of tissue penetration) was categorized
as partial thickness (including dermis and epidermis but not the fascia), full thickness
(extension through the fascia into subcutaneous structure), or unknown thickness based
on a classification presented in the supplemental appendix. We refer to this variable as
“wound classification”. Variables that were infrequently coded were not considered for the
model. For example, prior studies found ankle brachial index, nutritional status, HbA1c,
and antibiotic therapy to be predictors of wound healing, but they were infrequently
documented in our dataset. [25,26,27] While the wound care centers usually collect such
information during a patient’s intake assessment, it was documented as free text in the
electronic medical records and could not be leveraged for our analyses.

Model Construction
We used logistic regression models to predict the probability of a wound being healed by
the end of week twelve as a function of our three categories of variables: demographic
characteristics, patient level clinical characteristics and wound characteristics. Using the
training dataset, we entered variables individually and retained only those that increased
the predictive accuracy of the model, as the large sample size implied that many variables
55

were statistically significantly correlated with our outcome without contributing
meaningfully to the predictive accuracy.
We added variables in a stepwise fashion and examined their contribution to model
performance using area under the receiver operating characteristic curve (AUC) and the
Akaike information criterion (AIC). The AUC is a measure of how well the model predicts
the outcome.17 An AUC of 0.5 means the model performs no better than a random guess,
while an AUC of 1.0 means perfect prediction, and values above 0.7 are regarded as
acceptable model fit. [28] Adding variables will always increase predictive power, but also
model complexity and risk of “overfitting”, i.e., optimizing the model to reflect the
information in a specific dataset, while making it less generalizable to different datasets.
Hence, the AIC was used to assess the contribution of a variable to the informational
quality of the model as it increases, if a variable contributes limited explanatory power
relative to its contribution to model fit. [29]
Having selected variables based on examination of AUC and AIC, we constructed
three logistic models of ascending complexity: the first model contained only demographic
characteristics, the second model added patient-level clinical characteristics, and the final
model wound characteristics.
As the magnitude of predicted odds ratios in logistic regression models cannot tell
the extent to which each variable drives the prediction, we used a classification tree model
with all variables of the final model to further assess the contribution of each explanatory
variable. [30] The classification tree model is a machine learning approach that
recursively splits the data into increasingly homogeneous groups using combinations of
predictor variables. For example, the algorithm might determine that all wounds below a
56

certain area in non-diabetic females have healed or that pressure ulcers with full thickness
in elderly patients that lasted more than 30 days before admissions did not heal. In other
words, the final nodes in the classification tree contain mostly healed or not healed
wounds after the data have been split multiple times using combinations of predictor
variables. Classification trees can outperform logistic models if there are strong
interactions between predictor variables. [31]
The relative variable importance metric derived from such classification tree
models is a number between 0 and 1. It is calculated in two steps. First, importance of
each variable is measured based on the change in the sum of the squares of residuals
(i.e., difference in predicted and actual values) when a predictor variable is used to split
a node. Then, relative importance of each variable is calculated by dividing each
variable’s importance by the highest variable importance among all predictors. [32] Like
logistic models, our classification tree model was developed and validated using 70% and
30% random samples, respectively.
All statistical analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC).
Logistic models and classification tree were created using PROC LOGISTIC and PROC
HPSPLIT commands, respectively. [33] The study was reviewed and determined to
qualify for the Human Research Protection Program Flexibility Policy by the Institutional
Review Board of the University of Southern California (UP-18-00477).

57

Results
Table 1 compares characteristics between wounds that healed and wounds that failed to
heal by the end of week twelve. About 59% of wounds healed by the end of week twelve.
Because of the large sample size, all variables, except for patient’s history of diabetes
and chronic obstructive pulmonary disease, showed statistically significant differences,
even when the absolute magnitude of the differences was small. Healed wounds were
smaller in area and depth and associated with lower prevalence of comorbidities such as
Alzheimer’s disease or dementia, coronary artery disease, and peripheral vascular
diseases. Pressure ulcers and arterial ulcers were less likely to heal than wounds of other
etiology.
The comparison of AUC and AIC for the three logistic regression models is shown
in Table 2. The predictive accuracy of a model using only demographic variables was
limited with an AUC of 0.556 and adding indicators for comorbid conditions only improved
the AUC to 0.605. The addition of variables capturing wound characteristics resulted in
substantial improvement in predictive accuracy with an AUC of 0.712. The AIC decreased
with the addition of comorbidity and wound characteristic variables, indicating that the
additional sets of variables were informative in predicting wound healing. AUCs and AICs
of three models are presented in Figure 1.
Table 3 shows the estimated odds ratios for the final model that used all wound
and patient level characteristics. Pressure ulcers, arterial ulcers, full thickness wounds,
and wounds in patients with multiple concurrent wounds were less likely to heal by the
end of week twelve.
58

The analysis of the relative variable importance based on the classification tree
model shows a similar pattern. As shown in Table 4, wound location had the highest
variable importance in predicting wound healing, and the importance of the other
variables is presented relative to it. Seven of the ten most important predictors were
wound level characteristics and eight of the ten least important predictors were patient-
level characteristics. Patient level characteristics had a relative variable importance of 0.2
or less, whereas wound characteristics had a relative variable importance around or
above 0.5. The classification tree model achieved an AUC of 0.717.

Discussion
Our study used data from a large electronic medical record to predict the probability of
wound healing as a function of patient and wound characteristics with both clinically
informed and machine learning approaches. Both methods achieved acceptable
predictive accuracy with AUCs above 0.70. Our results are consistent with previous
findings that wound level characteristics are better predictors of wound healing than
patient level characteristics, such as demographic information and comorbidities. [12-
14,30]
Our model’s performance is comparable or higher than previously published
models. For example, using the U.S. Wound Registry that combines electronic medical
records from over 50 wound care clinics, Horn et al. reported AUCs between 0.594 and
0.708 and Fife et al. reported an AUC of 0.648. , Using electronic medical records from
a large health system, Margolis et al. achieved AUCs of 0.70 and 0.71. [34,35,36] An
59

important distinction, however, is that our model applies to a broad range of wound
etiologies, whereas other models were restricted to a single etiology such as diabetic
neuropathic foot ulcers or venous leg ulcers.
Another important distinction is that our model exclusively relies on data collected
at the patient’s initial intake exam, whereas other studies added information on wound
healing progress from subsequent visits to increase predictive accuracy. Jung et al, for
example, used a machine learning approach and constructed a predictive model with an
AUC of 0.842, but the change in wound dimensions between the first and second visit
contributed substantially to the model’s performance. [19] Similarly, Horn et al.
demonstrated that adding information from subsequent visits increased the AUCs to
between 0.615 and 0.726 compared to their above-mentioned models with AUCs
between 0.594 and 0.708 that only used data collected during patient intake. [34] It is to
be expected that the initial healing is a strong predictor of future healing and may help to
identify patients, who show insufficient progress in wound healing and may require
additional interventions. Thus, models including information on healing progress have
utility for clinical decision making. However, the objective of our work was also to use
healing rates to compare quality of care between centers. As healing progress will be
influenced by initial treatment, we would argue that including it in a model might result in
confounding. [37]
Although an AUC greater than 0.70 represents acceptable predictive accuracy of
our model, the results point to room for improvement. A logical focus would be information
on wound characteristics, as our results show that they are better predictors of healing
than patient characteristics. One could explore additional wound level characteristics,
60

such as presence of biofilm, bioburden or activity patterns of metalloproteases, as those
have been shown to impact wound healing. [38,39]
In addition, as our results indicate that wound dimensions are an influential
predictor of healing, increasing the precision with which they are measured and thus
reducing random measurement error might improve predictive accuracy. At the moment,
the wound care centers contributing to our data typically estimate surface area based on
maximum width and length and approximate depth. Especially for larger and irregularly
shaped wounds, this approach can lead to error compared to methods like three-
dimensional digital imaging techniques. [40]
Furthermore, more complete documentation of patient-level characteristics could
improve predictive accuracy. BMI and smoking status, for example, are known predictors
of wound healing but were missing in 54% and 31% in our data, respectively. [41] Similarly,
measures of metabolic control, like HbA1c, ankle brachial index for perfusion or skin
temperature for inflammation are correlated with wound healing but not yet documented
routinely in our data. [25,36,42]
Conversely, we have no evidence that increasing the complexity of the model
structure will raise predictive accuracy. We tested several interaction effects with only
marginal improvement in the AUC. For example, adding interaction terms for wound
area*depth and wound location*classification to the final model improved the AUC by only
0.001. Thus, we have no evidence that the interrelationships between predictive variables
improve model performance. This result is consistent with the observation that the
classification tree model, whose predictive power relies heavily on interactions between
predictor variables, had a similar AUC as the final logistic model. Similarly, the use of
61

multi-level models that accounted for clustering of wounds by patients and patients by
centers did not improve predictive power meaningfully. In addition, our results suggest
that a more parsimonious prediction model using solely wound level characteristics
performs comparably to our current model. We tested a logistic model with wound
characteristics only, and the model achieved an AUC of 0.689.
Our analysis has important limitations. First, while our dataset is large, it only
contains data from one wound care EMR system and would have to be validated, ideally
prospectively, with data from other settings. Second, our data have been collected under
real-world conditions without validation and/or formal assessment of coding quality and
may be subject to measurement error. However, we would argue that this error is likely
to be random and thus reduce rather than overstate our predictive accuracy. Lastly, some
of our design decisions, such as exclusion of wounds with implausible dimensions and
creation of a unified grading scheme, may have introduced error or even bias. As with all
predictive models, performance needs to be tracked over time to account for changes in
coding practice and clinical care.

Innovation
Our findings show that wound healing can be predicted based on real-world data from
electronic medical records and a model based on prior evidence and clinical reasoning.
Our model can inform clinical decision making and form the basis for applications like
quality measure development, future research into clinical practice to determine sources
of variability and performance-based payment.

62

TABLES AND FIGURES
Table 1. Comparison of healed versus not-healed wounds
Wound status at 12
th
week
Healed
(n = 365,659;
58.9%)
Not healed
(n = 254,697;
41.1%)
P-value
Demographic characteristics
Age category <.0001
Less than 55 84840 (23.2%) 61769 (24.3%)
55 to 64 82030 (22.4%) 57910 (22.7%)
65 to 74 89277 (24.4%) 59054 (23.2%)
Greater than 75 109512 (30.0%) 75964 (29.8%)
BMI category <.0001
Less than 18.5 4180 (1.1%) 4717 (1.9%)
18.5 to 24 30725 (8.4%) 26925 (10.6%)
25 to 29 38736 (10.6%) 28764 (11.3%)
Greater than 30 99017 (27.1%) 54349 (21.3%)
Missing 193001 (52.8%) 139942 (54.9%)
Sex (Female) 153052 (41.9%) 108145 (42.5%) <.0001
Palliative 5176 (1.4%) 10754 (4.2%) <.0001
Smoking status <.0001
Never Smoker 118640 (32.5%) 77640 (30.5%)
Former Smoker 93523 (25.6%) 63343 (24.9%)
Current Smoker 41226 (11.3%) 33527 (13.2%)
Unknown 112270 (30.7%) 80187 (31.5%)
Clinical characteristics
Alzheimer’s disease/ dementia 13656 (3.7%) 13920 (5.5%) <.0001
Coronary artery disease 61053 (16.7%) 46068 (18.1%) <.0001
Congestive heart failure 55559 (15.2%) 40749 (16.0%) <.0001
Chronic obstructive pulmonary disease 43703 (12.0%) 30403 (11.9%) 0.8586
Diabetes 220893 (60.4%) 154372 (60.6%) 0.1120
Peripheral vascular diseases 77959 (21.3%) 67471 (26.5%) <.0001
Quadri/paraplegia 7681 (2.1%) 11872 (4.7%) <.0001
Hypertension 189764 (51.9%) 128880 (50.6%) <.0001
Number of concurrent wounds <.0001
One 169589 (46.4%) 91799 (36.0%)
Two 70977 (19.4%) 52024 (20.4%)
Three or more 125093 (34.2%) 110874 (43.5%)
Wound characteristics
Wound area (cm
2
) (Mean/SD) 6.4 (16.5) 9.7 (19.0) <.0001
Wound depth (mm) (Mean/SD) 2.1 (5.0) 4.1 (8.1) <.0001
Wound location <.0001
Pelvic 31801 (8.7%) 32759 (12.9%)
Upper leg 13345 (3.7%) 6980 (2.7%)
63

Lower leg 153995 (42.1%) 70559 (27.7%)
Foot 81564 (22.3%) 87863 (34.5%)
Toe 53544 (14.6%) 39462 (15.5%)
Amputation site 3614 (1.0%) 4199 (1.7%)
Other 27799 (7.6%) 12875 (5.1%)
Wound classification <.0001
Full thickness 157903 (43.2%) 132153 (51.9%)
Partial thickness 121364 (33.2%) 47575 (18.7%)
Superficial 30290 (8.3%) 16710 (6.6%)
Unknown 56102 (15.3%) 58259 (22.9%)
Wound etiology <.0001
Arterial ulcer 8844 (2.4%) 14842 (5.8%)
Diabetic ulcer 127129 (34.8%) 98792 (38.8%)
Pressure ulcer 54500 (14.9%) 57514 (22.6%)
Venous ulcer 97047 (26.5%) 44035 (17.3%)
Other 78139 (21.4%) 39514 (15.5%)
Necrotic wound tissue 97392 (26.6%) 88797 (34.9%) <.0001
Infected wound 124198 (34.0%) 99026 (38.9%) <.0001
Heavily exuding wound 98921 (27.1%) 81616 (32.0%) <.0001
Eschar formation 24500 (6.7%) 28960 (11.4%) <.0001
SD: standard deviation

Table 2. Model comparison
Model type AUC AIC
Model 1: Demographics only 0.556 249547.7
Model 2: Demographics + Clinical Characteristics 0.605 245548.4
Model 3: Demographics + Clinical + Wound Characteristics 0.712 225519.3
*AUC: area under the curve; AIC: Akaike information criterion

64

Figure 1. Comparison of area under the receiver operator curves

Brown: Demographics + Clinical Characteristics + Wound Characteristics
Green: Demographics + Clinical Characteristics
Red: Demographics

65

Table 3. Estimated odds ratios from logistic model
(Demographics + Clinical + Wound Characteristics)

Odds Ratio
Variable (Comparator vs Reference) Point
Estimate
95% Confidence
Interval
P-value
Depth 0.943 0.942 0.945 <.0001
Wound surface area 0.987 0.986 0.987 <.0001
Age category - Less than 55 (Reference)
Age category - 55 to 64 0.999 0.980 1.019 0.0110
Age category - 65 to 74 1.028 1.008 1.049 0.0245
Age category - Greater than 75 1.031 1.011 1.052 0.0064
BMI category - 18.5 to 24 (Reference)
BMI category – Between 25 to 29 1.091 1.060 1.123 <.0001
BMI category - Greater than 30 1.277 1.244 1.311 <.0001
BMI category - Less than 18.5 0.829 0.782 0.878 <.0001
BMI category – Missing 0.968 0.945 0.992 <.0001
Palliative (Yes vs No) 0.414 0.397 0.433 <.0001
Sex (Male vs Female) 1.125 1.110 1.140 <.0001
Smoking Status – Never Smoker (Reference)
Smoking Status – Current Smoker 0.868 0.849 0.889 <.0001
Smoking Status – Former Smoker 0.987 0.969 1.005 <.0001
Smoking Status – Unknown 0.817 0.800 0.834 <.0001
Dementia / Alzheimer’s (Yes vs No) 0.820 0.794 0.848 <.0001
Coronary artery disease (Yes vs No) 0.981 0.962 1.000 0.0466
Congestive heart failure (Yes vs No) 0.903 0.885 0.921 <.0001
Chronic obstructive pulmonary disease
(Yes vs No)
1.029 1.006 1.051 0.0115
Diabetes (Yes vs No) 1.114 1.095 1.134 <.0001
Peripheral vascular diseases (Yes vs No) 0.774 0.761 0.787 <.0001
Quadri/paraplegia (Yes vs No) 0.661 0.634 0.688 <.0001
Hypertension (Yes vs No) 1.048 1.031 1.066 <.0001
Number of concurrent wounds – One (Reference)
Number of concurrent wounds – Two 0.728 0.715 0.741 <.0001
Number of concurrent wounds – More
than Two
0.567 0.559 0.576 <.0001
Wound Location – Foot (Reference)
Wound Location – Pelvic 1.388 1.349 1.429 <.0001
Wound Location – Upper Leg 2.063 1.982 2.147 <.0001
Wound Location – Lower Leg 1.966 1.928 2.005 <.0001
Wound Location – Toe 1.319 1.292 1.346 <.0001
Wound Location – Amputation Site 1.117 1.054 1.183 <.0001
Wound Location – Other 2.019 1.953 2.086 <.0001
Wound Classification – Superficial (Reference)
Wound Classification – Full thickness 0.501 0.485 0.517 <.0001
Wound Classification – Partial thickness 0.886 0.856 0.918 <.0001
Wound Classification – Unknown 0.388 0.376 0.402 <.0001
Wound Type – Diabetic ulcer (Reference)
66

Wound Type – Arterial Ulcer 0.669 0.644 0.694 <.0001
Wound Type – Other 1.609 1.569 1.650 <.0001
Wound Type – Pressure Ulcer 0.857 0.832 0.883 <.0001
Wound Type – Venous Ulcer 1.460 1.425 1.495 <.0001
Necrotic wound tissue (Yes vs No) 0.784 0.771 0.797 <.0001
Infected wound (Yes vs No) 0.887 0.874 0.901 <.0001
Heavily exuding wound (Yes vs No) 0.909 0.894 0.925 <.0001
Eschar formation (Yes vs No) 0.790 0.770 0.811 <.0001

Table 4. Relative importance of variables
Relative Importance
Would location 1.0000
Wound surface area 0.9260
Wound classification 0.7329
Wound etiology 0.5550
Number of concurrent wounds 0.4919
Depth 0.3427
Palliative 0.2441
Necrotic wound tissue 0.2030
Peripheral vascular disease 0.2009
BMI 0.1846
Age 0.1198
Eschar formation 0.1145
Sex 0.1128
Smoking status 0.0943
Infected wound 0.0799
Diabetes 0.0585
Heavily exuding wound 0.0424
Quadri/paraplegia 0.0401
Alzheimer’s disease / dementia 0.0343
Congestive heart failure 0.0342
Coronary artery disease 0.0311
Chronic Obstructive Pulmonary Disease 0.0296
Hypertension 0.0262
*Area under the curve: 0.717

67

References
1. Lazarus GS, Cooper DM, Knighton DR, et al. Definitions and guidelines for
assessment of wounds and evaluation of healing. Wound Repair Regen
1994;2:165–170.
2. Coleman S, Nelson EA, Vowden P, et al. Development of a generic wound care
assessment minimum data set. J Tissue Viability 2017;26:226–240.
3. Fife CE, Carter MJ, Walker D, Thomson B. Wound care outcomes and
associated cost among patients treated in US out-patient wound centers: data
from the US wound registry. Wounds. 2012;24:10–7.
4. Nussbaum SR, Carter MJ, Fife CE, DaVanzo J, Haught R, Nusgart M, Cartwright
D. An Economic Evaluation of the Impact, Cost, and Medicare Policy Implications
of Chronic Nonhealing Wounds. Value Health. 2018 Jan;21(1):27-32.
5. Snyder RJ, Fife C, Moore Z. Components and Quality Measures of DIME
(Devitalized Tissue, Infection/Inflammation, Moisture Balance, and Edge
Preparation) in Wound Care. Send to Adv Skin Wound Care. 2016
May;29(5):205-15.
6. Fortington LV, Geertzen JH, van Netten JJ, Postema K, Rommers GM, Dijkstra
PU. Short and long term mortality rates after a lower limb amputation. Eur J Vasc
Endovasc Surg. 2013;46:124-131.
7. Murphy PS, Evans GR. Advances in wound healing: a review of current wound
healing products. Plast Surg Int. 2012;2012:190436.
68

8. Gainza G, Villullas S, Pedraz JL, Hernandez RM, Igartua M. Advances in drug
delivery systems (DDSs) to release growth factors for wound healing and skin
regeneration. Nanomedicine. 2015 Aug;11(6):1551-73.
9. Greer N, Foman NA, MacDonald R, Dorrian J, Fitzgerald P, Rutks I, Wilt TJ.
Advanced wound care therapies for nonhealing diabetic, venous, and arterial
ulcers: a systematic review. Ann Intern Med. 2013 Oct 15;159(8):532-42.
10. Warriner RA, 3rd, Carter MJ. The current state of evidence-based protocols in
wound care. Plast Reconstr Surg 2011;127(Suppl. 1):144S–153S.
11. Hospital Readmissions Reduction Program (HRRP). Centers for Medicare and
Medicaid Services. Available on https://www.cms.gov/medicare/medicare-fee-
for-service-payment/acuteinpatientpps/readmissions-reduction-program.html
[Accessed March 2019]
12. Innovation Models. Centers for Medicare and Medicaid Services.
https://innovation.cms.gov/initiatives/#views=models [Accessed March 2019]
13. Ince P, Game FL, Jeffcoate WJ. Rate of healing of neuropathic ulcers of the foot
in diabetes and its relationship to ulcer duration and ulcer area. Diabetes Care.
2007 Mar;30(3):660-3.
14. Margolis DJ, Allen-Taylor L, Hoffstad O, Berlin JA. Diabetic neuropathic foot
ulcers and amputation. Wound Repair Regen. 2005 May-Jun;13(3):230-6.
15. Margolis DJ, Allen-Taylor L, Hoffstad O, Berlin JA. Diabetic neuropathic foot
ulcers: the association of wound size, wound duration, and wound grade on
healing. Diabetes Care. 2002 Oct;25(10):1835-9.
69

16. Margolis DJ, Allen-Taylor L, Hoffstad O, Berlin JA. Diabetic neuropathic foot
ulcers: predicting which ones will not heal. Am J Med. 2003 Dec 1;115(8):627-31.
17. Kurd SK, Hoffstad OJ, Bilker WB, Margolis DJ. Evaluation of the use of
prognostic information for the care of individuals with venous leg ulcers or
diabetic neuropathic foot ulcers. Wound Repair Regen. 2009 May-Jun;17(3):318-
25.
18. Guihan M, Sohn MW, Bauman WA, Spungen AM, Powell-Cope GM, Thomason
SS, Collins JF, Bates-Jensen BM. Difficulty in Identifying Factors Responsible for
Pressure Ulcer Healing in Veterans With Spinal Cord Injury. Arch Phys Med
Rehabil. 2016 Dec;97(12):2085-2094.e1.
19. Jung K, Covington S, Sen CK, Januszyk M, Kirsner RS, Gurtner GC et al. Rapid
identification of slow healing wounds. Wound Repair Regen. 2016 Jan; 24(1):
181–188.
20. Jung K, Shah NH. Implications of non-stationarity on predictive modeling using
EHRs. J Biomed Inform. 2015 Dec; 58: 168–174.
21. Ennis WJ, Hoffman RA, Gurtner GC, Kirsner RS, Gordon HM. Wound healing
outcomes: Using big data and a modified intent-to-treat method as a metric for
reporting healing rates. Wound Repair Regen. 2017 Aug;25(4):665-672.
22. Arundel C, Buckley H, Clarke E, et al. Negative pressure wound therapy versus
usual care for Surgical Wounds Healing by Secondary Intention (SWHSI trial):
study protocol for a randomized controlled pilot trial. Trials 2016; 17:535.
70

23. Marston WA, Hanft J, Norwood P, Pollak R. The efficacy and safety of
dermagraft in improving the healing of chronic diabetic foot ulcers: results of a
prospective randomized trial. Diabetes Care 2003; 26:1701–1705
24. Shaw J, Hughes CM, Lagan KM, Bell PM, Stevenson MR. An Evaluation of
Three Wound Measurement Techniques in Diabetic Foot Wounds. Diabetes
Care Oct 2007, 30 (10) 2641-2642.
25. Brownrigg JR, Hinchliffe RJ, Apelqvist J, Boyko EJ, Fitridge R, Mills JL et al.
Performance of prognostic markers in the prediction of wound healing or
amputation among patients with foot ulcers in diabetes: a systematic review.
Diabetes Metab Res Rev. 2016 Jan;32 Suppl 1:128-35.
26. Zimny S, Pfohl M. Healing times and prediction of wound healing in neuropathic
diabetic foot ulcers: a prospective study. Exp Clin Endocrinol Diabetes. 2005
Feb;113(2):90-3.
27. Jones KR, Fennie K, Lenihan A. Chronic Wounds: Factors Influencing Healing
Within 3 Months and Nonhealing After 5-6 Months of Care. Wounds. 2007
Mar;19(3):51-63.
28. Zou KH, O'Malley AJ, Mauri L. Receiver-operating characteristic analysis for
evaluating diagnostic tests and predictive models. Circulation. 2007 Feb
6;115(5):654-7.
29. Arnold TW. Uninformative Parameters and Model Selection Using Akaike’s
Information Criterion. Journal of Wildlife Management 74(6):1175–1178; 2010.
71

30. Tonidandel S, LeBreton JM. Relative Importance Analysis: A Useful Supplement
to Regression Analysis. Journal of Business and Psychology Vol. 26, No. 1
(March 2011), pp. 1-9.
31. Culter DR. What Statisticians Should Know About Machine Learning. Paper 883-
2017. SAS Global Forum 2017. Available on
https://support.sas.com/resources/papers/proceedings17/0883-2017.pdf
[Accessed March 2019]
32. SAS/STAT® 14.1 User’s Guide The HPSPLIT Procedure. SAS Institute.
Available from
https://support.sas.com/documentation/onlinedoc/stat/141/hpsplit.pdf [Accessed
March 2019]
33. Rothman L. Predictive Modeling with SAS (for Health). SAS Institute. Available
from
https://www.sas.com/content/dam/SAS/en_ca/User%20Group%20Presentations/
Victoria-User-Group/Rothman-PredictiveModelingforHealth.pdf [Accessed March
2019]
34. Horn SD, Fife CE, Smout RJ, Barrett RS, Thomson B. Development of a wound
healing index for patients with chronic wounds. Wound Repair Regen. 2013 Nov-
Dec;21(6):823-32.
35. Fife CE, Horn SD, Smout RJ, Barrett RS, Thomson B. A Predictive Model for
Diabetic Foot Ulcer Outcome: The Wound Healing Index. Adv Wound Care (New
Rochelle). 2016 Jul 1;5(7):279-287.
72

36. Margolis DJ, Allen-Taylor L, Hoffstad O, Berlin JA. The accuracy of venous leg
ulcer prognostic models in a wound care system. Wound Repair Regen. 2004
Mar-Apr;12(2):163-8.
37. Sheehan P, Jones P, Caselli A, Giurini JM, Veves A. Percent change in wound
area of diabetic foot ulcers over a 4-week period is a robust predictor of complete
healing in a 12-week prospective trial. Diabetes care. 2003;26(6):1879-1882.
38. Tuttle MS. Association Between Microbial Bioburden and Healing Outcomes in
Venous Leg Ulcers: A Review of the Evidence. Adv Wound Care (New Rochelle).
2015 Jan 1;4(1):1-11.
39. McCarty SM, Percival SL. Proteases and Delayed Wound Healing. Adv Wound
Care (New Rochelle). 2013 Oct;2(8):438-447.
40. Bilgin M, Gunes UY. A comparison of 3 wound measurement techniques: effects
of pressure ulcer size and shape. J Wound Ostomy Continence Nurs.
2013;40(6):590-593.
41. Anderson K, Hamm RL. Factors That Impair Wound Healing. J Am Coll Clin
Wound Spec. 2012 Dec; 4(4): 84–91.
42. Brownrigg JR, Hinchliffe RJ, Apelqvist J, Boyko EJ, Fitridge R, Mills JL, et al.
Performance of prognostic markers in the prediction of wound healing or
amputation among patients with foot ulcers in diabetes: a systematic review.
Diabetes Metab Res Rev. 2016 Jan;32 Suppl 1:128-35.

73

Supplemental Appendices

Supplemental Appendix – Chapter 1

ICD-9 Codes
Groups Category ICD-9-CD codes
Disease Non-Hodgkin’s lymphoma 196.x, 200.xx, 202.xx
Transplant Type
Autologous 41.01 41.04 41.07, 41.09
Allogeneic 41.02 41.03, 41.05, 41.08
Stem Cell
Source
Peripheral blood 41.04 41.05, 41.07, 41.08
Bone marrow 41.01, 41.02, 41.03, 41.09
Conditioning Total body irradiation 92.24 92.26 92.29
Hematologic Febrile neutropenia
288.00 288.03 288.09 with 780.60 or
780.61
Infectious
Bacteremia 038.xx, 790.7
Clostrium Difficile 8.45
Pneumonia 480.xx – 486.xx
Fungal infection 112.0 112.5 112.89 112.9 117.3 117.9 118
Severe sepsis/shock 995.92 785.52
Immunologic Graft-versus-host disease 279.5x
Respiratory
Intubation/ventilation 96.01–96.05 93.90 96.70 96.71 96.72
Acute respiratory failure 518.8x 799.1
Renal Acute renal failure 584.x 586
Cardiac Acute heart failure 428.0 428.21 428.31 428.41
Gastrointestinal Mucositis 528.0x
Parenteral nutrition 99.15

74

Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Autologous HSCT
Mean hospital cost Mean length of stay
Variable Coefficient P-value Coefficient P-value
AGE
18-44 Reference Reference
45/54 -5583 0.152 -1.170 0.281
55/64 -3236 0.424 -1.490 0.216
>65 -7126 0.068 -0.666 0.565
Male vs Female 2462 0.249 0.255 0.566
RACE
White Reference Reference
Black -4443 0.135 -0.593 0.351
Hispanic 13397 0.002 0.625 0.380
Asian -2641 0.542 -0.083 0.928
Others/missing -3661 0.135 -3.612 <0.001
CCI 2 vs 3+ 3201 0.133 0.688 0.137
Hospital region
Northeast Reference Reference
Midwest -15614 <0.001 -1.527 0.049
South -13382 <0.001 -0.713 0.373
West 344 0.938 -0.956 0.381
Hospital bed size
Large Reference Reference
Medium 1662 0.570 0.344 0.669
Small 20477 <0.001 0.128 0.828
Hospital ownership
Private/non-profit Reference Reference
Private/for-profit -177 0.973 0.567 0.486
Government 14234 <0.001 1.843 0.024
PB vs BM -3779 0.309 -0.660 0.326
TBI -7752 0.052 -3.374 0.003
Febrile neutropenia 4322 0.020 0.935 0.010
Bacteremia 3589 0.171 2.943 <0.001
C. Difficile 5482 0.166 1.659 0.018
Pneumonia 13429 0.022 4.576 0.002
Fungal infection 8166 0.056 1.480 0.095
Severe sepsis/shock 19233 0.281 6.393 0.232
Intubation/mech.
ventilation
46897 0.002 6.488 0.003
Acute respiratory failure 15384 0.306 2.346 0.419
Acute renal failure 15750 0.004 3.355 0.009
Acute heart failure -5103 0.290 0.642 0.543
Mucositis 6617 0.001 1.416 0.003
TPN 29462 <0.001 6.196 0.002
Death -25611 0.234 -9.748 0.025
_constant 61154 <0.001 19.025 <0.001
*This table shows an incremental increase in hospital cost and length of stay for each independent
variable after controlling for other variables. Adjusted R-squared are 0.2946 and 0.2545 for cost and
length of stay models, respectively.
75

Multivariate Models for the Incremental Effect of Complications on
Hospital Cost and Length of Stay, Allogeneic SCT
Mean hospital cost Mean length of stay
Variable Coefficient P-value Coefficient P-value
AGE
18-44 Reference Reference
45/54 7400 0.423 0.006 0.996
55/64 -13576 0.157 -1.678 0.270
>65 -2003 0.862 -0.828 0.675
Male vs Female -8219 0.216 0.057 0.959
RACE
White Reference Reference
Black 19118 0.212 4.862 0.087
Hispanic 53454 0.009 5.116 0.069
Asian -28174 0.024 -1.857 0.457
Others/missing 1223 0.897 -0.892 0.661
CCI 2 vs 3+ -8213 0.328 -0.841 0.539
Hospital region
Northeast Reference Reference
Midwest -14329 0.064 -0.375 0.805
South 8709 0.342 -0.849 0.626
West 14752 0.235 0.001 >0.999
Hospital bed size
Large Reference Reference
Medium 28219 0.134 3.740 0.156
Small 18018 0.062 0.459 0.756
Hospital ownership
Private/non-profit Reference Reference
Private/for-profit -19869 0.162 0.144 0.956
Government 35419 <0.001 5.287 0.001
PB vs BM 40687 0.016 7.087 0.003
TBI -7093 0.505 1.868 0.260
Febrile neutropenia 8676 0.190 3.185 0.004
Bacteremia -5402 0.580 1.398 0.348
C. Difficile -7040 0.503 -0.162 0.920
Pneumonia 8068 0.650 4.446 0.091
Fungal infection 15479 0.500 2.123 0.557
Severe sepsis/shock 39534 0.246 5.826 0.312
GVHD 42990 0.002 10.126 <0.001
Intubation/mech.
Ventilation
23129 0.345 0.914 0.843
Acute respiratory failure 33953 0.114 -0.460 0.911
Acute renal failure 34836 0.002 7.315 <0.001
Acute heart failure -3263 0.902 -1.017 0.733
Mucositis -4259 0.548 0.802 0.432
TPN 33372 0.016 4.193 0.017
Death 95835 0.004 12.342 0.038
_constant 74919 <0.001 18.279 <0.001
*This table shows an incremental increase in hospital cost and length of stay for each independent
variable after controlling for other variables. Adjusted R-squared are 0.5131 and 0.4531 for cost and
length of stay models, respectively.

76

Supplemental Appendix – Chapter 2
Total Cost of Hospitalization from AHRQ HCUPnet

Adverse Events Rates and Expected Costs by Treatment
Regorafenib
% Cost ($) Expected ($)
Fatigue 9% $ 10,603 $ 954
Hand-foot syndrome 17% $ 13,376 $ 2,274
Diarrhea 7% $ 10,301 $ 721
Anorexia 3% $ 9,711 $ 291
Hypertension 7% $ 9,035 $ 632
Oral mucositis 3% $ 14,803 $ 444
Rash/Desquamation 6% $ 13,376 $ 803
Fever 1% $ 6,785 $ 68
Vomiting 1% $ 10,301 $ 103
Thrombocytopenia 3% $ 17,821 $ 535
Proteinuria 1% $ 10,453 $ 105
Anemia 2% $ 11,633 $ 233

TAS-102
% Cost ($) Expected ($)
Nausea 2% $ 10,301 $ 206
Vomiting 2% $ 10,301 $ 206
Anorexia 4% $ 9,711 $ 388
Fatigue 4% $ 10,603 $ 424
Diarrhea 3% $ 10,301 $ 309
Abdominal pain 2% $ 10,301 $ 206
Fever 1% $ 6,785 $ 68
Asthenia 3% $ 10,603 $ 318
Febrile neutropenia 4% $ 23,435 $ 937
Neutropenia 38% $ 23,435 $ 8,905
Leukopenia 21% $ 23,435 $ 4,921
Anemia 18% $ 11,633 $ 2,094
Thrombocytopenia 5% $ 17,821 $ 891

Best Supportive Care
% Cost ($) Expected ($)
Nausea 1% $ 10,301 $ 103
Anorexia 5% $ 9,711 $ 486
Fatigue 6% $ 10,603 $ 636
Abdominal pain 4% $ 10,301 $ 412
Asthenia 3% $ 10,603 $ 318
Anemia 3% $ 11,633 $ 349
Episode DRG 2017$ SE (Cost) Distribution
Fatigue 947 $ 10,603 189 Normal
Hand-foot syndrome 606 $ 13,376 842 Normal
Diarrhea 391 $ 10,301 92 Normal
Anorexia 640 $ 9,711 136 Normal
Hypertension 304 $ 9,035 176 Normal
Oral Mucositis 157 $ 14,803 531 Normal
Rash/Desquamation 606 $ 13,376 842 Normal
Fever 864 $ 6,785 102 Normal
Asthenia 947 $ 10,603 189 Normal
Vomiting 391 $ 10,301 92 Normal
Thrombocytopenia 813 $ 17,821 659 Normal
Proteinuria 695 $ 10,453 372 Normal
Anemia 811 $ 11,633 138 Normal
Abdominal pain 391 $ 10,301 92 Normal
Febrile Neutropenia 808 $ 23,435 792 Normal
Leukopenia 808 $ 23,435 792 Normal
Nausea 391 $ 10,301 92 Normal
Neutropenia 808 $ 23,435 792 Normal
77

Parameters and Distribution in PSA
Description
Point
Estimates
Distribution SE SOURCE
Median Overall Survival – regorafenib
6.4 CORRECT
Median Progression Free Survival – regorafenib
1.9 CORRECT
Median Overall Survival – TAS-102
7.1 RECOURSE
Median Progression Free Survival – TAS-102
2.0 RECOURSE
Median Overall Survival – BSC
5.3 RECOURSE
Median Progression Free Survival – BSC
1.7 RECOURSE
Median Overall Survival – regorafenib (CONCUR)
8.8 CONCUR
Median Progression Free Survival – regorafenib (CONCUR)
3.2 CONCUR
Rate, Stable to Progressive – regorafenib
0.365 Normal 0.055 Estimated
Rate, Stable to Dead – regorafenib
0.108 Normal 0.016 Estimated
Rate, Progressive to Dead – regorafenib
0.154 Estimated
Rate, Stable to Progressive – TAS-102
0.347 Normal 0.052 Estimated
Rate, Stable to Dead – TAS-102
0.098 Normal 0.015 Estimated
Rate, Progressive to Dead – TAS-102
0.136 Estimated
Rate, Stable to Progressive – BSC
0.408 Normal 0.061 Estimated
Rate, Stable to Dead – BSC
0.120 Normal 0.018 Estimated
Rate, Progressive to Dead – BSC
0.169 Estimated
Rate, Stable to Progressive – regorafenib (CONCUR)
0.217 Normal 0.032 Estimated
Rate, Stable to Dead – regorafenib (CONCUR)
0.079 Normal 0.012 Estimated
Rate, Progressive to Dead – regorafenib (CONCUR)
0.124 Estimated
Drug - regorafenib 40 mg 84 tabs $ 7,219 VA FSS 11/2017
Drug - TAS-102 Monthly Cost $ 8,265
Normal Estimated*
Drug - TAS-102 15/6.14 mg 20 tabs $ 2,070
VA FSS 11/2017
Drug - TAS-102 20/8.19 mg 20 tabs $ 2,755
VA FSS 11/2017
Office visit, established patient, level 4
$ 79
CMS Facility Price 2017
Complete cbc w/auto diff wbc $ 14
CMS Mid-Point 2017
Comprehensive Metabolic Panel $ 20
CMS Mid-Point 2017
Monthly Number of MD Visits 2
Uniform [1,4] Assumption: Uniform
Fatigue (MS-DRG 947) $ 13,729
AHRQ HCUPnet*
Hand-foot syndrome (MS-DRG 606) $ 17,090
AHRQ HCUPnet*
Diarrhea (MS-DRG 391) $ 13,277
AHRQ HCUPnet*
Anorexia (MS-DRG 640) $ 12,562
AHRQ HCUPnet*
Hypertension (MS-DRG 304) $ 11,413
AHRQ HCUPnet*
Oral Mucositis (MS-DRG 157) $ 18,442
AHRQ HCUPnet*
Rash/desquamation (MS-DRG 606) $ 17,090
AHRQ HCUPnet*
Fever (MS-DRG 864) $ 8,620
AHRQ HCUPnet*
Asthenia (MS-DRG 947) $ 13,729
AHRQ HCUPnet*
Vomiting (MS-DRG 391) $ 13,277
AHRQ HCUPnet*
Thrombocytopenia (MS-DRG 813) $ 20,345
AHRQ HCUPnet*
Proteinuria (MS-DRG 695) $ 13,690
AHRQ HCUPnet*
Anemia (MS-DRG 811) $ 14,808
AHRQ HCUPnet*
Abdominal pain (MS-DRG 391) $ 13,277
AHRQ HCUPnet*
Febrile Neutropenia (MS-DRG 808) $ 28,571
AHRQ HCUPnet*
Leukopenia (MS-DRG 808) $ 28,571
AHRQ HCUPnet*
Nausea (MS-DRG 391) $ 13,277
AHRQ HCUPnet*
Neutropenia (MS-DRG 808) 28,571
AHRQ HCUPnet*
Monthly Patient's Time in Hours - Stable Disease
10 Lognormal 1.5 Assumption
Monthly Patient's Time in Hours - Progressive Disease
20 Lognormal 3 Assumption
Monthly Caregiver's Time in Hours- Stable Disease
80 Lognormal 12 Assumption
Monthly Caregiver's Time in Hours - Progressive Disease
160 Lognormal 24 Assumption
78

Compensation Wage for an Average Individual $ 35 BLS 2017
Compensation Wage for Caregiver (Medical home aides) $ 12
BLS 2017
Daily Compensate Wage for Patient during Hospitalization $ 560
Calculation
Health State Utility - Stable Disease, regorafenib 0.73 Beta 0.011 CORRECT
Health State Utility - Progressive Disease, regorafenib 0.59 Beta 0.014 CORRECT
Health State Utility - Stable Disease, TAS-102 0.73 Beta 0.011 CORRECT
Health State Utility - Progressive Disease, TAS-102 0.59 Beta 0.013 CORRECT
Health State Utility - Stable Disease, BSC 0.74 Beta 0.017 CORRECT
Health State Utility - Progressive Disease, BSC 0.59 Beta 0.021 CORRECT

*Transition rates were estimated from median overall and progression free survival using the
DEALE method.
*Drug cost for TAS-102 was estimated based on the body surface area distribution during the
clinical trail
*Costs of hospitalization was obtained from 2014 AHRQ HCUPnet and they were inflated to
2017 US dollars using the consumer price index in health care.
*During hospitalization, a patient’s indirect cost calculation was based on 16 hours a day for
each day in hospital.

Body Surface Area Distribution and Cost of Lonsurf (Minimum dose: 40 mg)
% Patients # of 20 tabs bottles
Mean RECOURSE Dose 15 mg 20 mg Total cost
<1.07 1.07 0 35 1 1 $ 4,825
1.07 - 1.22 1.15 0.13 40 0 2 $ 5,510
1.23 - 1.37 1.3 2.38 45 3 0 $ 6,210
1.38 - 1.52 1.45 9.25 50 2 1 $ 6,895
1.53 - 1.68 1.6 19.88 55 1 2 $ 7,580
1.69 - 1.83 1.75 27 60 0 3 $ 8,265
1.84 - 1.98 1.9 21.38 65 3 1 $ 8,965
1.99 - 2.14 2.05 12.63 70 2 2 $ 9,650
2.15 - 2.29 2.2 5.75 75 1 3 $ 10,335
>=2.30 2.3 1.63 80 0 4 $ 11,020

*BSA distribution was used to vary drug cost in probability sensitivity analysis for TAS-102

79

Supplemental Appendix – Chapter 3

AUC and AIC in training and validation sets
Training set Validation set
Model type AUC
(95% Confidence
interval)
AIC AUC AIC
Model 1:
Demographics
only
0.557
(0.555 – 0.559)
581939.0 0.556 249508.4
Model 2:
Demographics +
Clinical
Characteristics
0.604
(0.602 – 0.605)
573031.9 0.604 245452.8
Model 3:
Demographics +
Clinical + Wound
Characteristics
0.725
(0.723 – 0.726)
518334.4 0.726 221776.9

ICD-9/10 codes and clinical codes used for identification of co-morbidities
Comorbidity Variable descriptors in Healogics’ EMR system and
ICD-9/10 codes
Dementia /
Alzheimer’s
 Dementia
 331, F03, G30
Coronary artery
disease
 Coronary Artery Disease, Myocardial Infarction
 410-414, I20-I25
Congestive heart
failure
 428, I50
 Congestive Heart Failure
Chronic obstructive
pulmonary disease
 COPD
 490-492, 496, J41-J44
Diabetes

 Is Diabetic Patient, Type I Diabetes, Type II
Diabetes
 250, E08-E13, Z79.4
Peripheral vascular
diseases
 Peripheral Arterial Disease, Peripheral Venous
Disease
 443, 447, 448, I70, I73, I79.9
Plegia  Paraplegia, Quadriplegia
Hypertension  Hypertension

80

Classification of Wound Stages
iHeal Classification Model Classification
NULL Unknown
Category/Stage I Superficial
Category/Stage II Superficial
Category/Stage III Partial thickness
Category/Stage IV Full thickness
Deep Tissue Pressure Injury Persistent non-blanchable
deep red, maroon or purple discoloration
Full thickness
Eschar covered Unknown
Full Thickness Full thickness
Full Thickness With Exposed Support Structures Full thickness
Full Thickness Without Exposed Support Structures Full thickness
Grade 0 Superficial
Grade 1 Partial thickness
Grade 2 Full thickness
Grade 3 Full thickness
Grade 4 Full thickness
Grade 5 Full thickness
Medical Device Related Pressure Injury Unknown
Mucosal Membrane Pressure Injury Unknown
NA Unknown
Partial Thickness Partial thickness
Stage 1 Pressure Injury Superficial
Stage 2 Pressure Injury Superficial
Stage 3 Pressure Injury Partial thickness
Stage 4 Pressure Injury Full thickness
Suspected Deep Tissue Injury Full thickness
Unable to visualize wound bed Unknown
Unclassifiable Unknown
Unstageable Pressure Injury Obscured full-thickness
skin and tissue loss
Unknown
Unstageable/Unclassified Unknown

Abstract (if available)

Abstract The field of health economics and outcomes research (HEOR) continues to evolve as patients, health care providers, and payers seek to make a judicious choice between cost, quality, and access. Taken together, these three papers illustrate how HEOR studies can be used to assist decision making of the stakeholders. First paper reports the cost and safety of hematopoietic stem cell transplantation, which serve as a standard of care treatment for relapsed and refractory non-Hodgkin’s lymphoma, and assist comprehensive evaluation of new CAR-T therapy in the absence of head-to-head clinical trials. Second paper used a three state Markov model to examine the cost-effectiveness of TAS-102 and regorafenib in metastatic, refractory colorectal cancer in the United States in light of recent clinical evidence and underscores the importance of managing treatment-related toxicities and maintaining health-related quality of life. Third paper develops and validates severity-adjusted prediction models for chronic wound healing that can be used to implement value-based reimbursement by estimating the differences between actual and predicted healing rates across wound care centers.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

New approaches using probabilistic graphical models in health economics and outcomes research

PDF

Advances and applications for economic evaluation methods in health technology assessment (HTA)

PDF

Economic, clinical, and behavioral outcomes from medical and pharmaceutical treatments

PDF

Three essays on comparative economic and clinical outcomes of antipsychotics using retrospective claims data

PDF

Outcomes of antibiotic use among children with acute respiratory tract infections

PDF

Pharmacotherapy among patients with type 2 diabetes

PDF

Evaluating cancer treatments with health economics methods

PDF

Controlling for biases from measurement errors in health outcomes research: A structural equation modeling approach

PDF

Impact of pharmacy-based transitional care on healthcare utilization and costs

PDF

Discriminating changes in health using patient-reported outcomes

PDF

Essays in health economics

PDF

Treatment of hepatitis C in era of direct acting antivirals (DAAs): selected essays in health economics

PDF

The price and clinical impact of specialty drug management in the US healthcare system

PDF

Effects of a formulary expansion on the use of atypical antipsychotics and health care services by patients with schizophrenia in the California Medicaid Program

PDF

Hepatitis C in the post-interferon era: selected essays in health economics

PDF

Essays in health economics and provider behavior

PDF

Value in oncology care and opportunities for improvement

PDF

Value in health in the era of vertical integration

PDF

Documenting the impact of under-treatment for chronic disorders

PDF

Essays on health insurance programs and policies

Asset Metadata

Creator Cho, Sang Kyu (author)

Core Title Health economics and outcomes research for informed decision making in rapidly evolving therapeutic areas

School School of Pharmacy

Degree Doctor of Philosophy

Degree Program Health Economics

Publication Date 05/07/2020

Defense Date 03/18/2020

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

Tag Health Economics,OAI-PMH Harvest,pharmacoeconomics

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor McCombs, Jeffrey (committee chair), Hay, Joel (committee member), Mattke, Soeren (committee member)

Creator Email sangkcho@usc.edu,skc2152@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-301749

Unique identifier UC11664318

Identifier etd-ChoSangKyu-8450.pdf (filename),usctheses-c89-301749 (legacy record id)

Legacy Identifier etd-ChoSangKyu-8450.pdf

Dmrecord 301749

Document Type Dissertation

Rights Cho, Sang Kyu

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