University of Southern California Dissertations and Theses

On the electrophysiology of multielectrode recordings of the basal ganglia and thalamus to improve DBS therapy for children with secondary dystonia

(USC Thesis Other)

On the electrophysiology of multielectrode recordings of the basal ganglia and thalamus to improve DBS therapy for children with secondary dystonia

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content On the Electrophysiology of Multielectrode Recordings of
The Basal Ganglia and Thalamus to Improve DBS
Therapy for Children with Secondary Dystonia

by

Juan Enrique Argüelles Morales

A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BIOMEDICAL ENGINEERING)

August 2020
ii

Acknowledgments
I am very thankful to all the people who have contributed to the completion of my
doctoral studies. First and foremost, I want to thank my loving parents for being the
unconditional platform in which I stand every step along my way, for trusting all my
decisions, and for being my best example of kindness, integrity, perseverance and hard
work. These values made me who I am today and keep guiding my everyday actions. I
also wish to thank my sisters and partners in crime, Sandra and Nayelli, for being always
there for me when I need their carrying advice and support.
I am very lucky and thankful to have my girlfriend and best friend Angie with me
throughout all these years. She is my constant source of encouragement and inspiration.
Without her, the difficult times would have been tougher and the good ones less
enjoyable. Thank you, Angie, for making my life better.
I would like to express my sincere gratitude to my advisor Terry Sanger for his
patient guidance and support throughout (and beyond) my PhD. His teachings, unique
mix of medical and engineering expertise and sharp perspective, have shaped my
research thinking and contributed to my academic training. To my lab partners and friends
Matteo Bertucco, Shinichi Amano, Maryam Abolfath-Beygi, Ruta Deshpande,Estefania
Hernandez and Diana Ferman for the enriching discussions and constant feedback.
Thanks to all the members of the Sanger lab for making it the best work environment that
one could ever ask.
Finally, I would like to express my deepest gratitude to Professor Francisco Valero-
Cuevas for being the spark that initiated this adventure. His USC-Mexico Summer
Internship program exposed me for the first time to research and sparked my interest in
iii

pursuing a PhD in the BME department at USC. For your advice and sincere friendship,
thank you, Francisco. I hope this great program continues bringing talented students
eager to succeed in their professional carriers.

iv

Contents
Acknowledgments......................................................................................................... ii
List of Tables .............................................................................................................. viii
List of Figures .............................................................................................................. ix
Abstract ....................................................................................................................... xiv
1 Introduction ............................................................................................................... 1
Opportunities and limitations in the study of deep brain chronic multielectrode
recordings for movement disorders ............................................................................. 2
2 Inpatient depth multielectrode recordings for DBS target selection in
subjects with secondary dystonia ........................................................................... 5
Clinical rationale .......................................................................................................... 5
Patient selection .......................................................................................................... 6
Electrode implantation ................................................................................................ 7
Recording electrodes .................................................................................................. 7
Recording system and recording configuration ........................................................... 9
3 Spike detection from low impedance microelectrode recordings ...................... 11
Introduction ............................................................................................................... 12
v

Signal preprocessing for multielectrode recordings .............................................. 12
The Teager-Kaiser nonlinear energy operator for spike detection ........................ 14
Feature extraction and spike clustering ................................................................ 16
Methods .................................................................................................................... 18
Multiple Sample Nonlinear Energy Operator (MSNEO) ........................................ 19
Spike detection using MSNEO ............................................................................. 21
Feature extraction and clustering ......................................................................... 22
Simulated spike trains and background noise construction .................................. 23
Common noise reduction ...................................................................................... 25
Algorithm performance ......................................................................................... 25
Results ...................................................................................................................... 26
Common noise reduction ...................................................................................... 26
MSNEO spectrogram ............................................................................................ 27
Spike detection performance ................................................................................ 28
Discussion................................................................................................................. 30
Implications for closed-loop deep brain stimulation .............................................. 31
Implications for online spike detection .................................................................. 32
How important is accurate spike isolation? ........................................................... 32
4 Patterns of neural activation of the GPi and thalamic nuclei in children with
secondary dystonia may provide evidence for a selective
inhibition/disinhibition model of the basal ganglia .............................................. 34
vi

Introduction ............................................................................................................... 35
Methods .................................................................................................................... 37
Subjects ................................................................................................................ 37
Data collection ...................................................................................................... 38
Data analysis ........................................................................................................ 38
Results ...................................................................................................................... 41
Spike waveforms .................................................................................................. 41
Firing rate statistics ............................................................................................... 44
Firing rates in GPi showed positive correlation with thalamic firing rates and
muscle activation .................................................................................................. 47
Discussion................................................................................................................. 50
Unbiased sample of basal ganglia and thalamus ................................................. 51
Selective inhibition/disinhibition model for the function of basal ganglia ............... 52
5 Thalamic evoked potentials in response to GPi-DBS show frequency
dependent modulation............................................................................................ 56
Introduction ............................................................................................................... 57
Methods .................................................................................................................... 59
Subjects: inpatient intracranial recordings ............................................................ 59
Subjects: outpatient EEG recordings .................................................................... 60
Data collection ...................................................................................................... 61
vii

Data analysis ........................................................................................................ 62
Results ...................................................................................................................... 65
Artifact reduction by reversed polarity stimulation ................................................ 65
Low frequency GPi-DBS elicited Vo EPs .............................................................. 66
Low frequency GPi-DBS does not elicit a cortical response ................................. 69
Vo response to GPi-DBS was frequency dependent ............................................ 72
Discussion................................................................................................................. 72
Attenuated propagation of DBS depolarization may indicate similarities with
ablation ................................................................................................................. 73
6 Overall Conclusions ............................................................................................... 75
Bibliography ................................................................................................................ 78

viii

List of Tables
Table 4.1: Demographics, temporary recording lead locations, and final selected
targets. ............................................................................................................... 39
Table 4.2: Median firing rates of the aggregate data ..................................................... 44
Table 4.3: Summary of the firing rate statistics. Values indicate the group median.
KS = Kolmogorov-Smirnov. KW = Kruskal-Wallis. .............................................. 45
Table 5.1: Demographics, temporary recording lead locations, and final selected
targets. ............................................................................................................... 60
Table 5.2: Demographics and implanted DBS targets................................................... 63

ix

List of Figures
Figure 2.1: Schematic of the Ad-Tech MM16C lead. ....................................................... 9
Figure 2.2: Schematic of the unity gain amplifier (buffer). ............................................. 10
Figure 3.1: Example of the resulting frequency decomposition and detection signal
using MSNEO. Top: sample spike waveform extracted from sharp
microelectrode recordings. Middle: MSNEO spectrogram. Bottom: gray
traces indicate the energy content at each frequency band. Red trace is
the detection signal computed as the product of all frequency bands. ............... 22
Figure 3.2: Proposed spike sorting algorithm. The discrete multichannel signal
X(n) is high-pass filtered (HPF) to remove low frequency field potentials,
then, the common median reference is computed from all channels and
subtracted from each HPF signal, the resulting signal is input to the operator
Ψ(xn, k), as well as band-pass filtered for spike waveform extraction (BPFs).
The total energy per frequency band around the spike index Σωkcan be
computed and input to the superparamagnetic clustering (SPC) algorithm.
The spike indices are found based on the detection signal ∏Ψn,k and
assigned with a probability P(s) which is an input to the SPC. Spike
waveforms are extracted from the BPFs and three first two principal
components are computed (PCA) and input to the SOC. ................................... 24
Figure 3.3: Percentage of error introduced by each common subtraction method.
Left: average error for 1 minute of simulated recording and firing rate of 10
spikes/second. Right: heatmaps of the recording length and spike frequency. .. 27
Figure 3.4: Sample detection signals for several SNR levels (columns). Top: noisy
spike waveforms. Middle: estimated MSNEO spectrogram. Bottom:
frequency bands (gray) and detection signal (red). Note that the detection
signal is robust to increased noise levels. .......................................................... 28
x

Figure 3.5: Receiver Operator Characteristics (RO) curve comparing the
performance of the three spike detection methods for several SNR
conditions. Note that MSNEO outperforms sNEO and AT in all cases
and for SNR>1 the algorithm shows close to perfect discrimination. .................. 29
Figure 3.6: ROC curves comparing the performance of the spike detectors at the
output of the spike sorter for several SNR conditions. ........................................ 30
Figure 3.7: Contribution of the normalized peak value P(s) of the detection signal
to the improvement of spike identification for a SNR = 0.75. Left: SPC
output using 4 principal components. Right: SPC output using 3 principal
components and P(s). Note that the elongated P(s) feature cluster of the
spike (green) provides better differentiation from the noise cluster (red) than
PC4. ................................................................................................................... 31
Figure 3.8: Left: spike waveform and the MSNEO output for several values for k.
Right: Area under the ROC curve for spike detection with MSNEO for
several sample shifts k and SNR = 1. Note that the performance of the
detector significantly improves as k increases, with maximum peak at k = 3,
which corresponds to the peak frequency of the spike waveform (1 kHz). ......... 33
Figure 4.1: Example of isolated spike waveforms with SNR > 1. .................................. 42
Figure 4.2: Example of a multi-unit cluster (top row) and a well isolated single unit
(bottom row) and their respective firing rate distributions (right column).
Red lines indicate the interquartile range of distribution. .................................... 43
Figure 4.3: Spike (left) and spike-like artifact (right). Note that the artifact has a high
SNR and firing rates (FR) within the physiological range which makes it
difficult to exclude using any automatic criteria. .................................................. 43
Figure 4.4: GPi firing rate distributions per condition. Two medians are significantly
different at the 5% significance level if their intervals (red triangles). * p<0.05,
** p<0.01. ............................................................................................................ 46
xi

Figure 4.5: Vim firing rate distribution per condition (** p<0.01) .................................... 46
Figure 4.6: Vo/VA firing rate distribution per condition (** p<0.01) ................................ 47
Figure 4.7: Spike raster (top) and EMG envelope (bottom) of about 1 hour and 9
minutes of recordings from subject 1. Note the high spike activity during the
periods of active EMG, as well as the widespread positive correlation
between GPi and thalamic nuclei. ...................................................................... 48
Figure 4.8: Comparison between median firing rates at rest and during voluntary
movement. Red areas indicate that the median firing rate increased with
movement, and blue areas indicate that the median firing rates decreased
with movement. .................................................................................................. 49
Figure 4.9: Correlation coefficient matrix between firing rates across brain areas for
subject 2. GPi1 and GPi2 correspond to two different electrodes implanted in
the anterior and posterior portions the GPi. ........................................................ 50
Figure 4.10: Power spectral density of EMG (black) and x (blue) and y (red)
components of the task. The EMG PSD was computed from the envelopes of
the anterior deltoid (AntDelt), biceps (Bi), lateral deltoid (LatDelt), posterior
deltoid (PostDelt), supraspinatus, triceps (Tri), and the wrist extensor (WE)
and flexor (WE) groups. ...................................................................................... 53
Figure 4.11: Comparison of the power spectral density of the LFP recording in GPi
(left column) and Vim (right column) during the performance of the rhythmic
task at 0.5 Hz (black trace), and during resting state (red trace 土 standard
deviation). Note the two peaks in GPi above and below the task-relevant
frequency, and the peak in Vim at the task-relevant frequency. ......................... 54
Figure 5.1: Schematic of the depth electrode and stimulation polarities. Artifact
reduction was achieved by reversing the stimulation polarity between
stimulation blocks of ~1200 interpulse epochs. Big black squares represent
xii

the macroelectrodes from 1 to 6, and small circles in between represent
the microelectrodes. ........................................................................................... 62
Figure 5.2: Example of the artifact reduction method. Left: Upsampled cathodic,
anodic, and average EP signals. The blue trace represents the average
of ~1200 stimulus applied to macroelectrode 1 negative (-, cathode) and
macroelectrode 2 positive (+, anode). The red trace shows the response
to the same number of stimuli but with inverted stimulation polarity, i.e. 1+2-.
The black trace is the average response of both. Right: Vo EP (black trace)
which occurs within 2 ms of GPi-DBS. ............................................................... 66
Figure 5.3: Vo EP in response to 9 Hz GPi-DBS. Red and blue traces represent
the two stimulation polarities. The black traces are the average response.
Each plot corresponds to a different depth along the recording lead, being
the top left plot the most distal microelectrode pair and the bottom right plot
the most proximal relative to the tip of the lead. ................................................. 67
Figure 5.4: Vim EP (left) and VPLa (right) responses to 9 Hz GPi-DBS. ....................... 68
Figure 5.5: Propagation of the Vo EP along the recording lead. The heatmap shows
the average Vo EP signal per microelectrode group indicated in the left
schematic of the recording electrode. Note the time delay of the EP response
as it distances from the stimulation target. ......................................................... 69
Figure 5.6: Aggregate cortical EP in responses to GPi-DBS. Gray areas indicate
the standard error of the mean, black traces are the average EP signal, and
red dots indicate when the signal was significantly different from zero (above
and below). ......................................................................................................... 70
Figure 5.7: Aggregate cortical EP in responses to Vo-DBS. ......................................... 70
Figure 5.8: Topographical heatmap of the cortical EP in response to (from top to
bottom) Vo-DBS, VPL-DBS, and GPi-DBS. Each column represents the
response to different stimulating electrode pair. ................................................. 71
xiii

Figure 5.9: Vo EP in response to GPi-DBS frequency swab from distal (top row) to
proximal (bottom row) microelectrodes in reference to the tip of the lead.
Left: average Vo EP traces. Red dots indicate the peak of the EP. Right: EP
amplitudes corresponding to each GPi stimulation frequency. Note that the
EP amplitude decreases as the stimulation frequency increases. ...................... 72
Figure 5.10: EP amplitude is attenuated as the stimulation period is closer to the
repolarization period of the downstream nuclei. ................................................. 73

xiv

Abstract
For patients suffering from severe dystonia, deep brain stimulation (DBS) may be
an effective treatment to alleviate motor symptoms. Although effective, the level of
efficacy varies from patient to patient, likely due to differences in the etiology and level of
injury. The challenges to improve DBS therapy rely on two fundamental questions: what
are the neural patterns of abnormality leading to the development of dystonic symptoms?
and, how does electrical stimulation can effectively normalize these patterns? We recently
developed an inpatient multielectrode recording and stimulation protocol for the selection
of DBS target in pediatric patients with movement disorders. While this protocol was
designed to solve a clinical problem, it provides an unprecedented opportunity for
studying the electrophysiology of brain areas involved in the control of movement and
how neural patterns of activation correlate with motor behavior, moreover, it offers the
possibility to study the response of the neural tissue to electrical stimulation and how DBS
may modify abnormal neural patterns causing movement disorders. In order to explore
these questions, it is essential to first understand the electrophysics of these specific
recordings. Depth multielectrodes approved for human use consist of arrays of blunt
microelectodes which physical properties differ from standard sharp microelectrodes
commonly used for spike recordings. Large microelectrodes can record average activity
from a relatively large volume of neurons, while small microelectrodes can detect more
localized neural single unit activity. The microelectrodes used in this recording setup are
larger than the standard, but small enough to register spiking activity. This unique
characteristic results in low signal-to-noise ratio (SNR) recordings in which the spike
waveforms are buried in the background neural noise.
xv

First, we developed a new methodology to detect single unit activity from low SNR
recordings. The spike detection rate and spike classification were improved by introducing
a generalization of the Teager-Kaiser nonlinear energy operator which accounts not only
for the spike amplitude but also for the instantaneous energy content of the spike
waveform. The proposed algorithm was efficient and therefore suitable for processing the
terabytes of data commonly collected during the inpatient multielectrode recording
protocol. It was also robust to high noise content and to the selection of the detection
threshold.
Once the necessary spike detection methodology was established, firing rates
were analyzed for multiple basal ganglia and thalamic nuclei in pediatric subjects with
secondary dystonia. The correlation between the neural activation of GPi (internal globus
pallidus) and thalamic neurons was explored during the execution of voluntary movement.
Two relevant observations arose from this analysis. First, the vast majority of neurons in
GPi and thalamic nuclei fired with a median frequency below 5 Hz. This contrasts with
typical reports of intraoperative GPi discharge rates between 10 and 70 Hz, likely biased
towards fast firing neurons given the short recording times in the intraoperative setting.
Second, increased widespread activation of the GPi positively correlated with thalamic
activation during voluntary movement. This observation contrasts with the traditional rate
model of the basal ganglia which suggests that activation of the GPi results in widespread
thalamic inhibition, which leads to suppression of cortical motor commands to finally
inhibit movement. Instead, we propose a new model in which the basal ganglia can
selectively inhibit different dynamics by control of the inhibitory output from GPi and
xvi

therefore regulating ongoing thalamocortical activity for the selection of the appropriate
motor command.
Finally, we analyzed the effects of neural depolarization due to DBS in motor
areas. We present the application of a stimulation polarity reversal technique to reduce
artifact contamination in short latency evoked potential (EP) recordings. This technique
was applied to intracranial microelectrode and electroencephalography (EEG) recordings
to study the propagation of the DBS pulse through the brain. Thalamic and cortical EPs
in response to low frequency stimulation of the GPi (GPi-DBS) were analyzed, then the
GPi-DBS frequency was varied, and the thalamic EPs were measured. We found that low
frequency GPi-DBS elicited an evoked response in its projecting thalamic nuclei,
however, no cortical response was observed. Thalamic EPs in response to GPi-DBS had
a peak amplitude within 2 ms after stimulation onset, which decreased as the stimulation
frequency increased, showing the frequency dependent modulatory effects of DBS on the
neural network. These findings have not been previously reported, probably because
most studies of intracranial EPs record from the implanted DBS macroelectrodes, which
large surface area average a large volume of neural activity, therefore losing spatial and
temporal resolution. The observations resulting from this study are consistent with a
model in which blocking downstream propagation of DBS may be one way in which DBS
can function as a lesion, substituting patterns of variation of neural activation with
meaningful information, for regular neural patterns with no variation and therefore
irrelevant information.
1

1 Introduction
Dystonia is a movement disorder characterized by involuntary muscle contractions
and abnormal postures. When the symptoms are severe, the quality of life of people
suffering from this condition can be significantly reduced. Most lesions responsible for
symptomatic dystonia involve the basal ganglia or thalamus. In a meta-analysis of 240
patients with lesions affecting the basal ganglia and causing movement disorders, 36%
exhibited dystonia (Bhatia and Marsden 1994). Furthermore, dystonia was also observed
in 30% of patients with movement disorders associated with lesions of the thalamus and
subthalamic region (Lee and Marsden 1994). The difficulty in treating dystonia comes
from the fact that no consistent pattern of anatomical pathology has been found,
moreover, there is emerging evidence that a cerebellar deficit may contribute to the
symptoms.
Dystonia can be classified according to its etiology. Primary dystonia occurs in
patients who have no signs of structural abnormality in the central nervous system.
Secondary dystonia is observed when there is a demonstrable exogenous, structural or
metabolic cause. Deep brain stimulation (DBS) of the internal segment of the globus
pallidus (GPi) has been established as an effective surgical procedure to treat primary
dystonia, however, it has been less effective to treat secondary dystonia, likely due to the
heterogeneity of causes leading to the development of the motor symptoms. Recently,
we have developed a new inpatient neuromodulation procedure for DBS target selection
for patients with secondary dystonia, where temporary depth electrodes are used to test
stimulation and recording at multiple possible targets. This new protocol has proven
2

effective at improving DBS target selection (Sanger et al. 2018), and it has open a
valuable opportunity to record in vivo from the suspected DBS targets in the human brain.
Depth electrodes approved for chronic use in humans differ significantly from those
typically used in traditional electrophysiology studies, therefore, the spatial and temporal
characteristics of the recording signals are different and unique to these specific depth
electrodes. In order to improve DBS therapy, it is necessary to understand the patterns
of abnormality leading to the development of the dystonic symptoms, therefore, it is
essential to understand the electrophysiology of the recording electrodes and establish
the appropriate methodology to analyze the neural signals.
Opportunities and limitations in the study of deep brain chronic
multielectrode recordings for movement disorders
The opportunity to record neural activity directly inside the human brain is limited
to few clinical situations. Methods of neuroimaging, such as functional magnetic
resonance imaging (fMRI), though immensely popular, are indirect ways of registering
neural activity based on measurements of regional hemodynamic changes in the brain.
The major limitations of fMRI are its spatial and temporal resolutions. For instance, it
cannot demonstrate how individual neurons and small neural assemblies contribute to
process and transmit sensorimotor information, and it lacks the millisecond resolution
which is often important in probing the timing of neuronal activity.

3

Information transmission in the nervous system can be inferred by the analysis of
the changes in the extracellular potentials recorded by electrodes inserted in the neural
tissue. Neural activity can be measured at different spatial and temporal resolutions
depending on the characteristics of the electrode probe. While small and sharp
microelectrodes can record action potentials (spikes) from a single neuron located near
the electrode’s tip, large macroelectrodes can record average spiking and subthreshold
activity from the volume of neurons surrounding the electrode’s surface.
The study of electrophysiological activity of deep brain structures in humans such
as basal ganglia, has been limited to single sharp microelectrode recordings, usually
during the implantation of deep brain stimulation (DBS) leads for the treatment of
movement disorders. Current advances on simultaneous multielectrode recordings offer
a valuable tool to study complex brain processes that involve large neural networks.
Depth multielectrode recordings in humans have been used to detect epileptic foci in
patients undergoing resection surgery for refractory epilepsy, however, most depth
multielectrodes available for use in humans involve bundles of microwires that differ from
typical microelectrodes in shape, size and impedance. Microwires have cylindrical
shapes, with a surface area usually often twice the size of a microelectrode, and with
impedances that are typically one order of magnitude lower. These characteristics can
affect the electrode’s ability to isolate single units (spike waveforms generated by a single
source neuron). Although these low impedance microelectrodes can detect single unit
activity, a significant percentage of the detected neurons may correspond to multi-unit
and local field potentials (LFP), which carrie information about the population of neurons
surrounding the electrode. The relevance of the distinction between single- and multi-unit
4

depends on the kind of questions to be investigated. While it may be less critical for
machine-interface applications, it is particularly relevant when studying pairwise
correlation of spike trains or rate coded behaviors.
Depth electrodes approved for chronic use in humans differ significantly from those
typically used in traditional electrophysiology studies, therefore, the spatial and temporal
characteristics of the recording signals are different and unique to these specific depth
electrodes. In order to improve DBS therapy, it is necessary to understand the patterns
of abnormality leading to the development of the dystonic symptoms, therefore, it is
essential to understand the electrophysiology of the recording electrodes and establish
the appropriate methodology to analyze the neural signals.

5

2 Inpatient depth multielectrode recordings for DBS target
selection in subjects with secondary dystonia
This chapter describes the procedure for in vivo recordings of deep brain nuclei
based on the use of stereo EEG depth multielectrodes commonly used for identification
of epileptic foci. In this novel protocol, temporary leads were implanted and used to test
stimulation and record from multiple possible DBS targets, including the internal globus
pallidus (GPi), subthalamic nucleus (STN), and several subnuclei of the thalamus. An
overview of the clinical rationale for the setup is provided. The principal motivation for the
development of this protocol was to assess the best possible target for DBS implantation
for each individual child with secondary dystonia. This major medical endeavor represents
a valuable opportunity to collect unique data and to explore fundamental neuroscience
questions, however, given the novelty of the procedure, it is necessary to establish the
adequate recording and signal processing methodology. In this context, simultaneous
recordings of neural activity were collected from all implanted brain areas while the
patients were fully awake and performing behavioral tasks that would not be possible to
perform otherwise during the traditional stereotactic procedure for DBS implantation. In
brief, all medical and technical aspects of this setup are described.
Clinical rationale
Contrary to primary dystonia in which electrical stimulation of the GPi has been
highly effective to alleviate primary motor symptoms, in secondary dystonia the optimal
DBS target is not well defined, and it is likely to vary from subject to subject. This situation
is analogous to resection surgery for refractory epilepsy where the epileptogenic focus is
6

unknown and thus it is imperative to localize it with extreme precision prior to the surgery.
Unlike resection, DBS is a reversible procedure, nonetheless, the precise target
localization is critical for efficacy. We have developed a similar technique for identification
of epileptic foci to identify targets for deep brain stimulation and to record neural activity
from key relevant brain areas involved in the coordination of movement. This new protocol
has led to improved clinical outcomes in patients with secondary dystonia (Sanger et at.
2018).
Patient selection
Inclusion criteria for surgical includes:
1) Presence of dystonia that results in significant limitation of function or
interference with care.
2) Identifiable potential stimulation targets on MRI.
3) Failure of symptomatic medical therapy.
Low level of cognitive function or the presence of other movement disorders are
not exclusion criteria. Dystonia is diagnosed by a pediatric movement disorder specialist
(Terrence D. Sanger) based on published criteria (Knights et al. 2013; Jethwa et al. 2010).
All patients underwent informed consent for surgical procedures in accordance with
standard hospital practice. Patients or parents of minor patients also sign informed
consent for research use of electrophysiological data (University of Southern California
Human Subjects Institutional Review Board approval UP-13-00521, 15 November 2013
7

to 12 September 2018) and HIPAA authorization for research use of protected health
information.
Electrode implantation
We have previously described the details of the electrode implantation elsewhere
(Sanger et at. 2018). In brief, a summary of the key aspects of the procedure are here
summarized. Potential targets for stimulation were identified by consultation with the
Department of Neurology and Department of Neurosurgery of the Children’s Hospital Los
Angeles and candidate targets were based on reports in the literature of efficacy in
dystonia through either ablation or stimulation in thalamic and basal ganglia nuclei. Each
child had a preoperative anatomical scan performed. A head frame was attached under
general anesthesia, and computer tomography (CT) scan was performed. The CT scan
was then co-registered to the preoperative magnetic resonance imaging (MRI), and
stimulation targets on MRI were referenced to the frame coordinates. Up to 10 Ad-Tech
MM16C depth electrodes (Ad-Tech Medical Instrument Corp., Oak Creek, WI, USA) were
placed through drill-holes using frame guidance. The frame was removed after the
surgery, and a postoperative CT scan was performed to confirm location and to ensure
lack of hemorrhage or other complications.
Recording electrodes
Standard microelectrodes used to record single units from deep brain structures
are not designed for chronic implantation and their use is limited to intraoperative settings.
Clinically approved stereo EEG multielectrode leads that can be implanted and
8

externalized for chronic recordings are significantly different from standard
microelectrodes. While sharp microelectrodes are designed to approach point current
source potential and to minimize the damage to the neural tissue, stereo EEG leads are
designed for stimulation and local field potentials (LFP) recordings. In consideration of
these limitations, the choice of the recording electrode was made by prioritizing the clinical
outcome during the procedure. The temporary recording electrode must be similar to the
DBS permanent implant in order to emulate the effects of clinical stimulation and to better
predict the effects of DBS treatment. Two options were considered. The Benhke-Fried
depth electrode consists of 8 macroelectrodes distributed along the lead and a bundle of
9 microwires that protrude from the tip of the lead and insert further down to minimize
damage to the tissue of the recording region. Although the large surface area of the
macroelectrodes are suitable for high current stimulation, and microwires are a close
approximation to the standard intraoperative microelectrode, they protrude 20mm away
from the nearest macroelectrode, therefore, these microelectrodes would be located
away from the target area in which the macroelectrode must be placed in order to
stimulate the intended region. An alternative option is the Ad-Tech MM16C lead which
consists of six platinum macroelectrodes and ten platinum-iridium microelectrodes
distributed along the 1.3 mm diameter lead (figure 2.1). The macroelectrodes are similar
in shape and size to the permanent DBS electrodes, however, the spacing between the
contacts is significantly larger (5mm vs 1.5mm). The microelectrodes are microwires of
50um diameter with cylindrical tip that protrudes slightly from the lead’s insulator.
Although the geometry and configuration of the microelectrodes represent some issues
to consider when analyzing the corresponding recordings, the MM16C lead was chosen
9

given that it provides the best tradeoff between emulating the permanent DBS electrodes
and its capability for multielectrode recordings around the intended target and along the
trajectory.

Figure 2.1: Schematic of the Ad-Tech MM16C lead.

The Ad-Tech MM16C macro and microelectrode impedances were measured
using electrochemical impedance spectroscopy (EIS) in a bath of phosphate buffered
saline solution (PBS). The mean microelectrode impedance was 87.86 kΩ (std = 9.43 kΩ)
and the mean macroelectrode impedance was 318 Ω (single electrode measurement).
Recording system and recording configuration
The recording leads were connected to active Ad-Tech Cabrios™. The cabrios
connect each microelectrode to a unity gain amplifier to reduce noise interference
between the recording site and the front-end amplifier, effectively shielding the
microelectrode channels. The macroelectrodes bypass this buffer to allow electrical
stimulation. The cabrios were connected to a PZ5M-256 Medically Isolated NeuroDigitizer
connected to an RZ2 BioAmp Processor which connected to an RS4 Data Streamer unit
(System 3, Tucker-Davis Technologies Inc., Alachua, FL, USA). This system was chosen
because of the wide dynamic range of the PZ5 amplifiers that prevent them from
10

saturating during stimulation protocols, and because it is modular, which allowed
upgrading different components throughout the iterative process of setting up this
protocol.

Figure 2.2: Schematic of the unity gain amplifier (buffer).

All microelectrodes were referenced to the macroelectrode most distal to the tip of
the lead (macroelectrode 6 in figure 2.1) and sampled at 24 kHz. Similarly, all
macroelectrodes were referenced to a separate distal microelectrode. The preferred
reference location was GPi, however, the final reference was selected based on the most
distal macroelectrode that reduced the overall background noise level in the recording
microchannels. A scalp EEG electrode attached to the forehead was used as the system
ground.
Surface electromyography (EMG) was collected using a 16 channels Delsys
Trigno™ system (Delsys Corp., Natick, MA, USA). The EMG sensors were attached to
the biceps, triceps, wrist flexor and extensor groups, quadriceps, medial hamstrings,
tibialis anterior, and medial gastrocnemius muscles. EMG was sampled at 2 kHz and
synchronized with the neural recordings using an external trigger.
11

3 Spike detection from low impedance microelectrode
recordings
Single unit analysis has been the cornerstone for neuroscience to understand how
the nervous system processes and transmits information. Traditionally, single units have
been recorded from sharp microelectrodes that can capture extracellular action potentials
(also known as single units or spikes) generated by the neurons near the tip. A single
electrode can record several spike waveforms together with background neural activity
that includes synaptic potentials and unidentifiable spikes from distant neurons (Quian
Quiroga and Panzeri 2009). These spike waveforms can be detected and classified based
on their shape in a process known as spike sorting. Sharp microelectrodes allow clear
observation of the spike trains (high signal-to-noise ratio, or SNR), which facilitates single
unit isolation. Recently, the recording technology has evolved toward multielectrode
arrays with insulated blunt microwires or silicon probes which physical properties affect
the spatial and temporal characteristics of the recorded signals (Garonzik et al. 2002;
Jonker et al. 2013; Camuñas-Mesa and Quiroga 2013; Humphrey and Schmidt, n.d.).
While large electrodes can record average activity from a relatively large volume of
neurons, small microelectrodes can detect more localized neural activity and single units.
The microelectrodes used in this recording setup are larger than the standard, but small
enough to register spiking activity (Misra et al. 2014; Kano et al. 2008; (Yokoyama et al.
1998). This results in low SNR signal in which the spike waveforms are buried in the
background neural noise as well as other sources of interference. When faced with low
SNR conditions, spike sorting algorithms tend to perform poorly, detection rates are low,
12

and spikes form different sources tend to be clustered together in what is called multi
units.
In this chapter, a new spike sorting method to isolate single units from low SNR
signals is presented. The spike detection rate and cluster classification are improved by
introducing a generalization of the Teager-Kaiser nonlinear energy operator which
accounts not only for the spike amplitude but also for the instantaneous energy content
of the spike waveform. In spite of the imponderables affecting the neural recordings such
as the impedance and shape of the electrode, its placement relative to nearby neuron,
and the inflammatory process in response to the electrode insertion, the proposed
algorithm increases the probability of spike detection by accentuating the desired signal
component of the noisy recordings. The algorithm is efficient and robust to high noise
content and the selection of the detection threshold, which makes it suitable for
multielectrode recordings and real-time applications. Next, the general approach for spike
sorting is introduced. Then, the specific methodology and proposed algorithm is explained
and tested on synthetic data.
Introduction
Signal preprocessing for multielectrode recordings
Little attention has been focused on the necessary signal processing in preparation
for spike sorting. Improving the SNR in the recording channels is essential to increase
detection rate, maximize cluster differentiation, and decrease artifactual correlations
between spike trains. Several sources of noise can affect the quality of the signal. When
13

using multielectrode arrays, the sources of noise can be local (specific to each electrode),
and shared (across the electrode array). Primary sources of local noise include high and
low frequency extracellular local field potentials from the activity of the mixture of neurons
surrounding the electrode, as well as the Johnson–Nyquist noise, that is characteristic of
any electrode–electrolyte interface and it is generated by the thermal agitation of the
charge carriers inside the electrical conductor at equilibrium (which happens regardless
of any applied voltage). Shared sources of noise include local field potentials, spikes, and
the Johnson–Nyquist noise from the reference electrode, motion artifacts generated in
the wiring harness during the experimental procedure, 60Hz harmonics from the electrical
power, and other high frequency electromagnetic interference (EMI).
Local noise can be attenuated by filtering the signal. A high-pass filter removes
low frequency fluctuation and it is essential for setting a threshold amplitude for spike
detection. A low-pass filter reduces interference from spikes generated by the pool of
active neurons distant to the electrode and EMI. With respect to the filter type, a non-
causal zero-phase filter is preferred for offline applications because it can be easily
implemented without introducing filtering artifacts. Causal filters are typically implemented
on the acquisition hardware to remove low frequency components, but they have a
nonlinear phase response that can distort the signal of interest, therefore the low cutoff
frequencies must be far from the frequencies of interest.
Shared sources of noise can be attenuated by removing common patterns of
fluctuation across the recorded signals. Common techniques include multiple linear
regression, signal whitening and common average referencing. Multiple linear regression
(MLR) is used to compute how much a linear combination of the N-n channels can predict
14

the noise present in channel n using some optimization criteria (i.e. least squares), then,
an iterative process subtracts the estimated noise from each n channel. Whitening refers
to a linear transformation from correlated signals x with covariance Σ, into a set of white
noise vectors z with uncorrelated components, such that z = Wx, with the whitening matrix
W satisfying the condition W
T
W = Σ. Zero-phase component analysis (ZCA) was first
introduced by (Bell and Sejnowski 1997). It uses the solution W = Σ
-½
and singular value
decomposition of the empirical covariance of x, Σ = UΛU
T
, such that W = UΛ
1/2
U
T
, were
U and Λ are the eigenvectors and eigenvalues of Σ. Although computationally efficient,
this method assumes constant variance and can introduce errors when applied to long
recordings with non-stationary background noise. Common average referencing (CAR) is
a simple and efficient method that computes the empirical mean across channels at every
time instance, therefore it only depends on present values (Ludwig et al. 2009), however,
the mean can be biased by outliers such as high amplitude spikes or motion artifacts that
are not common to all channels. A more robust solution to approximate the common
background noise is the median. The common median reference (CMR) is less biased by
outliers and it contains less interference from spikes present in channels with active units,
reducing the possibility of introducing artifactual spikes into the less active channels. It is
also computationally efficient and simple to implement.
The Teager-Kaiser nonlinear energy operator for spike detection
Traditional spike detectors rely on a voltage threshold which is based on some
assumption of the background noise distribution. Static threshold is computed from the
entire distribution, whereas adaptive thresholds use small windows to compute piecewise
distributions which are particularly useful when the recording times are long and the noise
15

is non-stationary. More sophisticated classes of spike detectors include time-frequency
decomposition of the neural signal using a family of wavelet functions (Kim and Kim
2003), and parametric mixture models that account for non-stationarity changes in the
signal which tend to change the spike shape causing cluster drifts and therefore affecting
the performance of the sorting algorithm (Shan, Lubenov, and Siapas 2017; Calabrese
and Paninski 2011). Despite the choice of spike detector, the sensitivity and specificity of
the method significantly decrease with low SNR conditions (Wild et al. 2012). For this
reason most algorithms are typically tested for neural recordings with SNR above 0 dB
(Obeid and Wolf 2004; Ekanadham, Tranchina, and Simoncelli 2014; Wild et al. 2012),
however, advances in multielectrode recordings required to address this issue.
One approach to improve detection rate in signals with high noise content, consist
in maximizing the distance between the background noise and the spike amplitude
distributions. A common technique is to transform the detection signal using energy
operators. A typical choice is to square and smooth the neural signal. A more integral
solution that accounts not only for the spike amplitude but also for the frequency
characteristics of the spike waveform, can be found in the study of nonlinear speech
modeling. In their work, Teager and Keiser studied speech signals from the point of view
of the energy required to generate them (H. Teager 1980; H. M. Teager and Teager 1990;
Kaiser 1990). They pointed out that the average energy content in one second of a 10 Hz
sinusoid wave is the same that the average energy in one with frequency of 100 Hz in the
same time period, however, the required energy to generate a 100 Hz sinusoid is much
larger. Based on this observation, Teager described that in a harmonic motion 𝑥 (𝑡 ), the
16

instantaneous energy required to generate an oscillating signal is given by the square of
the product of the signal’s amplitude A and the signal’s frequency 𝜔 , such that
𝑥 (𝑡 ) = 𝐴 𝑐𝑜𝑠 (𝜔𝑡 + 𝜑 ); 𝐸 [𝑥 (𝑡 )] ∼ 𝐴 2
𝜔 2

where 𝜑 is the arbitrary initial phase. If 𝑥 is a discrete signal, the harmonic motion equation
can be expressed in terms of its sampling rate 𝑓 𝑠 and the oscillation frequency 𝜔
𝑥 𝑛 = 𝐴 𝑐𝑜𝑠 (𝛺𝑛 + 𝜑 ); 𝛺 =
2𝜋 𝑓 𝑠 𝜔
where n is the sample instance and 𝛺 is the digital frequency of the oscillation in
radians/sample. From the above equation, it can be shown that if 𝛺 is small, then
𝐴 2
𝛺 2
≈ 𝑥 𝑛 2
− 𝑥 𝑛 +1
𝑥 𝑛 −1
= 𝛹 (𝑥 𝑛 )
where 𝛹 (𝑥 𝑛 ) is the Teager-Kaiser nonlinear energy operator (NEO). The extracellular
action potential is characterized by a large and fast depolarization followed by a period of
a relatively slow repolarization. The NEO output is proportional to the product of the
instantaneous frequency and amplitude of the spike waveform, therefore, exalting the
action potential transient (Mukhopadhyay and Ray 1998). Moreover, the implementation
of this operator is computationally efficient and can be applied to multichannel processing
and online spike detection.
Feature extraction and spike clustering
The first algorithms used to identify spike waveforms from neural recordings
include window discriminators and template matching. The window discriminators are
defined in time and amplitude and the waveforms crossing them are assigned to a
particular unit (Rodrigo Quian Quiroga 2012). Template matching uses a reference spike
waveform and finds similarities between waveforms based on a distance metric (Gerstein
17

and Clark 1964). These algorithms are computationally efficient, but require constant
manual adjustment, especially under low SNR conditions, making them unfeasible for
multichannel processing. More robust methods involve mapping a measure of individual
spike shape to a point in multi-dimensional feature space for subsequent classification
into clusters (neurons). Principal component analysis (PCA) is a dimensionality reduction
method that maps each data point of the spike shape (dimensions) into principal
components (orthogonal axis) aligned in the direction of maximum variation. Typically,
the first three or four principal components are used as inputs to the clustering algorithm
to classify the spike waveforms. PCA is the most common technique used in spike sorting
and has proven adequate for high SNR recordings. Alternatively, the spike can be
decomposed into a set of basis wavelet functions that can be dilated and shifted to best
capture the time-frequency characteristics of the waveform
𝜓 𝑎 ,𝑏 (𝑡 ) = |𝑎 |
−
1
2
𝜓 (
𝑡 − 𝑏 𝑎 )
where a and b are the scale and translation parameters. The choice of the basis wavelet
is a balance between the best fit of the spike waveform and the computational cost.
Algorithms that use complex wavelets that fit the action potential waveform are
computationally more expensive than ones that use simpler Haar wavelets which tend to
overlook important features of the spike waveform.
Clustering algorithms can classify data points in the multidimensional feature
space based on a measurement of similarity between two or more classes. Typically, this
measurement is the distance between the spike features, under the principle that nearby
points are more related than points that are further apart. The superparamagnetic
18

clustering (SPC) algorithm was first proposed by Blatt in 1996 ((Blatt, Wiseman, and
Domany 1996; R. Quian Quiroga, Nadasdy, and Ben-Shaul 2004), and implemented for
spike classification by Quiroga in 2004 (Blatt, Wiseman, and Domany 1996; R. Quian
Quiroga, Nadasdy, and Ben-Shaul 2004). The algorithm follows the model of spins in a
crystalline lattice changing their state as a function of the temperature. With high
temperatures, spins will randomly change their state regardless of the interaction strength
between them (paramagnetic phase). At low temperatures, the spine glass changes its
state together (ferromagnetic phase). At mid-range temperatures, spins that share some
interaction strength will be more likely to change their state simultaneously, in which is
called the superparamagnetic phase. The implementation of the algorithm relies on K-
nearest-neighbors to compute the level of “attraction” between data points which changes
based on only one parameter, the temperature. The advantage of this algorithm is that it
does not assume any structure of the underlying distribution of the data and it uses only
the temperature parameter to adjust the cluster size.
Methods
The goal of the algorithm presented here is to increase the probability of spike
detection and correct classification by compounding the high energy content and
broadband frequency features of the spike waveform. The proposed algorithm is a
generalization of the Teager-Kaiser nonlinear energy operator for the estimation of the
instantaneous energy of a time-varying signal. This generalization incorporates multiple
sample shifts of the nonlinear energy operator (MSNEO), to estimate the spike’s
instantaneous energy content in discrete frequency bands, which can be used as the
19

spike detection signal as well as the input features to the clustering algorithm. In brief, the
deduction of the operator is explained.
Multiple Sample Nonlinear Energy Operator (MSNEO)
The Teager-Kaiser NEO is a particular case of high-order energy operators that
quantifies the instantaneous differences in the relative rate of change between a
continuous-time signal and its first derivative. The discretization of this operator uses one-
sample symmetric differences to approach derivatives. Kaiser described the discrete
NEO output as an estimation of the instantaneous energy required to produce a harmonic
motion that can be calculated by the difference in the relative energy between one sample
point n and two adjacent n 土1 samples along any given signal
𝛹 (𝑥 𝑛 ) = 𝑥 𝑛 2
− 𝑥 𝑛 +1
𝑥 𝑛 −1

The algorithm proposed here generalizes this idea to a higher order symmetric
differences, such that
𝛹 (𝑥 𝑛 ,𝑘 ) =
𝑥 𝑛 2
− 𝑥 𝑛 +𝑘 𝑥 𝑛 −𝑘 𝑘 2

Similarly to Kaiser’s original work (Kaiser 1990), this operator can be deduced from
any three given points in the harmonic motion
𝑥 𝑛 = 𝐴 𝑐𝑜𝑠 (𝛺𝑛 + 𝜑 ) , 𝑥 𝑛 −𝑘 = 𝐴 𝑐𝑜𝑠 (𝛺 (𝑛 − 𝑘 ) + 𝜑 ) , 𝑥 𝑛 +𝑘 = 𝐴 𝑐𝑜𝑠 (𝛺 (𝑛 + 𝑘 ) + 𝜑 )
where n is the sample instance, k is the sample shift, 𝛺 =
2𝜋 𝑓 𝑠 𝜔 is the digital frequency in
radians/sample, 𝜔 is the instantaneous analog frequency of oscillation in Hz, and 𝑓 𝑠 is the
sampling frequency in Hz. By the trigonometric identity
𝑐𝑜𝑠 (𝛼 + 𝛽 ) 𝑐𝑜𝑠 (𝛼 + 𝛽 ) =
1
2
[𝑐𝑜𝑠 (2𝛼 ) + 𝑐𝑜𝑠 (2𝛽 )]
20

we obtain
𝑥 𝑛 +𝑘 𝑥 𝑛 −𝑘 =
𝐴 2
2
[𝑐𝑜𝑠 (2𝛺𝑛 + 2𝜑 ) + 𝑐𝑜𝑠 (2𝑘𝛺 )]
using
𝑐𝑜𝑠 (2𝛼 ) = 2 𝑐𝑜 𝑠 2
𝛼 − 1 = 1 − 2𝑠𝑖 𝑛 2
𝛼
results in
𝑥 𝑛 +𝑘 𝑥 𝑛 −𝑘 = 𝐴 2
𝑐𝑜 𝑠 2
(𝛺𝑛 + 2𝜑 )−𝐴 2
𝑠𝑖 𝑛 2
(𝑘𝛺 )
Substituting 𝑥 𝑛 in the above equation leads to
𝐴 2
𝑠𝑖 𝑛 2
(𝑘𝛺 ) = 𝑥 𝑛 − 𝑥 𝑛 −𝑘 𝑥 𝑛 +𝑘
Note that the above equation is independent of the initial phase 𝜑 . If 𝛺 → 0, then
𝐴 2
𝑘 2
𝛺 2
≈ 𝑥 𝑛 − 𝑥 𝑛 −𝑘 𝑥 𝑛 +𝑘 ⇒ 𝛹 (𝑥 𝑛 ,𝑘 ) =
𝑥 𝑛 − 𝑥 𝑛 −𝑘 𝑥 𝑛 +𝑘 𝑘 2

where 𝛹 (𝑥 𝑛 ,𝑘 ) is the generalization of the Teager-Kaiser NEO for k symmetric differences.
This approximation can be expressed in terms of the Taylor series of 𝑠𝑖𝑛𝑒 for a small
angle 𝛽
𝑠𝑖𝑛 (𝛽 ) ≈ 𝛽 −
𝛽 3
6

where the second term is a rough estimation of the error. For 𝛽 =
𝜋 4
the relative error is
always below 11%, and the definition of 𝛺 provides the relationship between k and the
instantaneous frequency of oscillation 𝜔
𝛽 = 𝑘𝛺 = 𝑘 2𝜋 𝑓 𝑠 𝜔 =
𝜋 4
⇒ 𝜔 𝑘 =
𝑓𝑠 8𝑘
Therefore, the operator 𝛹 (𝑥 𝑛 ,𝑘 ) provides an approximation of the energy content
at frequency bands which are a function of the sample shifts. In other words, the output
21

of 𝛹 (𝑥 𝑛 ,𝑘 ) is a spectrogram that contains the estimated local energy of the signal 𝑥 at
frequency bands 𝜔 𝑘 for every sample 𝑛 .
Spike detection using MSNEO
To compute the detection signal, the raw input signal was high-pass filtered with a
cutoff frequency of 200 Hz to remove local field potentials and other slow fluctuations.
Linear filters can correlate artifactual trancients with spike waveforms, especially if the
cutoff frequencies are close to the bounds of the spike bandwidth, hence, no low-pass
filter was used. The values of k were computed for a typical spike bandwidth between
300 and 4000 Hz, then, the output of the operator at each frequency band was convolved
with a triangular (Bartlett) window of 1 ms duration. This particular window was chosen to
sum the energy content of the signal in a time period equivalent to the duration of a typical
action potential, moreover, this operation smooths the output signal and the shape of the
window contributes to improve peak detection. The detection signal was then computed
as the product of all the frequency bands convolved with at every sample
𝑋 𝑛 = ∏
𝑘 (𝛹 (𝑥 𝑛 ,𝑘 ) ⊗ 𝑊 )
This operation results in a signal which peaks are exponentially larger for
broadband high energy transients (such as spikes) compared to narrowband low energy
fluctuations, providing maximum separation between the spike and background noise
distributions, therefore increasing the chances of correct spike detection. All peaks higher
than the median of 𝑋 𝑛 were considered candidate spikes. Spike waveforms were
extracted for a time window from -0.5 to 1.5 ms, interpolated by a factor of two using cubic
22

splines, realigned by maximum absolute peak, and downsampled back to 24 kHz using
linear interpolation.

Figure 3.1: Example of the resulting frequency decomposition and detection signal using
MSNEO. Top: sample spike waveform extracted from sharp microelectrode recordings. Middle:
MSNEO spectrogram. Bottom: gray traces indicate the energy content at each frequency band.
Red trace is the detection signal computed as the product of all frequency bands.
Feature extraction and clustering
The peak values associated with the candidate spikes were log transformed and
divided into 100 bins of equal width which were assigned to values from 0 to 1. This
generates a distribution which values are associated with the probability of the waveform
to be a spike. Additionally, the sum of the energy of each frequency bands of the MSNEO
23

spectrogram can be computed and used as input features. These features provide
information about the energy and frequency content of the spike. In order to include
information about the spike shape, the first two principal components were added to the
feature space. The spike features were classified using the superparamagnetic clustering
(SPC) algorithm which was implemented as described by Quiroga (R. Quian Quiroga,
Nadasdy, and Ben-Shaul 2004), using the subroutine do_clustering of the Wave_clus
software package (Chaure, Rey, and Quian Quiroga 2018). The implementation of the
algorithm is a Monte Carlo iteration on simulated interactions between each spike in the
multidimensional feature space and its K-nearest-neightbors. The probability of two spike
features changing state simultaneously is a function of the ratio between the Euclidean
distance of its K-nearest-neighbors and the temperature, which is a free parameter that
is systematically increased to maximize the cluster size, and spike features that are close
to each other change their state together and are clustered together.
Simulated spike trains and background noise construction
To quantify the performance of the spike sorting algorithm under low SNR
conditions, synthetic spike trains were constructed with a typical sampling rate of 24 kHz.
A spike waveform was extracted from standard microelectrode recordings of the GPi
obtained during stereotactic placement of DBS. Spike trains of 4 minutes long were
created by drawing time events from a Poisson distribution with rates randomly selected
from 1 to 30 spikes per second and minimum refractory period of 1.5 ms. Ten channels
of Gaussian noise (mu = 0, 1<𝜎 <2) were generated to simulate the effects of the variability
24

Figure 3.2: Proposed spike sorting algorithm. The discrete multichannel signal X(n) is high-pass
filtered (HPF) to remove low frequency field potentials, then, the common median reference is
computed from all channels and subtracted from each HPF signal, the resulting signal is input to
the operator 𝛹 (𝑥 𝑛 ,𝑘 ), as well as band-pass filtered for spike waveform extraction (BPFs). The
total energy per frequency band around the spike index 𝛴𝜔
𝑘 can be computed and input to the
superparamagnetic clustering (SPC) algorithm. The spike indices are found based on the
detection signal ∏Ψn,k and assigned with a probability P(s) which is an input to the SPC. Spike
waveforms are extracted from the BPFs and three first two principal components are computed
(PCA) and input to the SOC.

in electrode impedances. Shared sources of noise were added together in a common
reference signal, including reference electrode impedance (Gaussian, mu = 0, std = 1),
60Hz power noise, and artifact transients modeled as exponential decays with aleatory
time constants. Neural cross-talk between Ad-Tech microelectrodes can be neglected
due to the relatively large space between them. It has been shown that the detection
radius of a standard microelectrode is approximately 150um (Quian Quiroga and Panzeri
2009). The Ad-Tech microelectrodes are distributed around the circumference of the
1.3mm diameter insulator, therefore, given the space between them (~ 2mm) it is unlikely
25

that two or more separate electrodes can record from the same neuron. Spike trains were
normalized by the standard deviation of the spike waveform and added to the simulated
background noise which was band-passed filtered (200 Hz to 10 kHz) and scaled based
on the target SNR level
𝑥 = 𝑥 𝑠 𝑝𝑖𝑘𝑒 𝑡𝑟𝑎𝑖𝑛 +
1
𝑆𝑁𝑅 𝑥 𝑛𝑜𝑖𝑠𝑒
Common noise reduction
Four methods for common noise reduction were tested: multiple linear regression
(MLR); channel whitening by zero-phase component analysis (ZCA); common average
reference (CAR), and common median reference (CMR). The robustness of each method
to spike interference, recording length and transient artifacts was tested by varying the
spike density in each channel, as well as the simulated recording time from 1 to 20
minutes, and comparing the relative error between the simulated signals free of common
noise and the common-subtracted signals.
Algorithm performance
To compare the MSNEO performance, two other methods for spike detection were
tested: the standard absolute amplitude threshold (AT), and the smoothed Teager-Kaiser
nonlinear energy operator (sNEO). The AT detection signal was computed as the
absolute value of the band-pass filtered (300 - 4000 Hz) mixture of background noise and
spikes. The threshold was calculated as
𝑡 ℎ𝑟𝑒𝑠 ℎ𝑜𝑙𝑑 = 𝑇 𝑚𝑒𝑑𝑖𝑎𝑛 (
| 𝑥 𝐴𝑇
|
0.6745
)
26

where T is a free parameter typically between 1 and 4. The sNEO detection signal was
computed as the convolution of the one sample NEO output with a Bartlett window of 1
ms length
𝑥 𝑠𝑁𝐸𝑂 = ( 𝑥 2
𝑛 − 𝑥 𝑛 −1
𝑥 𝑛 +1
) ⊗ 𝑊
The detection threshold was calculated as T times the standard deviation from the
mean M. For comparison at the final output of the spike sorter, the detected spike
waveforms were clustered in the space of principal components using the SPC algorithm
previously described.
Results
Common noise reduction
Figure 3.3 shows the percentage of error introduced by the different methods
tested for common subtraction and channel decorrelation. Overall, MLR performed poorly
compared to the other methods, therefore it was excluded for future analysis. The
robustness of ZCA, CAR and CMR was tested for changes in firing rate and recording
length. Long recording lengths imply higher number of spikes and increased probability
of artifactual transients (such as motion artifacts). Figure 3.2 shows that ZCA and CAR
performed well for short recording times and firing rates below 20 sps, however, CMR
was more robust to transients and introduced less error for higher firing rates and
recording time. Hence, CMR was the common subtraction method used for the rest of the
analysis.
27

Figure 3.3: Percentage of error introduced by each common subtraction method. Left: average
error for 1 minute of simulated recording and firing rate of 10 spikes/second. Right: heatmaps of
the recording length and spike frequency.
MSNEO spectrogram
Figure 3.4 shows the MSNEO estimation for a typical spike waveform sampled at
24 kHz. The heat maps show the comparison between the MSNEO spectrogram and the
continuous wavelet transform (CWT) for the spike waveform shown in the top left corner.
The MSNEO spectrogram was computed for a spike frequency bandwidth between 300
and 4000Hz according to 𝑘 = 𝑓 𝑠 /8𝜔 , i.e., 1 ≤ 𝑘 ≤ 10. The bottom left graph shows the
comparison between the multitaper power spectral density (PSD) and the total MSNEO
frequency content. Note that the estimation error is higher for the highest frequency
corresponding to 𝑘 = 1. This can be explained by the error introduced by the
discretization of the signal which is a function of the sampling frequency. The relationship
between 𝑘 and 𝜔 is given by 𝜔 = 𝑓 𝑠 /8𝑘 , meaning that if 𝑘 = 1, the sampling rate must be
8 times higher than the instantaneous frequency of oscillation for the error to be less than
11%. The CWT estimation shows that the mean spike waveform has high energy content
in frequencies close to 4 kHz, which would require 𝑓𝑠 ≥ 32 kHz for a more accurate
estimation at higher frequency bands. It has been reported that the spectrum of the
variability between different instances of a spike waveform in the raw microelectrode
recordings falls exponentially for frequencies above 1 kHz (Fee, Mitra, and Kleinfeld
28

1996), as shown by the PSD in figure 3.4, therefore, a sampling rate of 24 kHz may be
sufficient for the spectral estimation of most type of neurons.
Spike detection performance
The MSNEO detection signal was robust to noise interference. Figure 3.5 shows
examples of a spike waveform and additive noise at different SNR levels within the spike’s
frequency bandwidth (200 - 5000 Hz). From the middle (spectrograms) and bottom
(detection signal) panels of the figure, it can be observed that the product operator
amplifies the broadband energy content of the spike spectrum, while attenuating the high
energy but narrowband frequency content of the background noise, therefore maximizing
the distance between the spike and background noise distributions and increasing the
probability of correct spike detection.

Figure 3.4: Sample detection signals for several SNR levels (columns). Top: noisy spike
waveforms. Middle: estimated MSNEO spectrogram. Bottom: frequency bands (gray) and
detection signal (red). Note that the detection signal is robust to increased noise levels.
29

Figure 3.5 shows the receiver operator characteristics (ROC) curves at the output
of each spike detector. The performance of the MSNEO spike detection was compared
to the sNEO and the standard AT operators. MSNEO Improved the chances of detection
even at very low SNR levels. For SNR ≥ 1, the area under the SNR curve was close to 1.
The performance of the algorithm was also tested at the output of the SPC. Figure 3.6
shows the ROC curves for the spike clusters. The MSNEO improved spike classification
for low SNR levels and was more robust to the selection of detection thresholds for higher
SNR levels. AT was more accurate for correct spike classification given the right
threshold, but its sensitivity falls quickly for lower thresholds. Spike classification can be
further improved by adding relevant extra features to the clusters. Figure 3.7 shows that
by including the peak detection as an input feature, the SPC can more easily isolate true
spikes from spike-like waveforms.

Figure 3.5: Receiver Operator Characteristics (RO) curve comparing the performance of the
three spike detection methods for several SNR conditions. Note that MSNEO outperforms
sNEO and AT in all cases and for SNR>1 the algorithm shows close to perfect discrimination.
30

Discussion
Current technologies in microelectrodes for human recordings require robust algorithms
that can extract relevant information from noisy neural signals. Here, a new spike
detection algorithm to identify single unit activity from low SNR neural recordings was
presented. The algorithm estimates the instantaneous energy content of the signal at
discrete frequency bands to differentiate spike waveforms from the background noise
even under very low SNR conditions. The performance of the spike detector showed

Figure 3.6: ROC curves comparing the performance of the spike detectors at the output of the
spike sorter for several SNR conditions.

31

Figure 3.7: Contribution of the normalized peak value P(s) of the detection signal to the
improvement of spike identification for a SNR = 0.75. Left: SPC output using 4 principal
components. Right: SPC output using 3 principal components and P(s). Note that the elongated
P(s) feature cluster of the spike (green) provides better differentiation from the noise cluster
(red) than PC4.

significant improvement compared to existing ones, and it was more robust to the
changes of the detection threshold which selection represents a recurring problem in
spike sorting (Ekanadham, Tranchina, and Simoncelli 2014). Furthermore, the simplicity
of the algorithm makes it suitable for the analysis of high frequency multichannel
recordings as well as for online implementation.
Implications for closed-loop deep brain stimulation
In recent years, there has been a growing interest in closed-loop deep brain
stimulation (DBS) for automatic adjustment of the stimulation setting based on the
therapeutic needs of each individual patient. Currently, local field potentials (LFP)
recorded from the stimulating macroelectrode are used to detect oscillatory activity which
correlates with features of the specific movement disorders (Bouthour et al. 2019). The
current technology in DBS electrodes is evolving toward high density electrode leads
which smaller electrodes with the potential to detect single unit activity. While the LFP
signal recorded from macroelectrodes can incorporate average activity of a large volume
32

of neurons surrounding the electrode, single unit detection can provide information about
more localized neural activity which may be relevant to adjust the stimulation settings to
target more specific regions.
Implications for online spike detection
The proposed recursive algorithm is computationally efficient for multichannel
offline processing, however, online applications may require an even simpler version of
the algorithm to improve detection rates with less computational time. Figure 3.9 shows
that an improvement of about 13% in the detection rate can be achieved when k = 3. This
means that if the neural signal is sampled at 24 kHz, the algorithm would only require 7
clock cycles to collect the necessary samples for the estimation of the detection signal to
increase the chances of correct spike identification. Moreover, the algorithm does not
depend on heavy filtering of the input signal because the estimation of the instantaneous
energy occurs at high frequencies and/or for selected frequency bands based on the
chosen number of sample shifts, therefore, the interference of low frequency fluctuation
is neglectable, much like the carrier signal in a frequency demodulation process.
How important is accurate spike isolation?
Spike isolation refers to the correct separation between spike waveforms from
different sources. The necessary level of accuracy in spike isolation depends on the
purpose of the spike detection. While an excellent isolation is crucial to study spike
synchronicity (Pazienti and Grün 2006), for brain computer interfaces the exact
identification of the neuron generating the spike may not be as relevant, in which case
33

the implementation of the detection algorithm without a classification stage can be
sufficient (Fraser et al. 2009).

Figure 3.8: Left: spike waveform and the MSNEO output for several values for k. Right: Area
under the ROC curve for spike detection with MSNEO for several sample shifts k and SNR = 1.
Note that the performance of the detector significantly improves as k increases, with maximum
peak at k = 3, which corresponds to the peak frequency of the spike waveform (1 kHz).

34

4 Patterns of neural activation of the GPi and thalamic nuclei
in children with secondary dystonia may provide evidence
for a selective inhibition/disinhibition model of the basal
ganglia
Current models of basal ganglia function are based on electrophysiological
recordings from single microelectrode in animals or anesthetized humans under the
constraints of the operating room. Our recently developed inpatient multielectrode
recording procedure provides a valuable opportunity to record from the basal ganglia and
thalamus of patients with secondary dystonia without the constraints of the operating
room. The methods for extracting single units from the low impedance microelectrodes to
which this recording setup is constrained have been established in the previous chapter.
In this chapter, firing rates from single unit recordings are analyzed for multiple basal
ganglia and thalamic nuclei in pediatric subjects with secondary dystonia. The correlation
between the neural activation of GPi (internal globus pallidus) and thalamic neurons is
explored during the execution of voluntary movement. Two relevant observations arise
from this analysis. First, the vast majority of the detected neurons in basal ganglia and
thalamus fire with median frequencies lower than 5 Hz, which contrasts with typical
reports of intraoperative GPi discharge rates between 10 and 70 Hz. Second, increased
widespread activation of the GPi positively correlated with thalamic activation during
voluntary movement. This observation contrasts with the traditional rate model of the
basal ganglia which suggests that activation of the GPi results in thalamic inhibition, which
leads to suppression of cortical motor commands to finally inhibit movement. Finally, a
new model of basal ganglia function that accounts for these observations is discussed.
35

This model provides a framework for the modulatory function of basal ganglia over
ongoing thalamocortical dynamics as opposed to widespread thalamic
inhibition/disinhibition.
Introduction
Typically, recordings of subcortical brain areas in humans have been performed
by advancing single microelectrodes through the brain along a fixed trajectory usually
during the implantation of DBS electrodes. These recordings are usually collected either
from the main basal ganglia output or from its projected nuclei in thalamus, but rarely from
both simultaneously, which has made it difficult to convey hypothesis on how these
subcortical nuclei process sensorimotor information to coordinate movement, and how
patterns of abnormality can cause movement disorders. One of the most commonly
accepted models of basal ganglia, the rate model, is based on observation of single
electrode recordings from subjects with movement disorders for whom differences in firing
rates can be correlated with a specific feature of the motor symptoms.
The traditional rate model of basal ganglia function predicts that changes in the
discharge rate of the internal segment of the globus pallidus (GPi, the main basal ganglia
output which is gabaergic and inhibitory) through the differential modulation of the direct
and indirect pathways, would result in inhibition or disinhibition of the projected areas of
the motor thalamus in order to select the appropriate motor command that satisfies the
demands of the task (Albin, Young, and Penney 1989; DeLong 1990; Nelson and Kreitzer
2014; Nambu 2007). This model implies two major assumptions. First, high GPi discharge
rates will inhibit movement, whereas low GPi discharge rates will facilitate movement,
36

therefore, decreased basal ganglia output would correlate with uncontrolled movement.
Although low firing rates have been observed in patients with dystonia (Zhuang, Li, and
Hallett 2004; J. L. Vitek et al. 1999; McClelland et al. 2016), it has been shown that GPi
activity during voluntary limb movements always displays increased rather than
decreased discharge rates (DeLong 1972; Nambu 2007). Moreover, this model is
inconsistent with observations of the alleviation of dystonic symptoms followed by
pallidothomy (Ondo et al. 1998), and normal firing rates in patients with hyperkinetic
(excessive movement) disorders (Tang et al. 2005). The second assumption is that the
information necessary to execute the correct motor command is encoded in the firing rate.
While abnormal patterns of GPi discharge rate have been associated with basal ganglia
dysfunction, the exact pattern of abnormality is still a subject of debate.
Simultaneous multielectrode recordings can help clarifying some of these
confounds by measuring the neural activity from multiple brain areas involved in the
coordination of movement. Current technology on depth multi electrodes for human use
is limited to low impedance microelectrodes which differ significantly from standard sharp
electrodes. In the previous chapter, the MSNEO algorithm to extract single units from low
impedance microelectrodes was presented. Here, MSNEO is applied to extract spikes
from basal ganglia and thalamic recordings from multiple low impedance microelectrodes
and the neural activity is analyzed based on firing rates. First, the characteristics of the
detected spike waveforms are described, then, spike rates are compared with published
reports in a similar subject population using standard intraoperative electrodes. Finally,
observations of spiking activity in GPi and thalamus during the execution of voluntary
37

movement are contrasted with the traditional rate model, and a new model of basal
ganglia function based on these recordings is discussed.
Methods
Subjects
Twelve pediatric subjects (11 with dystonia and 1 with Tourette’s syndrome) were
implanted with up to five Ad-Tech leads. Surgical implantation as well as all recording
procedures were performed at the Children’s Hospital Los Angeles. All patients signed
informed consent for surgical procedures in accordance with standard hospital practice.
Patients or parents of minor patients also sign informed consent for research use of
electrophysiological data (University of Southern California Human Subjects Institutional
Review Board approval UP-13-00521, 15 November 2013 to 12 September 2018) and
HIPAA authorization for research use of protected health information. Table 5.1 shows
the demographics, recording targets, and final implantation targets that were selected
based on the evaluation of electrical stimulation outcomes and electrophysiological
recordings. The diagnosis and primary dystonic symptoms were assessed by the
pediatric movement disorder specialist (T. D. Sanger). Abnormal MRI refers to the
presence of noticeable injury in preoperative imaging studies. The implanted target areas
include the internal globus pallidus (GPi), subthalamic nucleus (STN), and thalamic nuclei
VA, Voa/Vop (Vo, for conciseness), Vim, and VPL (Hirai and Jones 1989). Voa/Vop and
VA are the output projection areas from GPi, Vim is the projection area from cerebellar
nuclei, and VPLa is the projection area from the spinothalamic and lemniscal sensory
systems.
38

Data collection
Data were collected between the second and third day after implantation in
sessions of one hour minimum during which subjects were fully awake. Blocks of resting
state periods of four minutes each were intercalated with blocks of active finger-to-noise
reaching tasks of similar duration. Video was recorded continuously and synchronized to
the microelectrode recordings. Surface electromyography (EMG) was recorded using up
to 16 Delsys DE2.1™ or Delsys Trigno™ electrodes (Delsys Corp., Natick, MA, USA)
placed on the biceps, triceps, wrist flexor and extensor groups, quadriceps, medial
hamstrings, tibialis anterior, and medial gastrocnemius muscles.
Raw microelectrode recordings were sampled at 24 kHz and stored for offline
analysis. All microelectrodes were referenced to the macroelectrode most distal to the tip
of the implanted lead. The prefered reference location was GPi, however, the final
reference was selected based on the most distal macroelectrode that reduced the overall
background noise level in the recording microchannels. A scalp EEG electrode attached
to the forehead was used as system ground.
Data analysis
The neural recordings were processed using the MSNEO spike sorting algorithm
previously described in chapter 4 which was implemented on Matlab (The MathWorks,
Inc., Natick, Massachusetts, US). The common median reference was computed for all
the microchannels in the same lead after applying a digital high-pass filter (200 Hz, 6th
order, Butherworth). This common subtraction method significantly reduced the number
of spike-like artifacts and decreased the background noise amplitude by a factor of two
39

Table 4.1: Demographics, temporary recording lead locations, and final selected targets.
ID Age Diagnosis
Type of
dystonia
Primary
Symptom
MRI
Recording
targets
1 15.6
Unilateral
idiopathic dystonia
Secondary Hypertonia Abnormal
GPi, VoaVop,
Vim, VPLa
2 8.6 CP Secondary Hypertonia Abnormal
GPi, Vo, Vim,
VA
3 10.6
Autoimmune
hemidystonia
Secondary Hypertonia Abnormal
GPi, Vo, Vim,
VA
4 15.1 CP Secondary Hypertonia Normal
GPi, Vo, Vim,
VA
5 12.4 Idiopathic dystonia Primary Hyperkinesia Normal
GPi, Vo, Vim,
VA
6 10.6
ADCY5
Mutation/generaliz
ed dystonia
Primary Hypertonia Abnormal
GPi, STN, Vo,
Vim, Vm, VA
7 20.8 CP Secondary Hypertonia Abnormal
GPi, STN, Vo,
Vim, VPLa
8 17.9 CP Secondary Hypertonia Abnormal
GPi, Vo, Vim,
VA
9 14.5 CP Secondary Hypertonia Abnormal
GPi, Vo, Vim,
VA
10 9.8 CP Secondary Hyperkinesia Abnormal
GPi, Vo, Vim,
VA
11 14.4 Idiopathic dystonia Secondary Hypertonia Normal
GPi, STN, NA,
Vo, Vim, CMP
12 15.5 Tourette Syndrome Hyperkinesia Abnormal
GPi, STN, NA,
Vo, Vim, CMP
40

(measured as the ratio between the standard deviation before and after the common
subtraction). Motion artifacts were common, and they had the characteristic of occurring
at the same time in all or most of the microchannels in the same lead, therefore,
timestamps that repeated in more than five microchannels in the same lead in a time
window of 0.1 ms were ignored. Furthermore, an artifact detection threshold of 10 times
the standard deviation of the common-subtracted signals was implemented, and any
event that crossed this level was ignored. The spike waveforms were obtained from
filtering the common-subtracted signals (bandpass 300 - 4000 Hz, 6th order Butterworth)
which were then extracted based on the output timestamps of the MSNEO spike sorter
and stored for a time window of -0.5 to 1.5 ms. The waveforms were then interpolated by
a factor of two using cubic splines, realigned by the absolute maximum peak, resampled
back to 24 kHz, and clustered in the space of principal components. This realignment was
crucial to improve cluster differentiation. After spike sorting, all waveforms were visually
inspected, and artifactual spikes were discarded.
Firing rates were calculated as the median of the inverse of the interspike interval
distribution of each spike cluster and compared across conditions (type of dystonia,
primary dystonic symptom, and MRI). Statistical significance between groups in the same
condition was performed using the two-sample Kolmogorov-Smirnov, which is a
nonparametric hypothesis test that compares the distance between the two cumulative
distributions and it does not assume any distribution.
To analyze the correlation between neural activity across brain areas during
voluntary movement, cross-correlation coefficients were calculated by combining all the
spike cluster in each electrode and convolving them with an exponential decay with 1.5
41

ms time constant (to mimic to the refractory period of a typical action potential). This
method was selected to compute the correlation between the average spike activity per
brain area rather than individual spike trains.
To compute the EMG envelope, the raw recordings were filtered (bandpass 20 -
200 Hz, 4th order Butterworth), and rectified by absolute value. A moving median with a
200 ms window was implemented to filter artifactual transients caused by motion artifacts.
Finally, the moving median output was smoothed by applying a linear low pass filter (20
Hz, 4th order Butterworth).
Results
Spike waveforms
Neural recordings from 1036 microelectrodes were analyzed. Examples of
extracted spike waveforms with high SNR are shown in figure 4.1. The spike SNR was
calculated as the ratio between the powers of the average waveform and the background
noise. The aggregate data shows that, on average, two spike clusters (neurons) were
identified per microelectrode, with median peak-to-peak amplitude of 44.44 uV (min =
16.10 uV, max= 178.60 uV) and median SNR of 1.62 (min = 0.5, max = 9.70). This results
contrast with typical reports of spike waveforms recorded with sharp microelectrodes with
mean amplitudes of hundreds of uV (Dhawale et al. 2017; Suner et al. 2005), but can be
expected given the low impedance of the Ad-Tech microelectrodes (Camuñas-Mesa and
Quiroga 2013).
42

Figure 4.1: Example of isolated spike waveforms with SNR > 1.

Well isolated single units were characterized by high SNR and narrow firing rate
distributions (Figure 4.2). Multi-units are spikes from different sources that are clustered
together, usually from neurons far from the electrode with shapes that are
indistinguishable from the background neural noise. While these spike compounds may
carry significant information regarding the neural activation of the volume of neurons
surrounding the electrode, they tend to have broad firing rate distributions and very high
firing rates that can bias this analysis, therefore, an automatic rejection criteria was
implemented: clusters with SNR < 1 and/or 5% or more of their ISI that follow within 3 ms,
were excluded from further analysis. A final visual inspection of the spike waveforms was
necessary to discard artifactual clusters. Figure 4.3 provides an example of these
common confounds.
43

Figure 4.2: Example of a multi-unit cluster (top row) and a well isolated single unit (bottom row)
and their respective firing rate distributions (right column). Red lines indicate the interquartile
range of distribution.

Figure 4.3: Spike (left) and spike-like artifact (right). Note that the artifact has a high SNR and
firing rates (FR) within the physiological range which makes it difficult to exclude using any
automatic criteria.
44

Firing rate statistics
For comparison with the existing literature, firing rates were analyzed during the
resting periods. The aggregate firing rates per brain area are summarized in table 4.2.
The results indicate a predominance of low firing rate neurons in the basal ganglia and
thalamus, which contrasts with typical reports of mean GPi discharge rates between 30
and 60 Hz in patients with primary dystonia (Vitek et al. 1999; Zhuang, Li, and Hallett
2004; Nambu et al. 2011). However, other studies on a similar subject population have
shown low GPi firing rates that range from 9 to 13 Hz (McClelland et al. 2016), and
bursting activity with frequencies between 0.4 and 0.5 Hz that correlate with spontaneous
muscle activation (Zhuang, Li, and Hallett 2004).
Table 4.2: Median firing rates of the aggregate data
Brain area N Median (Hz) Mean (Hz) Range (Hz)
GPi 396 2.8 14.1 0.1 - 62
STN 37 10 24.3 0.2 - 41.8
VIM 357 3.3 16.2 0.2 - 90.4
VPLa 87 3.3 12.8 0.2 - 29.6
Vo/VA 406 4.1 16.9 0.1 - 48.8

Table 4.3 summarizes the firing rate statistics per subject group and conditions.
For comparison between two groups, the Kolmogorov-Smirnov (KS) test was used. For
comparison between more than two groups, one-way ANOVA (Kruskal-Wallis, KW test)
was performed. Overall, the primary motor symptom was not a factor for differences in
firing rates, likely due to the fact that most dystonia patients present both motor
symptoms, but only the most prominent symptom was considered.
45

Table 4.3: Summary of the firing rate statistics. Values indicate the group median. KS =
Kolmogorov-Smirnov. KW = Kruskal-Wallis.
MRI
Primary motor
symptom
Diagnosis

Abnormal
(Hz)
Normal
(Hz)
KS
(p<)
Hyperkinesia
(Hz)
Hypertonia
(Hz)
Primary
(Hz)
Secondary
(Hz)
Tourette
(Hz)
KW
(p<)
GPi 2.4 3.4 0.05 2.8 2.7 4.2 2.6 1 0.05
Vim 2.6 6 0.01 3.3 2.8 3.3 3.4 2.3
Vo/
VA
4 4.4 4.1 4.24 3.3 5 2.2 0.05

Figure 4.4 shows the GPi firing rate distributions per subject group and condition.
Subjects with abnormal MRI showed significantly lower firing rates than the patients with
no signs of brain injury; subjects with secondary dystonia showed significantly lower firing
rates compared to subjects with primary dystonia diagnosis, and no significant difference
was found between patients with differences in primary motor symptoms.
Figure 4.5 shows the Vim firing rate distributions per subject group and condition.
Only subjects with differences in MRI showed significant differences in firing rates (p <
0.01). Similarly to GPi, subjects with abnormal MRI had lower firing rates than subjects
with normal scans. Vo and VA are both output projection areas from GPi, therefore, for
this analysis they were combined together. Figure 4.6 shows only significant difference
between subjects with different diagnosis (p < 0.01). The secondary dystonia group
showed higher firing rates than the subject with Tourrete’s syndrome.
46

Figure 4.4: GPi firing rate distributions per condition. Two medians are significantly different at
the 5% significance level if their intervals (red triangles). * p<0.05, ** p<0.01.

Figure 4.5: Vim firing rate distribution per condition (** p<0.01)
47

Figure 4.6: Vo/VA firing rate distribution per condition (** p<0.01)
Firing rates in GPi showed positive correlation with thalamic firing rates and muscle
activation
The principal advantage of our inpatient microelectrode recording setup over
traditional intraoperative recordings is the possibility to study neural patterns of activation
during the execution of voluntary movement. Dystonia is a movement disorder that is
triggered by voluntary attempts to move. According to the most common model of basal
ganglia function (the ‘rate model’), decreased firing rates in GPi output would cause
uncontrolled muscle contractions and ‘excessive movement’ due to increased thalamic
disinhibition (Jerrold L. Vitek 2002). In this study, it was found that during the execution
of voluntary movement, both, GPi and thalamic nuclei, increase spiking activity. Figure
4.7 shows the spike raster of well isolated single units (SNR > 1) and the corresponding
EMG envelope for one recording session of about one-hour duration in which subject 1
was asked to perform a series of unconstrained reaching tasks. Periods of active
48

Figure 4.7: Spike raster (top) and EMG envelope (bottom) of about 1 hour and 9 minutes of
recordings from subject 1. Note the high spike activity during the periods of active EMG, as well
as the widespread positive correlation between GPi and thalamic nuclei.
49

movement and rest can be clearly observed from changes in EMG amplitude. From the
figure, it can be observed that periods of high EMG activity correspond to a widespread
activation of the GPi and thalamic nuclei. Importantly, spiking activity dropped significantly
during low EMG amplitude segments that correspond to the periods at which the subject
was at complete rest. Figure 4.8 shows the percentage of neurons with significant change
in firing rate during the execution of voluntary movement compared to the periods of
complete rest (KS test between active and at rest firing rate distributions, p<0.05). It can
be observed that the vast majority of neurons in all implanted nuclei increased firing rates
during movement.

Figure 4.8: Comparison between median firing rates at rest and during voluntary movement.
Red areas indicate that the median firing rate increased with movement, and blue areas indicate
that the median firing rates decreased with movement.

50

To analyze the correlation between firing rates in multiple brain areas using
standard linear methods, the spike trains were convolved with an exponential decay.
Figure 4.9 shows an example of correlated firing patterns between GPi and Vo.

Figure 4.9: Correlation coefficient matrix between firing rates across brain areas for subject 2.
GPi1 and GPi2 correspond to two different electrodes implanted in the anterior and posterior
portions the GPi.

Discussion
Studies of human basal ganglia and thalamus have presented a significant
challenge because they are often limited to intraoperative recordings and because the
weak correlation between subcortical neural patterns and measurable features of motor
output such as EMG and limb kinematics (DeLong 1972; J. L. Vitek et al. 1999; J. W.
Mink and Thach 1991; Schmidt and Berke 2017). Here, it was shown that chronic
multielectrode recordings can help to overcome some of these limitations.
Identification of single units from low impedance microelectrodes was corroborated
by comparing firing rates to reported ones in a similar subject population using standard
51

microelectrode intraoperative recordings. In a study of 22 subjects with secondary
dystonia (McClelland et al. 2016), the median firing rate of GPi neurons was reported to
be 9.6 Hz in a distribution with a range of 1 - 66 Hz. Here, the median GPi firing rate was
found to be much lower (2.8 Hz) but the range of the distribution was very similar (0.1 -
62 Hz). Moreover, consistent with the same report, subjects with evidence of injury
(abnormal MRI) showed decreased firing rates. This indicates that the presence of slow
firing neurons does not appear to be a result of our specific recording method, instead,
they appear to be a feature of secondary dystonia.
Unbiased sample of basal ganglia and thalamus
The high number of low firing rate neurons found in this study rarely has been
reported, likely due to the short recording times in the intraoperative setting. It is also
possible that these results reflect the fact that no active cells were used as a reference to
determine the location of the recording electrode at the time of implantation, in which case
these results show an unbiased sample of the basal ganglia and thalamic activity.
Traditional single unit recordings have been performed by advancing the microelectrodes
until neural activity is detected. This method can lead to a form of selection bias where
not only will recordings be made primarily from high rate neurons, but also from cells
responding to the specific stimuli or conditions present at the time of the recording (Harris
et al. 2016). In this work, sampling bias was ameliorated by performing continuous
recordings using fixed electrodes over very long time periods, which lead to sufficient
spike numbers to define clusters even for cells with predominantly low firing rate.
52

Selective inhibition/disinhibition model for the function of basal ganglia
Despite discrepancies in the firing rates, the important observation from these
simultaneous recordings is the widespread activation of both GPi and thalamus, and the
observation that both areas increase the firing rate during movement. This is not predicted
by the rate model, in which the inhibitory output of GPi would be expected to correlate
inversely with motor thalamus. The output projections from basal ganglia to thalamus are
entirely inhibitory, originating in the GPi and projecting to secondary motor thalamic nuclei
including Voa/Vop and VA. Although it is possible that these inhibitory signals can
themselves propagate due to rebound depolarization (Zhuang, Li, and Hallett 2004; Rovó,
Ulbert, and Acsády 2012), it may be more likely that cortical drive to thalamic neurons
causes depolarization that is subsequently inhibited by the incoming signals from GPi.
Thus, thalamic projections back to cortex create a cortico-thalamo-cortical loop that could
be modulated by the inhibitory inputs from GPi. In a pilot study performed in subject 1
during the inpatient multielectrode recording procedure, the subject was asked to perform
a rhythmic task that consisted of drawing a figure 8 on an iPad. The task was paced by a
metronome, so that the desired task-relevant frequency was known. As a result of this
task constraint, undesired task-irrelevant frequencies represent potential movements that
must be suppressed in order to execute the task correctly (Lunardini et al. 2015). From
the frequency spectrum of the local field potential recordings (LFP), the presence of task-
relevant and task-irrelevant components can be determined in each muscle and each
region of the brain. Figure 4.10 shows the power spectral density of the EMG and
kinematics of the task, where arm muscles coordinate to produce the task-relevant
dynamics necessary to perform the rhythmic task with a pace set to 30 bpm (0.5 Hz).
53

Figure 4.11 shows the LFP power spectral density of the contralateral GPi and Vim. GPi
shows two peaks above and below the task frequency, while Vim peaks at the exact task-
relevant frequency, thus, indicating GPi inhibition of the task-irrelevant frequencies and
focusing motor performance on task-relevant dynamics.

Figure 4.10: Power spectral density of EMG (black) and x (blue) and y (red) components of the
task. The EMG PSD was computed from the envelopes of the anterior deltoid (AntDelt), biceps
(Bi), lateral deltoid (LatDelt), posterior deltoid (PostDelt), supraspinatus, triceps (Tri), and the
wrist extensor (WE) and flexor (WE) groups.

Prior work has shown that GPi in healthy primates has tonic firing, suggesting that
inhibition or disinhibition of thalamo-cortical loops could be a mechanism for basal ganglia
to control motor function (Miller and DeLong 1987; Raz et al. 2001; Saxena et al. 2019).
Moreover, surround inhibition has been proposed as a feature of basal ganglia function
(Jonathan W. Mink 1996). Evidence of this center-surround organization in the motor
system has been observed in finger muscles during motor activation, where active
muscles show increased excitability while neighboring muscles are inhibited (Sohn and
54

Hallett 2004). In focal hand dystonia, patients have shown deficient surround inhibition
during movement initiation (Beck et al. 2008).

Figure 4.11: Comparison of the power spectral density of the LFP recording in GPi (left column)
and Vim (right column) during the performance of the rhythmic task at 0.5 Hz (black trace), and
during resting state (red trace 土 standard deviation). Note the two peaks in GPi above and
below the task-relevant frequency, and the peak in Vim at the task-relevant frequency.

While the rate model of basal ganglia output would suggest that dystonia is
associated with widespread disinhibition of thalamus, the spike and LFP data presented
here suggests that the pathophysiology will be associated with abnormal patterns of
disinhibition or inhibition. While no data from healthy human subjects are available for
comparison, this is consistent with a model in which a process that would normally reduce
the thalamic activity is functioning but perhaps incompletely in children with secondary
dystonia. Based on these observations, we propose a new model in which the basal
ganglia can select or inhibit different dynamics by control of the inhibitory output from GPi.
In particular, we propose that the cortico-thalamo-cortical loops modulate cortical
55

dynamics by driving changes in cortical activity based on past activity in both sensory and
motor areas. This provides a mechanism for the influence of sensory information on
movement and provides a mechanism for time-varying control of motor outputs. Failure
to properly map the sensory input into the correct motor output would result in
superimposed dynamics that might include involuntary postures, or movements that
respond inappropriately to sensory inputs and thus appear random. Hypertonia and
hyperkinetic involuntary movements are the hallmarks of dystonia, so this provides a
hypothesis of the link between basal ganglia injury and the different features of different
types of dystonia.

56

5 Thalamic evoked potentials in response to GPi-DBS show
frequency dependent modulation
Evoked potentials (EPs) are a fundamental electrophysiology tool to analyze the
neural response to electrical stimulation. When a stimulation pulse is delivered, the
recording EP signals are a combination of the neural response and the stimulation artifact,
and if the EP occurs at short latencies, it can be occluded by the artifact. In order to study
the stimulation effects on the neural tissue, it is essential to differentiate the EP from
artifactual waveforms. This chapter presents the use of a polarity reversal technique to
reduce artifact contamination in EP recordings. Moreover, this technique is applied to
intracraneal microelectrode and electroencephalography (EEG) recordings to study the
propagation of the DBS pulse through the brain areas involved in the motor control, with
the goal of providing insights into the mechanism by which DBS can effectively reduce
symptoms of movement disorders. Thalamic and cortical EPs in response to low
frequency stimulation of the GPi (GPi-DBS) are analyzed, then the GPi-DBS frequency is
varied, and the thalamic EPs are measured. We found that low frequency GPi-DBS
elicited an evoked response in its projecting thalamic nuclei, but no cortical response
could be observed. Thalamic EPs in response to GPi-DBS had a peak amplitude within
2 ms after stimulation onset, which decreased as the stimulation frequency increased,
showing the frequency dependent modulatory effects of DBS. These findings have not
been previously reported, probably because most studies of intracraneal EPs record
neural signals from the implanted DBS macroelectrodes, which large surface area
average a large volume of neural activity therefore losing spatial and temporal resolution.
Here, the relevance of low impedance microelectrode recordings in the study of the
57

electrophysiology of DBS is shown and the methods for processing these signals are
established.
Introduction
Deep brain stimulation of the internal segment of the globus pallidus (GPi-DBS),
has been an effective treatment to alleviate dystonic symptoms of patients for whom
medication has failed to treat effectively. Currently, DBS settings are programmed by a
clinician whose goal is to maximize symptom suppression while minimizing negative side
effects. Although effective, changes in motor symptoms are not immediate, the level of
long-term improvement varies from subject to subject, and the specific settings are likely
to be subject dependent based on the specific etiology or level of injury. Recently, closed-
loop DBS has been proposed as an alternative to automatically adjust DBS settings
based on some biomarker specific to the movement disorder (Johnson et al. 2016).
Regardless the programming method, it is evident that in order to improve DBS therapy,
it is necessary to understand the effects of electrical stimulation in the neural tissue.
Several hypotheses have been proposed to explain the physiological mechanisms by
which DBS is effective. The similarity in outcomes between DBS and pallidotomy has led
to the proposition that DBS inactivates or inhibits the structures being stimulated. Studies
have suggested that DBS inhibits the cell bodies of neurons surrounding the electrode,
by activation of presynaptic terminals, while simultaneously stimulating the output of local
neurons by initiation of action potentials in the axon distant from the soma (W. Grill 2001;
McIntyre et al. 2004). Additionally, it has been shown that the level of improvement in the
motor symptoms is frequency dependent. Typical stimulation frequencies to treat dystonia
58

vary from 60 to 185 Hz , moreover, there has been reports of low frequency DBS that can
accentuate motor symptoms (Benabid et al. 1991).
DBS evoked potentials (DBS EP) are an electrophysiology tool commonly used to
study the neural response to electrical stimulation. Our recent development on
intracraneal multielectrode recordings has opened a window to study the effects of DBS
parameters on the propagation of the DBS EP through the motor loop. In this recording
setup, depth electrodes are implanted in GPi and its projected thalamic nuclei Voa/Vop
(Vo for conciseness). If GPi-DBS causes orthodromic propagation of the stimulation
pulse, an EP is expected to appear in Vo a few milliseconds after stimulation onset. When
the EP occurs at short latencies relative to the stimulation onset and the stimulating and
recording electrodes are close to each other, the EP can be occluded by the stimulation
artifact. A typical solution consists of ‘blanking’ the front-end amplifiers of the recording
system for the same time period as the stimulation pulse duration, this way, the
stimulation artifact does not saturate the amplifiers and the EP can be measured
(Montgomery 2006). However, this method requires precise synchronization between the
time of the stimulation pulse and the recording system, which is only achieved when both
the stimulator and amplifier share the same clock (i.e. they are in the same system). In
practice, stimulation pulses are usually delivered through the implanted DBS pulse
generator, and the EP is collected using a separate recording system. In spite of the
inconvenience that this represents when analyzing short latency DBS EP, it also provides
a signal which pulse shape characteristics are consistent with the clinically effective
pulses.

59

To overcome these limitations in recording and analyzing asynchronous short
latency DBS EP, a method to attenuate the DBS artifact is presented. The method
consists of inverting the simulation polarity between two stimulating contacts, which
inverts the polarity of the artifact while preserving the polarity of the neural response.
Then, several interpulse epochs are averaged together, effectively attenuating the artifact
while increasing the signal-to-noise ratio (SNR) of the neural response. The
implementation of this method is demonstrated on the analysis of Vo EP in response to
GPi-DBS for one subject during our inpatient multielectrode recording procedure.
Microelectrode recordings are collected for several stimulation frequencies, from single
pulse (9 Hz) to clinically effective frequencies (>100 Hz). Additionally,
electroencephalography evoked potentials (EEG EP) are collected postoperatively and
analyzed using the same polarity reversal method. Finally, a model for the propagation of
the DBS pulse is discussed.
Methods
Subjects: inpatient intracranial recordings
A pediatric subject with unilateral idiopathic dystonia underwent previous DBS
implantation of the GPi, Vo, and VPLa nuclei contralateral to the most affected left side
of the body. Based on reports that DBS is more effective when stimulation is delivered
biliterary to both hemispheres, the subject was given the option of a second DBS
implantation of the hemisphere contralateral to the least affected side of the body. After
acceptance, the subject was admitted for the inpatient neuromodulation procedure
(described in chapter 3). Multiple temporal electrodes were implanted in potential target
60

areas including the left GPi, left Voa/Vop (Vo, for conciseness), left Vim, and left VPL
(Hirai and Jones 1989). Voa/Vop are the output projection areas from GPi, Vim is the
projection area from cerebellar nuclei, and VPLa is the projection area from the
spinothalamic and lemniscal sensory systems. Surgical implantation as well as all
recording procedures were performed at the Children’s Hospital Los Angeles. The subject
signed informed consent for surgical procedures in accordance with standard hospital
practice. The subject and parents of the minor patients also sign informed consent for
research use of electrophysiological data (University of Southern California Human
Subjects Institutional Review Board approval UP-13-00521, 15 November 2013 to 12
September 2018) and HIPAA authorization for research use of protected health
information. Table 5.1 shows the demographics, recording targets, and final implantation
targets that were selected based on the evaluation of electrical stimulation outcomes and
electrophysiological recordings.
Table 5.1: Demographics, temporary recording lead locations, and final selected targets.
ID Age Diagnosis Recording targets Implantation targets
1 15
Unilateral idiopathic
dystonia
GPi, VoaVop, Vim, VPLa GPi+VoaVop+VPLa
Subjects: outpatient EEG recordings
An outpatient follow-up EEG study was conducted on 24 pediatric subjects that
underwent DBS implantation either through the inpatient neuromodulation procedure or
by the standard intraoperative stereotactic placement of the permanent DBS leads.
Subject details are shown in table 5.2. Subject 1 corresponds to the same subject as in
table 5.1. EEG recordings were collected between 1 and 3 months after activation of the
61

DBS pulse generator. The University of Southern California Institutional Review Board
approved the study protocol. All children’s parents gave informed written consent for
participation, and all children gave written assent. Authorization for analysis, storage, and
publication of protected health information was obtained from parents according to the
Health Information Portability and Accountability Act (HIPAA). The experiment was in
accordance with the Declaration of Helsinki.
Data collection
Data were collected while subjects were at complete rest. EPs were recorded from
the microelectrode signals sampled at 24 kHz and stored for offline analysis. All
microelectrodes were referenced to the macroelectrode most distal to the tip of the lead.
The preferred reference location was GPi, however, the final reference was selected
based on the most distal macroelectrode that reduced the overall background noise level
in the recording microchannels. A scalp EEG electrode attached to the forehead was used
as the system ground. Bipolar stimulation was applied systematically to each adjacent
macroelectrode pair in the implanted GPi lead using a Medtronic external neurostimulator
model 37022 connected to a N’Vision clinician programmer model 8840 (Medtronic Inc.,
Minneapolis, MN, USA). The stimulation polarity was interleaved between stimulation
blocks of 1200 pulses minimum, i.e., one block of cathodic phase first (e.g. 1-2+), followed
by a block of anodic phase first (e.g. 1+2-). Pulse amplitude and pulse width were fixed
to 3 V and 120 us, respectively, and stimulation frequencies were varied from 9, 29, 59,
89, 189 to 249 Hz. These frequencies were selected based on typical DBS clinical settings
and in order to avoid overlap with 60 Hz harmonics.
62

Figure 5.1: Schematic of the depth electrode and stimulation polarities. Artifact reduction was
achieved by reversing the stimulation polarity between stimulation blocks of ~1200 interpulse
epochs. Big black squares represent the macroelectrodes from 1 to 6, and small circles in
between represent the microelectrodes.

Postoperative EEG recordings were collected from all subjects while at rest using
the ASAlab™ system (ANT Neuro, Enschede, The Netherlands). An electrode cap (ANT
WaveGuard™) containing 32 Ag/AgCl electrodes was used and fitted with a chin-strap.
The electrodes were positioned according to the 10-20 System and signals were sampled
at 2048 Hz. Similarly, to the inpatient microelectrode recordings, the stimulation polarity
was interleaved between stimulation blocks of 1200 pulses minimum. In this case, only 9
Hz stimulation was applied to the permanent implanted leads to analyze DBS pulse
propagation to cortical areas. Pulse amplitude and pulse width were fixed to 3 V and 120
us, respectively.
Data analysis
A high-pass digital filter (200 Hz, 4th order, Butterworth) was applied to the
microelectrode intracranial recordings. A differential montage between adjacent
microelectrodes was implemented offline to reduce noise interference and to reduce the
spatial spread of the recorded field potentials. Stimulation artifacts were detected by
voltage level crossing using a threshold of 10 times the standard deviation. The interpulse
63

Table 5.2: Demographics and implanted DBS targets.
ID Age (years) Diagnosis Implanted targets
1 15 CP GPi+Vo+VPL
2 20 CP GPi+Vo
3 20 CP GPi+Vo
4 9 CP GPi+Vim
5 18 Primary dystonia GPi
6 11 CP GPi
7 18 CP GPi
8 7 Primary dystonia (Unknown) GPi+Vo
9 16 Angelman Syndrome GPi+Vo
10 11 Dopamine transporter GPi+Vo
11 7 Idiopathic dystonia GPi+Vim
12 18 CP GPi+STN
13 6 Hemolytic Uremic Syndrome GPi+Vo
14 15 Idiopathic dystonia GPi+Vim
15 15 CP GPi+Vim
16 18 Stroke GPi+Vo
17 17 CP GPi+STN
18 12 Primary dystonia (Unknown) GPi+Vim
19 11 ADCY5 GPi+Vim
20 21 Congenital Lymphangioma, CP GPi+STN
21 18 CP GPi+Vo
22 10 Kernicterus, G6PD deficiency GPi+Vim

64

epochs were upsampled by a factor of 100 using cubic spline interpolation, aligned by
absolute maximum peak, and downsampled back to 24 kHz using linear interpolation. At
least 1200 epochs were averaged to obtain the cathodic and anodic average signals. The
EP signal was obtained by averaging the two signals after fitting the anodic to the cathodic
using standard linear fitting. EP amplitude was measured from all thalamic nuclei as the
maximum peak amplitude 0.5 ms after the onset of the stimulation artifact.
A digital high-pass filtered (20 Hz, 4th order, Butterworth) was applied to the EEG
recordings. Interference from the hair and the passive scalp electrodes made the EEG
recordings more susceptible to noise, therefore, in order to improve artifact detection, the
EEG signals were double-differentiated so that the maximum rate of change could be
used to detect DBS artifacts by voltage level crossing with a threshold of 10 times the
standard deviation. A minimum of 1200 interpulse epochs were extracted from the original
time series. The interpulse epochs were upsampled by a factor of 10 using cubic spline
interpolation, aligned by absolute maximum peak, and downsampled back to 2.048 kHz
using linear interpolation. Similarly to the intracranial recordings, cortical EPs were
obtained by computing the average of the anodic and cathodic signals after linear fitting
of the anodic to the cathodic signal and their amplitudes were measured as the peak-to-
peak distance of the maximum and minimum peaks that were significantly different from
zero in a time window from 3 to 45 ms after stimulation onset. Significance was calculated
from the sample distribution of the EP signals at each time point for a 5% confidence.
65

Results
Artifact reduction by reversed polarity stimulation
Consider the stimulation artifact to be the dipole
𝛷 =
1
4𝜋𝜎
⋅
𝐼 ⋅ 𝑑 𝑐𝑜𝑠 (ө)
𝑟 2

where the electric field 𝛷 is a function of the conductivity 𝜎 of the tissue, the current source
amplitude I, the distance r between the source and the recording electrode, the distance
d separating the dipole sources, and the polar angle θ which is measured from the
recording site relative to the vector formed by the dipole sources. When the stimulation
polarity is inverted, the direction of the field is determined by θ. If every other parameter
is kept the same, θ = θ0+180
o
, causing the recorded electric field potential to invert. The
extracellular potentials can also be considered dipoles generated by the sum of
contributions from sources and current sinks of the individual neurons in the volume
conductor. Changes in the stimulation parameters will change the response of the neural
tissue (W. M. Grill 1999), however, neural depolarization is nonlinear with respect to
changes in stimulation polarity, i.e. neural cells either fire or not in response to changes
in the extracellular electric field. This polarity reversal technique takes advantage of this
nonlinearity of the neural response by inverting the polarity of the stimulation artifact while
preserving the polarity of the EP, therefore reducing artifact contamination, especially for
those EPs with short latencies. This method was effective at attenuating the capacitive
effect (artifact tail) at latacies in which the response occurred (Figure 5.2).
66

Figure 5.2: Example of the artifact reduction method. Left: Upsampled cathodic, anodic, and
average EP signals. The blue trace represents the average of ~1200 stimulus applied to
macroelectrode 1 negative (-, cathode) and macroelectrode 2 positive (+, anode). The red trace
shows the response to the same number of stimuli but with inverted stimulation polarity, i.e.
1+2-. The black trace is the average response of both. Right: Vo EP (black trace) which occurs
within 2 ms of GPi-DBS.

Although the polarity of extracellular potential may not be affected by the
stimulation angle, the relative position of the neurons with respect to the stimulating
electrode affects the way neurons are depolarized (McIntyre et al. 2004). Note from the
right plot in Figure 5.2 the difference in EP shape between the cathodic and anodic
phases of stimulation. This is explained by the change in the relative position of the
electrode’s cathode with respect to the recorded volume of neurons when the polarity is
reversed. Despite these differences between the neural response to cathodic and anodic
stimulation, this artifact reduction method allows to clearly differentiate the artifact
waveform from the EP response at latencies as short as 1 ms.
Low frequency GPi-DBS elicited Vo EPs
The polarity reversal technique was used to study the propagation of the DBS
pulse through the neural tissue. Figure 5.3 shows the Vo EP response (Vo EP) elicited
67

by low frequency (9Hz) GPi-DBS, with peak responses within 2 ms and repolarization
period of ~5 ms.

Figure 5.3: Vo EP in response to 9 Hz GPi-DBS. Red and blue traces represent the two
stimulation polarities. The black traces are the average response. Each plot corresponds to a
different depth along the recording lead, being the top left plot the most distal microelectrode
pair and the bottom right plot the most proximal relative to the tip of the lead.

EPs were not often present in the cerebellar and sensory thalamic nuclei
projections Vim and VPL (figure 5.4), consistent with the expected anatomical
connections between the implanted areas given that the majority of the GPi outputs
project to motor thalamic nuclei Vo and VA. However, long latency EPs at around 2 ms
could be observed in some electrodes of the VPLa lead. This could represent spread
thalamic depolarization by secondary interneurons with projections to sensory areas.
68

Figure 5.4: Vim EP (left) and VPLa (right) responses to 9 Hz GPi-DBS.

Vo EPs were larger for the electrodes closer to the target area, in which
macroelectrode 1 was implanted. From figure 5.5 it can be observed that Vo EP
amplitudes were larger for the electrode group closer to the target area and decreased
as the distance of the microelectrode relative to this area increased. Moreover, the latency
of the peak response increases with the distance from the target. It can also be observed
from the figure a second peak at around 5 ms, which may correspond to the refractory
period of the depolarized volume of neurons in Vo in response to the activation of a large
volume of axons in GPi due to the stimulation pulse (McIntyre et al. 2004).
69

Figure 5.5: Propagation of the Vo EP along the recording lead. The heatmap shows the average
Vo EP signal per microelectrode group indicated in the left schematic of the recording electrode.
Note the time delay of the EP response as it distances from the stimulation target.

Low frequency GPi-DBS does not elicit a cortical response
The postoperative follow up EEG study was conducted using the same polarity
reversal technique as described above through the macroelectrode pairs in the
permanent DBS leads. The results show that GPi-DBS did not elicit a consistent cortical
evoked response. Figure 5.6 shows the average EEG of all 24 subjects. Mean cortical
EP amplitudes lower than 1 uV. In addition to GPi, patients were implanted in a second
nucleus for which clinical stimulation showed the most benefits with the least amount of
negative side effects. Common targets were Vo and Vim. EEG EPs were also collected
for the low frequency 9 Hz stimulation applied to the thalamic nuclei. The results in figure
5.7 show that thalamic DBS produces a high amplitude EP in central and frontal EEG
channels (primary and supplementary motor cortex, respectively), ipsilateral to the
stimulation side, with consistent peak times at around 20 ms and mean peak amplitudes
of around 7 uV.
70

Figure 5.6: Aggregate cortical EP in responses to GPi-DBS. Gray areas indicate the standard
error of the mean, black traces are the average EP signal, and red dots indicate when the signal
was significantly different from zero (above and below).

Figure 5.7: Aggregate cortical EP in responses to Vo-DBS.
71

In patients with consistent EPs, the peak-to-peak amplitude and cortical
projections changed depending on the simulation contact pair (FIgure 5.8). Contacts that
produced a higher EP corresponded to the active clinical contacts.

Figure 5.8: Topographical heatmap of the cortical EP in response to (from top to bottom) Vo-
DBS, VPL-DBS, and GPi-DBS. Each column represents the response to different stimulating
electrode pair.
72

Vo response to GPi-DBS was frequency dependent
The Vo EP was analyzed for several GPi stimulation frequencies. Figure 5.9 shows
that the amplitude of the Vo EP decreases as the stimulation frequency increases. Figure
5.10 shows that maximum attenuation occurs when the period between stimulation pulses
is less than the observed repolarization period of ~5 ms (>189 Hz).

Figure 5.9: Vo EP in response to GPi-DBS frequency swab from distal (top row) to proximal
(bottom row) microelectrodes in reference to the tip of the lead. Left: average Vo EP traces.
Red dots indicate the peak of the EP. Right: EP amplitudes corresponding to each GPi
stimulation frequency. Note that the EP amplitude decreases as the stimulation frequency
increases.
Discussion
A reverse polarity method to analyze intracraneal and EEG EPs was presented.
We found that low frequency GPi-DBS elicits an evoked response in its projecting
73

thalamic nuclei Vo but no cortical response was observed. The topography of the Vo EP
shows a peak amplitude within 2 ms after stimulation onset, and a depolarization period
of ~3 ms followed by a repolarization peak at ~5 ms.

Figure 5.10: EP amplitude is attenuated as the stimulation period is closer to the repolarization
period of the downstream nuclei.

Attenuated propagation of DBS depolarization may indicate similarities with ablation
Previous observations of cortical EPs elicited by GPi stimulation have suggested
cortical modulation as a DBS mechanism to treat dystonia (Bhanpuri et al. 2014; Tisch et
al. 2008), probably due to rebound excitation of the thalamic neurons (Kim and Kim 2018).
Our analysis indicates that low frequency GPi-DBS can elicit a thalamic response, but the
EP does not propagate to cortical areas. These discrepancies may be due to the fact that
previous studies did not implement a method to separate EPs from artifactual waveforms,
therefore measuring artifactual responses that are similar in shape to the EP waveform.
Nonetheless, the observation that the DBS signal does not propagate to primary output
74

areas is intriguing. The Vo EP shows a first peak within 2 ms and a repolarization peak
around 5 ms after stimulation onset, probable indication of rebound excitation followed by
the depolarization of GPi efferent axons in response to the DBS pulse. While it is possible
that the Vo EP ‘dilutes’ through the multiple thalamocortical projections, and that the
spatial average of the mix of active and inactive projections in the network shows no EP
in the EEG recordings, this also means that DBS can be used as neuromodulation for
deep structures without generating significant sensory or motor side effects.
Vo EPs were attenuated with high frequency stimulation of the GPi, likely due to
persistent depolarization of the efferent projections from the stimulation site, suggesting
a form of blockade of the receiving thalamic nuclei. These observations are consistent
with a model in which blocking downstream propagation of DBS may be one of the ways
in which DBS can function as a lesion, substituting patterns of variation of neural
activation with meaningful information, for regular neural patterns with no variation and
therefore irrelevant information (Warren M. Grill, Snyder, and Miocinovic 2004; Sanger
2018).

75

6 Overall Conclusions
Depth electrode recordings for clinical evaluation of movement disorders represent
a unique opportunity to understand the human neural activity and brain function at
multiple levels. This work provides a framework for the analysis of chronic multielectrode
recordings for deep brain areas in awake patients for whom neural activity can be
associated with motor behavior. We presented a new spike detection algorithm that is
suited for low SNR recordings such as the ones collected with current available depth
multielectrodes for human implantation. This led to new physiological observations of
basal ganglia and thalamic activity that would not be possible to be observed during
standard intraoperative recordings. Furthermore, based on these observations, new
models for basal ganglia and DBS function were discussed. It is important to keep in mind
that these studies were performed in patients with neurological disorders, and that the
results presented here may be specific to dystonia. However, the homogeneity in the
results and the heterogeneous causes of dystonia of the subjects studied here, may
indicate that many of these observations might be part of a normal brain process.
The implications of this work in the improvement of DBS treatment for dystonia are
several. First, we established the signal processing tools for the analysis of depth
multielectrode recordings necessary to understand the pathophysiology of dystonia. The
current trend in multielectrode design is aiming to high-density arrays of blunt
microelectrodes which will require fast and proper signal processing methodologies
which, as discussed earlier in this work, differ from the traditional spike detection
techniques that have been developed over the years for a very specific type of electrodes.
76

For the past few years, several studies have been focused on the feasibility of decoding
movement and percepts from multiple levels of neural activity, from the single unit to multi-
unit and LFPs. Recently, closed-loop DBS have used online stereo EEG recordings to
detect biomarkers in the LFP (typically in the form of oscillations) to deliver targeted
stimulation to treat the motor symptoms of Parkinson’s disease and to stop epileptic
seizures. While epileptic events are characterized by sustained LFP oscillations, the
specific origin of the motor symptoms in movement disorders is unknown and identifying
its genesis may require detecting lower levels of neural activation such as asynchronous
events of spike coincidences that would not show at the LFP level. In this scenario, higher
specificity in the neural recordings together with efficient online algorithms to detect lower
levels of neural activity from multiple signals would be necessary for identification of the
correct biomarker to potentially improve DBS treatment.
Second, this work provided physiological evidence for a new hypothesis of basal
ganglia function that can link basal ganglia injury and the different features of different
types of dystonia. These observations provide a framework for new hypothesis of the role
of basal ganglia in motor control and expands the views on the commonly accepted but
highly constrained rate model. Finally, physiological evidence of DBS frequency
modulation of the neural circuit is provided, together with clinical observations of patients
with similar symptoms responding differently to the same stimulation frequency. This
frequency response suggests the possibility of an optimal stimulation frequency per
subject and the possibility of the use of DBS evoked potentials as biomarkers to guide
DBS programming. Furthermore, DBS EPs could be used in closed loop for personalized
treatment of dystonia and other movement disorders.
77

The role of brain-derived signals from depth multielectrodes is important not only
to understand the physiological mechanisms governing the control of movement, but also
in designing neural interfaces and improving clinically relevant devices that can have a
potential direct positive impact on the life of people suffering from neurological diseases
for whom current treatments are ineffective, with the goal of improving their quality of life.

78

Bibliography
Albin, Roger L., Anne B. Young, and John B. Penney. 1989. “The Functional Anatomy of
Basal Ganglia Disorders.” Trends in Neurosciences 12 (10): 366–75.
Beck, S., S. P. Richardson, E. A. Shamim, N. Dang, M. Schubert, and M. Hallett. 2008.
“Short Intracortical and Surround Inhibition Are Selectively Reduced during
Movement Initiation in Focal Hand Dystonia.” Journal of Neuroscience 28 (41):
10363–69.
Bell, Anthony J., and Terrence J. Sejnowski. 1997. “The ‘independent Components’ of
Natural Scenes Are Edge Filters.” Vision Research. https://doi.org/10.1016/s0042-
6989(97)00121-1.
Benabid, A. L., P. Pollak, C. Gervason, D. Hoffmann, D. M. Gao, M. Hommel, J. E. Perret,
and J. de Rougemont. 1991. “Long-Term Suppression of Tremor by Chronic
Stimulation of the Ventral Intermediate Thalamic Nucleus.” The Lancet 337 (8738):
403–6.
Bertucco, Matteo, and Terence D. Sanger. 2015. “Current and Emerging Strategies for
Treatment of Childhood Dystonia.” Journal of Hand Therapy: Official Journal of the
American Society of Hand Therapists 28 (2): 185–93; quiz 194.
Bhanpuri, Nasir H., Matteo Bertucco, Diana Ferman, Scott J. Young, Mark A. Liker, Mark
D. Krieger, and Terence D. Sanger. 2014. “Deep Brain Stimulation Evoked Potentials
May Relate to Clinical Benefit in Childhood Dystonia.” Brain Stimulation 7 (5): 718–
26.
79

Bhatia, K. P., and C. D. Marsden. 1994. “The Behavioural and Motor Consequences of
Focal Lesions of the Basal Ganglia in Man.” Brain: A Journal of Neurology 117 ( Pt
4) (August): 859–76.
Blatt, M., S. Wiseman, and E. Domany. 1996. “Superparamagnetic Clustering of Data.”
Physical Review Letters 76 (18): 3251–54.
Bouthour, Walid, Pierre Mégevand, John Donoghue, Christian Lüscher, Niels Birbaumer,
and Paul Krack. 2019. “Biomarkers for Closed-Loop Deep Brain Stimulation in
Parkinson Disease and beyond.” Nature Reviews. Neurology 15 (6): 343–52.
Calabrese, Ana, and Liam Paninski. 2011. “Kalman Filter Mixture Model for Spike Sorting
of Non-Stationary Data.” Journal of Neuroscience Methods 196 (1): 159–69.
Camuñas-Mesa, Luis A., and Rodrigo Quian Quiroga. 2013. “A Detailed and Fast Model
of Extracellular Recordings.” Neural Computation 25 (5): 1191–1212.
Chaure, Fernando J., Hernan G. Rey, and Rodrigo Quian Quiroga. 2018. “A Novel and
Fully Automatic Spike-Sorting Implementation with Variable Number of Features.”
Journal of Neurophysiology 120 (4): 1859–71.
DeLong, Mahlon R. 1972. “Activity of Basal Ganglia Neurons during Movement.” Brain
Research. 40, no. 1 (1972): 127–135.
DeLong, Mahlon, 1990. “Primate Models of Movement Disorders of Basal Ganglia
Origin.” Trends in Neurosciences 13 (7): 281–85.

80

Dhawale, Ashesh K., Rajesh Poddar, Steffen Be Wolff, Valentin A. Normand, Evi
Kopelowitz, and Bence P. Ölveczky. 2017. “Automated Long-Term Recording and
Analysis of Neural Activity in Behaving Animals.” eLife 6 (September 8, 2017).
Ekanadham, Chaitanya, Daniel Tranchina, and Eero P. Simoncelli. 2014. “A Unified
Framework and Method for Automatic Neural Spike Identification.” Journal of
Neuroscience Methods 222 (January): 47–55.
Fee, M. S., P. P. Mitra, and D. Kleinfeld. 1996. “Variability of Extracellular Spike
Waveforms of Cortical Neurons.” Journal of Neurophysiology 76 (6): 3823–33.
Fraser, George W., Steven M. Chase, Andrew Whitford, and Andrew B. Schwartz. 2009.
“Control of a Brain-Computer Interface without Spike Sorting.” Journal of Neural
Engineering 6 (5): 055004.
Garonzik, Ira M., Sherwin E. Hua, Shinji Ohara, and Frederick A. Lenz. 2002.
“Intraoperative Microelectrode and Semi-Microelectrode Recording during the
Physiological Localization of the Thalamic Nucleus Ventral Intermediate.” Movement
Disorders: Official Journal of the Movement Disorder Society 17 (S3): S135–44.
Gerstein, G. L., and W. A. Clark. 1964. “Simultaneous Studies of Firing Patterns in
Several Neurons.” Science 143 (3612): 1325–27.
Grill, W. 2001. “Extracellular Excitation of Central Neurons: Implications for the
Mechanisms of Deep Brain Stimulation.” Thalamus & Related Systems 1 (3): 269–
77.
81

Grill, Warren M., Andrea N. Snyder, and Svjetlana Miocinovic. 2004. “Deep Brain
Stimulation Creates an Informational Lesion of the Stimulated Nucleus.” Neuroreport
15 (7): 1137–40.
Grill, W. M. 1999. “Modeling the Effects of Electric Fields on Nerve Fibers: Influence of
Tissue Electrical Properties.” IEEE Transactions on Biomedical Engineering. vol. 46,
no. 8, pp. 918-928, Aug. 1999.
Harris, Kenneth D., Rodrigo Quian Quiroga, Jeremy Freeman, and Spencer L. Smith.
2016. “Improving Data Quality in Neuronal Population Recordings.” Nature
Neuroscience 19 (9): 1165–74.
Hirai, T., and E. G. Jones. 1989. “A New Parcellation of the Human Thalamus on the
Basis of Histochemical Staining.” Brain Research. Brain Research Reviews 14 (1):
1–34.
Humphrey, Donald R., and Edward M. Schmidt. n.d. “Extracellular Single-Unit Recording
Methods.” In Neurophysiological Techniques, II, 1–64.
Johnson, Luke A., Shane D. Nebeck, Abirami Muralidharan, Matthew D. Johnson,
Kenneth B. Baker, and Jerrold L. Vitek. 2016. “Closed-Loop Deep Brain Stimulation
Effects on Parkinsonian Motor Symptoms in a Non-Human Primate: Is Beta
Enough?” Brain Stimulation 9, no. 6 (November 2016): 892–896.
Jonker, Pascal K. C., J. Marc C. van Dijk, Arjen L. J. van Hulzen, Teus van Laar, Michiel
J. Staal, and H. Louis Journée. 2013. “The Added Value of Semimicroelectrode
Recording in Deep Brain Stimulation of the Subthalamic Nucleus for Parkinson
Disease.” Neurosurgical Focus 35 (5): E3.
82

Kaiser, J. F. 1990. “On a Simple Algorithm to Calculate the ‘Energy’ of a Signal.”
International Conference on Acoustics, Speech, and Signal Processing. 1990, pp.
381-384 vol.1.
Kim, Jeongjin, and Daesoo Kim. 2018. “Rebound Excitability Mediates Motor
Abnormalities in Parkinson’s Disease.” BMB Reports 51 (1): 3–4.
Kim, Kyung Hwan, and Sung June Kim. 2003. “A Wavelet-Based Method for Action
Potential Detection from Extracellular Neural Signal Recording with Low Signal-to-
Noise Ratio.” IEEE Transactions on Bio-Medical Engineering 50 (8): 999–1011.
Lee, M. S., and C. D. Marsden. 1994. “Movement Disorders Following Lesions of the
Thalamus or Subthalamic Region.” Movement Disorders 9, no. 5 (September 1,
1994): 493–507.
Ludwig, Kip A., Rachel M. Miriani, Nicholas B. Langhals, Michael D. Joseph, David J.
Anderson, and Daryl R. Kipke. 2009. “Using a Common Average Reference to
Improve Cortical Neuron Recordings from Microelectrode Arrays.” Journal of
Neurophysiology 101 (3): 1679–89.
Lunardini, Francesca, Serena Maggioni, Claudia Casellato, Matteo Bertucco, Alessandra
L. G. Pedrocchi, and Terence D. Sanger. 2015. “Increased Task-Uncorrelated
Muscle Activity in Childhood Dystonia.” Journal of Neuroengineering and
Rehabilitation 12 (June): 52.

83

McClelland, V. M., A. Valentin, H. G. Rey, D. E. Lumsden, M. C. Elze, R. Selway, G.
Alarcon, and J-P Lin. 2016. “Differences in Globus Pallidus Neuronal Firing Rates
and Patterns Relate to Different Disease Biology in Children with Dystonia.” Journal
of Neurology, Neurosurgery, and Psychiatry 87 (9): 958–67.
McIntyre, Cameron C., Warren M. Grill, David L. Sherman, and Nitish V. Thakor. 2004.
“Cellular Effects of Deep Brain Stimulation: Model-Based Analysis of Activation and
Inhibition.” Journal of Neurophysiology 91 (4): 1457–69.
Miller, William C., and Mahlon R. DeLong. 1987. “Altered Tonic Activity of Neurons in the
Globus Pallidus and Subthalamic Nucleus in the Primate MPTP Model of
Parkinsonism.” Advances in Behavioral Biology 32.
Mink, Jonathan W. 1996. “The Basal Ganglia: Focused Selection and Inhibition of
Competing Motor Programs.” Progress in Neurobiology 50 (4): 381–425.
Mink, J. W., and W. T. Thach. 1991. “Basal Ganglia Motor Control. I. Nonexclusive
Relation of Pallidal Discharge to Five Movement Modes.” Journal of Neurophysiology
65 (2): 273–300.
Montgomery, Erwin B., Jr. 2006. “Effects of GPi Stimulation on Human Thalamic Neuronal
Activity.” Clinical Neurophysiology: Official Journal of the International Federation of
Clinical Neurophysiology 117 (12): 2691–2702.
Mukhopadhyay, S., and G. C. Ray. 1998. “A New Interpretation of Nonlinear Energy
Operator and Its Efficacy in Spike Detection.” IEEE Transactions on Bio-Medical
Engineering 45 (2): 180–87.
84

Nambu, Atsushi. 2007. “Globus Pallidus Internal Segment.” In Gaba and the Basal
Ganglia - From Molecules to Systems, 160:135–50. Progress in Brain Research.
Elsevier.
Nelson, Alexandra B., and Anatol C. Kreitzer. 2014. “Reassessing Models of Basal
Ganglia Function and Dysfunction.” Annual Review of Neuroscience 37: 117–35.
Obeid, Iyad, and Patrick D. Wolf. 2004. “Evaluation of Spike-Detection Algorithms for a
Brain-Machine Interface Application.” IEEE Transactions on Bio-Medical Engineering
51 (6): 905–11.
Ondo, W. G., J. M. Desaloms, J. Jankovic, and R. G. Grossman. 1998. “Pallidotomy for
Generalized Dystonia.” Movement Disorders: Official Journal of the Movement
Disorder Society 13 (4): 693–98.
Pazienti, Antonio, and Sonja Grün. 2006. “Robustness of the Significance of Spike
Synchrony with Respect to Sorting Errors.” Journal of Computational Neuroscience
21 (3): 329–42.
Quian Quiroga, Rodrigo, and Stefano Panzeri. 2009. “Extracting Information from
Neuronal Populations: Information Theory and Decoding Approaches.” Nature
Reviews. Neuroscience 10 (3): 173–85.
Quiroga, Rodrigo Quian. 2012. “Spike Sorting.” Current Biology: CB 22 (2): R45–46.
Quiroga, R. Quian, Z. Nadasdy, and Y. Ben-Shaul. 2004. “Unsupervised Spike Detection
and Sorting with Wavelets and Superparamagnetic Clustering.” Neural Computation
16 (8): 1661–87.
85

Raz, Aeyal, Vered Frechter-Mazar, Ariela Feingold, Moshe Abeles, Eilon Vaadia, and
Hagai Bergman. 2001. “Activity of Pallidal and Striatal Tonically Active Neurons Is
Correlated in MPTP-Treated Monkeys But Not in Normal Monkeys.” The Journal of
Neuroscience 21, no. 3 (February 1, 2001): RC128:1–RC128:5
Rovó, Zita, István Ulbert, and László Acsády. 2012. “Drivers of the Primate Thalamus.”
The Journal of Neuroscience: The Official Journal of the Society for Neuroscience
32 (49): 17894–908.
Sanger, Terence D. 2018. “A Computational Model of Deep-Brain Stimulation for
Acquired Dystonia in Children.” Frontiers in Computational Neuroscience 12
(September): 77.
Sanger, Terence D., Mark Liker, Enrique Arguelles, Ruta Deshpande, Arash Maskooki,
Diana Ferman, Aprille Tongol, and Aaron Robison. 2018. “Pediatric Deep Brain
Stimulation Using Awake Recording and Stimulation for Target Selection in an
Inpatient Neuromodulation Monitoring Unit.” Brain Sciences 8, no. 7 (July 17, 2018).
Saxena, Shreya, Sridevi V. Sarma, Shaun R. Patel, Sabato Santaniello, Emad N.
Eskandar, and John T. Gale. 2019. “Modulations in Oscillatory Activity of Globus
Pallidus Internus Neurons During a Directed Hand Movement Task: A Primary
Mechanism for Motor Planning.” Frontiers in Systems Neuroscience 13 (April 30,
2019): 15.
Schmidt, Robert, and Joshua D. Berke. 2017. “A Pause-Then-Cancel Model of Stopping:
Evidence from Basal Ganglia Neurophysiology.” Philosophical Transactions of the
Royal Society of London. Series B, Biological Sciences 372 (1718).
86

Shan, Kevin Q., Evgueniy V. Lubenov, and Athanassios G. Siapas. 2017. “Model-Based
Spike Sorting with a Mixture of Drifting T-Distributions.” Journal of Neuroscience
Methods 288 (August): 82–98.
Sohn, Youngh, and Mark Hallett. 2004. “Surround Inhibition in Human Motor System.”
Experimental Brain Research. Experimentelle Hirnforschung. Experimentation
Cerebrale 158 (4).
Suner, Selim, Matthew R. Fellows, Carlos Vargas-Irwin, Gordon Kenji Nakata, and John
P. Donoghue. 2005. “Reliability of Signals from a Chronically Implanted, Silicon-
Based Electrode Array in Non-Human Primate Primary Motor Cortex.” IEEE
Transactions on Neural Systems and Rehabilitation Engineering: A Publication of the
IEEE Engineering in Medicine and Biology Society 13 (4): 524–41.
Tang, Joyce K. H., Elena Moro, Andres M. Lozano, Anthony E. Lang, William D.
Hutchison, Neil Mahant, and Jonathan O. Dostrovsky. 2005. “Firing Rates of Pallidal
Neurons Are Similar in Huntington’s and Parkinson’s Disease Patients.”
Experimental Brain Research. Experimentelle Hirnforschung. Experimentation
Cerebrale 166 (2): 230–36.
Teager, H. 1980. “Some Observations on Oral Air Flow during Phonation.” IEEE
Transactions on Acoustics, Speech, and Signal Processing 28, no. 5 (October 1980):
599–601.
Teager, H. M., and S. M. Teager. 1990. “Evidence for Nonlinear Sound Production
Mechanisms in the Vocal Tract.” Speech Production and Speech Modelling 55.
87

Tisch, Stephen, John C. Rothwell, Ludvic Zrinzo, Kailash P. Bhatia, Marwan Hariz, and
Patricia Limousin. 2008. “Cortical Evoked Potentials from Pallidal Stimulation in
Patients with Primary Generalized Dystonia.” Movement Disorders: Official Journal
of the Movement Disorder Society 23 (2): 265–73.
Vitek, Jerrold L. 2002. “Pathophysiology of Dystonia: A Neuronal Model.” Movement
Disorders 17, no. S3 (March 2002): S49–S62.
Vitek, J. L., V. Chockkan, J. Y. Zhang, Y. Kaneoke, M. Evatt, M. R. DeLong, S. Triche, K.
Mewes, T. Hashimoto, and R. A. Bakay. 1999. “Neuronal Activity in the Basal Ganglia
in Patients with Generalized Dystonia and Hemiballismus.” Annals of Neurology 46
(1): 22–35.
Wild, Jiri, Zoltan Prekopcsak, Tomas Sieger, Daniel Novak, and Robert Jech. 2012.
“Performance Comparison of Extracellular Spike Sorting Algorithms for Single-
Channel Recordings.” Journal of Neuroscience Methods 203 (2): 369–76.
Yokoyama, T., K. Sugiyama, S. Nishizawa, T. Tanaka, N. Yokota, S. Ohta, and K.
Uemura. 1998. “Neural Activity of the Subthalamic Nucleus in Parkinson’s Disease
Patients.” Acta Neurochirurgica 140, no. 12 (December 1998): 1287–1291.
Zhuang, Ping, Yongjie Li, and Mark Hallett. 2004. “Neuronal Activity in the Basal Ganglia
and Thalamus in Patients with Dystonia.” Clinical Neurophysiology: Official Journal
of the International Federation of Clinical Neurophysiology 115 (11): 2542–57.

Abstract (if available)

Abstract For patients suffering from severe dystonia, deep brain stimulation (DBS) may be an effective treatment to alleviate motor symptoms. Although effective, the level of efficacy varies from patient to patient, likely due to differences in the etiology and level of injury. The challenges to improve DBS therapy rely on two fundamental questions: what are the neural patterns of abnormality leading to the development of dystonic symptoms? and, how does electrical stimulation can effectively normalize these patterns? We recently developed an inpatient multielectrode recording and stimulation protocol for the selection of DBS target in pediatric patients with movement disorders. While this protocol was designed to solve a clinical problem, it provides an unprecedented opportunity for studying the electrophysiology of brain areas involved in the control of movement and how neural patterns of activation correlate with motor behavior, moreover, it offers the possibility to study the response of the neural tissue to electrical stimulation and how DBS may modify abnormal neural patterns causing movement disorders. In order to explore these questions, it is essential to first understand the electrophysics of these specific recordings. Depth multielectrodes approved for human use consist of arrays of blunt microelectodes which physical properties differ from standard sharp microelectrodes commonly used for spike recordings. Large microelectrodes can record average activity from a relatively large volume of neurons, while small microelectrodes can detect more localized neural single unit activity. The microelectrodes used in this recording setup are larger than the standard, but small enough to register spiking activity. This unique characteristic results in low signal-to-noise ratio (SNR) recordings in which the spike waveforms are buried in the background neural noise. First, we developed a new methodology to detect single unit activity from low SNR recordings. The spike detection rate and spike classification were improved by introducing a generalization of the Teager-Kaiser nonlinear energy operator which accounts not only for the spike amplitude but also for the instantaneous energy content of the spike waveform. The proposed algorithm was efficient and therefore suitable for processing the terabytes of data commonly collected during the inpatient multielectrode recording protocol. It was also robust to high noise content and to the selection of the detection threshold. Once the necessary spike detection methodology was established, firing rates were analyzed for multiple basal ganglia and thalamic nuclei in pediatric subjects with secondary dystonia. The correlation between the neural activation of GPi (internal globus pallidus) and thalamic neurons was explored during the execution of voluntary movement. Two relevant observations arose from this analysis. First, the vast majority of neurons in GPi and thalamic nuclei fired with a median frequency below 5 Hz. This contrasts with typical reports of intraoperative GPi discharge rates between 10 and 70 Hz, likely biased towards fast firing neurons given the short recording times in the intraoperative setting. Second, increased widespread activation of the GPi positively correlated with thalamic activation during voluntary movement. This observation contrasts with the traditional rate model of the basal ganglia which suggests that activation of the GPi results in widespread thalamic inhibition, which leads to suppression of cortical motor commands to finally inhibit movement. Instead, we propose a new model in which the basal ganglia can selectively inhibit different dynamics by control of the inhibitory output from GPi and therefore regulating ongoing thalamocortical activity for the selection of the appropriate motor command. Finally, we analyzed the effects of neural depolarization due to DBS in motor areas. We present the application of a stimulation polarity reversal technique to reduce artifact contamination in short latency evoked potential (EP) recordings. This technique was applied to intracranial microelectrode and electroencephalography (EEG) recordings to study the propagation of the DBS pulse through the brain. Thalamic and cortical EPs in response to low frequency stimulation of the GPi (GPi-DBS) were analyzed, then the GPi-DBS frequency was varied, and the thalamic EPs were measured. We found that low frequency GPi-DBS elicited an evoked response in its projecting thalamic nuclei, however, no cortical response was observed. Thalamic EPs in response to GPi-DBS had a peak amplitude within 2 ms after stimulation onset, which decreased as the stimulation frequency increased, showing the frequency dependent modulatory effects of DBS on the neural network. These findings have not been previously reported, probably because most studies of intracranial EPs record from the implanted DBS macroelectrodes, which large surface area average a large volume of neural activity, therefore losing spatial and temporal resolution. The observations resulting from this study are consistent with a model in which blocking downstream propagation of DBS may be one way in which DBS can function as a lesion, substituting patterns of variation of neural activation with meaningful information, for regular neural patterns with no variation and therefore irrelevant information.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Next generation neural interfaces for neural modulation and prosthesis

PDF

Facilitating myocontrol for children with cerebral palsy

PDF

Investigating the role of muscle physiology and spinal circuitry in sensorimotor control

PDF

Simulating electrical stimulation and recording in a multi-scale model of the hippocampus

PDF

Improving stimulation strategies for epiretinal prostheses

PDF

Design and application of a C-shaped miniaturized coil for transcranial magnetic stimulation in rodents

PDF

Detection and decoding of cognitive states from neural activity to enable a performance-improving brain-computer interface

PDF

Data-driven modeling of the hippocampus & the design of neurostimulation patterns to abate seizures

PDF

Nonlinear dynamical modeling of single neurons and its application to analysis of long-term potentiation (LTP)

PDF

Electromyography of spinal cord injured rodents trained by neuromuscular electrical stimulation timed to robotic treadmill training

PDF

Brain connectivity in epilepsy

PDF

Stimulation strategies to improve efficiency and temporal resolution of epiretinal prostheses

PDF

Neuromuscular dynamics in the context of motor redundancy

PDF

Decoding memory from spatio-temporal patterns of neuronal spikes

PDF

Corticomotor excitability of gluteus maximus: influence on hip extensor strength and hip mechanics

PDF

A million-plus neuron model of the hippocampal dentate gyrus: role of topography, inhibitory interneurons, and excitatory associational circuitry in determining spatio-temporal dynamics of granul...

PDF

From sensory processing to behavior control: functions of the inferior colliculus

Asset Metadata

Creator Argüelles Morales, Juan Enrique (author)

Core Title On the electrophysiology of multielectrode recordings of the basal ganglia and thalamus to improve DBS therapy for children with secondary dystonia

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 06/02/2020

Defense Date 12/02/2019

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

Tag deep brain stimulation,dystonia,electrophysiology,motor control,movement disorders,multielectrode recordings,neuromodulation,OAI-PMH Harvest,signal processing

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Sanger, Terence David (committee chair), Lee, Brian (committee member), Song, Dong (committee member), Valero-Cuevas, Francisco (committee member)

Creator Email jarguell@usc.edu,jenrique.arguelles@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-314246

Unique identifier UC11665303

Identifier etd-ArguellesM-8563.pdf (filename),usctheses-c89-314246 (legacy record id)

Legacy Identifier etd-ArguellesM-8563-0.pdf

Dmrecord 314246

Document Type Dissertation

Rights Argüelles Morales, Juan Enrique; Arguelles Morales, Juan Enrique

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