Intelligent Medical Devices for Brain State Classification
and Responsive Neuromodulation
by
Gerard O’Leary
A thesis submitted in conformity with the requirements
for the degree of Master of Applied Science
Edward S. Rogers Sr. Department of Electrical & Computer Engineering
University of Toronto
© Copyright by Gerard O’Leary 2018
ii
Intelligent Devices for Brain State Classification
and Responsive Neuromodulation
Gerard O’Leary
Master of Applied Science
Edward S. Rogers Sr. Department of Electrical & Computer Engineering
University of Toronto
2018
Abstract
Recent advances in implantable electronics, machine learning and fundamental neuroscience have
enabled the individualized artificial control of neural systems. This has led to breakthrough
therapies for neurological disorders such as epilepsy. This thesis describes the design and
implementation of a framework for chronically-implanted, intelligent devices for neural signal
recording and responsive stimulation. The low-power signal processing required to extract
pathological biomarkers is described, and the process of classifying brain states based on such
biomarkers is outlined. A hardware-efficient approach to time-series classification is introduced:
the exponentially decaying memory support vector machine (EDM-SVM). Biomarker extraction,
classification, and chronically-safe neurostimulation is integrated in an implantable integrated
circuit: the neural interface processor (NURIP). The continuous adaptation of NURIP over the
lifetime of an individual is essential to maximize classification accuracy and treatment efficacy.
The design of a patient-localized device training microserver is described to achieve this goal.
iii
This thesis was created through a brain’s struggle to understand itself,
and its naïve quest to build tools to investigate itself.
I dedicate this work to the brain in the hope that its pursuit is not futile,
but “the struggle itself toward the heights is enough to fill a man's heart”.
- Albert Camus
iv
Acknowledgments
Firstly, I would like to thank Professor Roman Genov for giving me the opportunity to pursue this
research, and for the trust, guidance and freedom he gave me to explore new directions. He enabled
all of the work in this thesis. I am extremely grateful to have worked with my co-supervisor, Dr.
Taufik Valiante. His relentless work ethic, his dedication to his craft, his pursuit of excellence in
his research and his humble genius are a continued source of inspiration to me. Thank you to Dr.
David Groppe for his modest brilliance. Our discussions have greatly influenced this work and our
collaborations have been as enjoyable as they have been productive.
I would like to thank Prof. Jose-Luis Perez Velazquez from the Hospital for Sick Children for his
work in laying the foundation of this research. Thank you to CMC microsystems for support and
generous access to CMOS fabrication processes. I would also like to thank my defense committee:
Professor Stark Draper, Professor Paul Chow and the committee chair, Professor Paul Yoo for
committing their time to contribute to this process.
Thank you to my fellow graduate students and friends at the University of Toronto. Particularly to
the souls of BA5158; Ahmed, Asish, Gairik, Maged, Navid, Reza and Wilfred. A special thanks
to Javid for late-night blackboard brainstorming. To Dr. Nikola Katic for his contagious laughter
and inspirational research ethic. I must thank Dr. Saber Amini for his disturbing entertainment. To
Firas and the lessons of the Stoics. Thank you to the extraordinary minds at the Valiante Lab. To
Chaim for his passion and work rate. To Marjan for bringing music to my ears. To Michael, where
every encounter brought a new adventure and a new story. To Victoria for her amazing diligence.
To Sara, Homeira and Sumayya. A special thanks to the summer students who have contributed to
this research; Akshay, Ali, Shahryar, Ted and Veronica.
I am very grateful to Laura Fujino and K.C. Smith for believing in Jin Hee and myself to lead the
Saratoga student volunteer team at the IEEE International Solid-State Circuits Conference each
year in San Francisco. Thank you to all the team members and leaders, past and present, for
sacrificing sleep and sanity for a smooth operation and unforgettable adventures.
Thank you to my parents and family for their perspectives and support throughout many
questionable decisions. To my father for his unquestionable morals, to my mother teaching me
many life lessons. Finally, to Erin, for her support, love and patience.
v
Table of Contents
Acknowledgments.......................................................................................................................... iv
Table of Contents .............................................................................................................................v
List of Tables ............................................................................................................................... viii
List of Figures ................................................................................................................................ ix
List of Abbreviations .................................................................................................................... xii
Chapter 1 Introduction..................................................................................................................1
1.1 Thesis Outline ......................................................................................................................1
1.2 Measuring Neural Activity ..................................................................................................3
1.2.1 Physiological Brain States .......................................................................................4
1.2.2 Pathological Brain States .........................................................................................5
1.3 Brain State Classification .....................................................................................................7
1.3.1 Machine learning .....................................................................................................7
1.3.2 Support Vector Machines ........................................................................................8
1.3.3 Quantifying Classifier Performance ......................................................................11
1.4 Electrical Neural Stimulation .............................................................................................12
1.4.1 History....................................................................................................................12
1.4.2 Mechanisms ...........................................................................................................13
1.4.3 Side-effects ............................................................................................................14
1.4.4 Charge Balancing ...................................................................................................14
1.5 Implantable Medical Devices for Neuromodulation..........................................................15
1.5.1 Open-Loop Devices ...............................................................................................15
1.5.2 Closed-Loop Devices .............................................................................................15
Chapter 2 On-Device Neural Signal Processing ........................................................................19
2.1 Spectral Features ................................................................................................................19
vi
2.2 Phase Locking Value .........................................................................................................20
2.2.1 Hardware Architecture ...........................................................................................22
2.2.2 CORDIC ................................................................................................................22
2.2.3 Design Optimizations.............................................................................................23
2.3 Cross-frequency Coupling .................................................................................................24
2.3.1 Introduction ............................................................................................................24
2.3.2 Measuring CFC ......................................................................................................25
2.3.3 System Architecture ...............................................................................................26
2.3.4 Modulation Signal Extraction ................................................................................26
2.3.5 CFC Processor Implementation .............................................................................27
2.3.6 Surrogate Analysis .................................................................................................31
2.3.7 Methods..................................................................................................................32
2.3.8 Results ....................................................................................................................32
2.3.9 Discussion ..............................................................................................................34
Chapter 3 Intelligent Closed-Loop Neuromodulation Devices ................................................35
3.1 Machine-Learning-Based Seizure Detection .....................................................................35
3.2 Neural Signal Time Series Classification ..........................................................................35
3.2.1 Exponentially Decaying Memory ..........................................................................36
3.3 Arbitrary Waveform Neurostimulation ..............................................................................37
3.4 NURIP: The Neural Interface Processor ............................................................................38
3.4.1 Architecture Overview ...........................................................................................38
3.4.2 Data Preprocessing.................................................................................................39
3.4.3 Feature Extraction ..................................................................................................40
3.4.4 Classification..........................................................................................................41
3.4.5 Responsive Neurostimulation ................................................................................42
3.4.6 Results ....................................................................................................................43
vii
3.4.7 Chip Micrograph and Comparison.........................................................................44
Chapter 4 Continuous Lifelong Device Personalization...........................................................45
4.1 The Need for Chronic Neural Recordings .........................................................................45
4.2 Existing Long-Term Monitoring Solutions .......................................................................46
4.2.1 NeuroPace Remote Monitor ..................................................................................46
4.2.2 Medtronic Patient Programmer ..............................................................................46
4.2.3 Medtronic Nexus ....................................................................................................47
4.3 Machine Learning Microserver for Neuromodulation Device Training ............................48
4.3.1 Introduction ............................................................................................................48
4.3.2 System Overview ...................................................................................................50
4.3.3 Feature Extraction ..................................................................................................51
4.3.4 One-Class Support Vector Machine ......................................................................52
4.3.5 Hardware Implementation .....................................................................................53
4.3.6 Training Method ....................................................................................................54
4.3.7 Results ....................................................................................................................56
Chapter 5 Conclusion ..................................................................................................................58
5.1 Contributions and Relevant Publications ...........................................................................58
5.2 Future Directions ...............................................................................................................59
5.2.1 Clinical Trial ..........................................................................................................59
5.2.2 Adaptive Stimulation Parameterization .................................................................61
5.2.3 Additional Biomarkers ...........................................................................................61
5.2.4 Towards Seizure Prediction ...................................................................................61
5.2.5 Alternative Stimulation Modalities ........................................................................62
References .....................................................................................................................................63
viii
List of Tables
Table 1: Understanding a classifier’s output ................................................................................ 12
Table 2: FPGA Resource Utilization ............................................................................................ 54
Table 3: EU Database System Performance ................................................................................. 57
ix
List of Figures
Figure 1: Early neural recordings using the Lippman electrometer. .............................................. 3
Figure 2: EEG analog to digital conversion. ................................................................................... 4
Figure 3: SVM hyperplane selection. ............................................................................................. 9
Figure 4: Maximum Hyperplane Margin. ....................................................................................... 9
Figure 5: SVM Slack Variable. ..................................................................................................... 11
Figure 6: Medtronic open-loop deep brain stimulator. ................................................................. 15
Figure 7: Neuropace RNS. ............................................................................................................ 16
Figure 8: Medtronic Activa PC+S. ............................................................................................... 18
Figure 9: Signal band energy. ....................................................................................................... 20
Figure 10: Phase locking value. .................................................................................................... 20
Figure 11: Phase locking value mean complex vector. ................................................................. 22
Figure 12: Phase locking value area reduction. ............................................................................ 23
Figure 13: Cross-frequency coupling. ........................................................................................... 25
Figure 14: CFC processing system architecture. .......................................................................... 26
Figure 15: Modulation signal extractor......................................................................................... 27
Figure 16: Cross-frequency phase locking value (CF-PLV) architecture. .................................. 28
Figure 17: Limitation of CF-PLV. ................................................................................................ 29
Figure 18: The heights ratio. ........................................................................................................ 29
Figure 19: MVL-MI complex vector series. ................................................................................. 30
x
Figure 20: Mean vector length modulation index extractor. ......................................................... 31
Figure 21: Slice electrode arrangement for signal acquisition..................................................... 32
Figure 22: Computed measures of PAC using MVL-MI and CF-PLV. ....................................... 33
Figure 23. CFC measurement latency. ......................................................................................... 34
Figure 24: EDM Array. ................................................................................................................. 36
Figure 25: NURIP architecture for responsive neuromodulation. ................................................ 39
Figure 26: Preprocessor and signal band filtering. ....................................................................... 40
Figure 27: Feature extraction and EDM-SVM classification. ...................................................... 41
Figure 28: Charge-balanced neuromodulation architecture. ......................................................... 42
Figure 29: EDM-SVM output and NURIP power consumption. ................................................. 43
Figure 30: NURIP SoC micrograph and comparison table. ......................................................... 44
Figure 31: Neuropace remote monitor. ......................................................................................... 46
Figure 32: Medtronic DBS device programmer. .......................................................................... 47
Figure 33: Training microserver concept. ..................................................................................... 50
Figure 34: Training microserver implementation overview. ........................................................ 50
Figure 35: EDM decay rate. .......................................................................................................... 51
Figure 36: One-class support vector machine. .............................................................................. 53
Figure 37: Prototype training microserver. ................................................................................... 53
Figure 38: SVM training time. ...................................................................................................... 55
Figure 39: Microserver processing example. ................................................................................ 56
xi
Figure 40: Proposed clinical trial setup. ....................................................................................... 60
Figure 41: Implementation of NURIP. ......................................................................................... 60
xii
List of Abbreviations
BRAM Block Random-Access Memory
CFC Cross Frequency Coupling
CMOS Complementary Metal–Oxide–Semiconductor
CORDIC COordinate Rotation DIgital Computer
DBS Deep Brain Stimulation
EDM Exponentially Decaying Memory
EEG Electroencephalography
EMU Epilepsy Monitoring Unit
FPGA Field Programmable Gate Array
NURIP NeUral Interface Processor
PAC Phase Amplitude Coupling
PD Parkinson’s Disease
PLV Phase Locking Value
RNS Responsive Neural Stimulation
SVM Support Vector Machine
VLSI Very-Large-Scale Integration
1
Chapter 1
Introduction
Recent advances in implantable electronics, machine learning and fundamental neuroscience have
enabled the individualized artificial control of neural systems. This has led to breakthrough
therapies for the treatment of neurological disorders such as epilepsy. This thesis presents the
design and implementation of a device-based framework for chronically-implanted, intelligent,
closed-loop neural signal recording and responsive stimulation. The process of extracting
biomarkers from recorded neural signals in low-power devices will be described, and the
implementation of three effective approaches will be outlined: signal band energy, phase
coherence, and cross-frequency coupling. An array of these biomarkers can be used to support the
classification of a wide array of normal physiological and pathological brain states. An effective
and efficient solution to performing time series classification in an implanted device will be
presented in the form of the exponentially decaying memory support vector machine (EDM-SVM).
The use of responsive neuromodulation upon the detection of a pathological state will be outlined,
and a safe approach to more precise stimulation using arbitrary waveforms will be presented. These
concepts are integrated on a single electronic system on chip (SoC) in the form of NURIP, the
neural interface processor. The path towards adapting machine learning based devices over the
lifetime of a patient will be outlined and a device-localized microserver solution will be presented.
Such developments are necessary to improve the efficacy of existing closed-loop devices to
maximize the quality of treatment for patients.
1.1 Thesis Outline
This thesis is organized as follows:
▪ Chapter 1 gives the necessary background information to put the proceeding chapters in
context. First, an introduction to neural signal recording and the concept of brain states is
detailed. A method to classify brain states is then outlined with an introduction to machine
2
learning and the support vector machine algorithm. An overview of electrical neural
stimulation is then presented. Finally, these areas are tied together with a discussion on
existing implantable neuromodulation devices.
▪ In Chapter 2, the signal processing required to extract pathological biomarkers on an
implantable device is described. Improvements upon the state of the art in phase-based
biomarkers is outlined. A novel hardware implementation for the extraction of Cross-
Frequency Coupling metrics is then presented.
▪ Chapter 3 provides a background and description of an implantable integrated circuit for
brain state classification and responsive neurostimulation known as NURIP, the Neural
Interface Processor. The process of classifying brain states based on an array of extracted
biomarkers is first introduced. A new time series classification approach, the Exponentially
Decaying Memory Support Vector Machine (EDM-SVM), is then outlined. The motivation
for more complex, chronically-safe stimulation waveforms is then presented. Finally, an
integrated circuit is presented which incorporates these concepts.
▪ In Chapter 4, the motivation for continuous device adaptation throughout the lifetime of a
patient is described. The design of a patient-localized microserver to support continuous
model training post-implantation is outlined.
▪ Chapter 5 will conclude the thesis by summarizing the results and discussing the significance
of this work. The thesis will then close by offering some insights for future investigation.
The remainder of this chapter will focus on the background needed to answer four key questions:
1) How can we measure brain states?
2) How can we intelligently classify brain states?
3) What is neuromodulation?
4) What are neurostimulation devices?
3
1.2 Measuring Neural Activity
In August 1875, Richard Caton reported to the British Medical Association in Edinburgh that he
had observed electrical impulses from the surfaces of living brains in the rabbit and monkey using
a Galvonometer (an electromechanical instrument for indicating electric current) [1]. The basis of
this activity was uncovered by Edgar Adrian in his 1928 publication "The Basis of Sensation" [2].
He would later go on to receive the Nobel Prize for his work revealing the function of neurons.
These original recordings visualized discharges in single nerve fibers using a Lippmann
electrometer, in which neural activity caused a corresponding jump in the mercury level inside a
calibrated tube (Figure 1).
Figure 1: Early neural recordings using the Lippman electrometer. The instrument displays a current
measurement in a similar manner to a mercury thermometer. The neuron’s action current causes a corresponding jump
in the mercury level in a calibrated tube [3].
Since this point, electronic recording systems have evolved to enable precise recording of many
brain regions simultaneously. The advent of CMOS integrated circuits brought with it new
instruments to convert analog signals to digital signals. A broad range of these analog to digital
convertors (ADCs) have been proposed for neural applications, a review of which can be found in
[4]. The key implication is that neural activity can be reliably recorded and stored for
computational analysis (Figure 2).
4
Figure 2: EEG analog to digital conversion. EEG activity can be digitized using electronic analog to digital
converters. Signal levels are quantized to binary values allowing the data to be processed and analyzed.
It is still impractical to record the electrical activity of all cells in the brain concurrently [5],
however, when analyzing neural networks at the macro level, emergent oscillations caused by
synchronized firing of local neural populations contains useful information about the brain’s
activity [6]. These oscillations, known as local field potentials, or LFPs, first came to light with
Hans Berger’s discovery of the Alpha wave (7–13 Hz), and the birth of electroencephalography
(EEG) in 1929 [7]. These oscillations are now understood to reflect the underlying activity of the
brain, and categorizing their patterns has led to fundamental insights into physiological and
pathological neural mechanisms [8].
1.2.1 Physiological Brain States
A brain state can be defined here as a snapshot of the measurable characteristics of the nervous
system in a behavioral state. For example, sleep is a recurring state of mind and body, characterized
by altered consciousness, relatively inhibited sensory activity and inhibition of nearly all voluntary
muscles [9]. This behavioral state (which itself contains sub-states known as sleep stages),
coincides with LFPs which have common characteristics across the human population. As an
individual falls asleep, the EEG initially transitions from a state of high-frequency, low-voltage
waves in the waking state to higher voltage, slower waves representing non-rapid eye movement
(NREM) sleep.
5
1.2.2 Pathological Brain States
The discovery of anomalous EEG patterns in pathological conditions such as epilepsy and
Parkinson’s disease has led to breakthrough therapies based on disrupting activity associated with
seizures and motor tremor respectively. The work presented in this thesis will focus on epilepsy,
with some work which is directly applicable to Parkinson’s Disease. A brief background on these
two disorders is presented in the following sections.
1.2.2.1 Epilepsy
Epilepsy is one of the most common neurological disorders, affecting approximately 50 million
people worldwide. A third of these individuals are not successfully treated with current anti-seizure
medications [10]. Epilepsy patients suffer from chronic unprovoked seizures, which can result in
broad spectrum of debilitating medical and social consequences. A seizure is a pathological brain
state defined by the International League against Epilepsy (ILAE) as a “transient occurrence of
signs and/or symptoms due to abnormal excessive or synchronous neuronal activity in the brain,”
and can come abruptly without warning [11].
Seizures can be categorized into clinical seizures and subclinical seizures. A clinical seizure results
in subjective symptoms or objective signs, while subclinical, or electrographic seizures appear
only on recorded EEG [12]. Clinical seizures can be further divided into partial seizures, where
seizures begin focally, and generalized seizures, where seizures begin in multiple brain regions
simultaneously. The anomalous EEG events found in epilepsy are referred to as ictal activity,
whereas interictal refers to activity between seizures which corresponds to over 99% of the
patient’s life.
Mesial temporal lobe epilepsy (MTLE) is the most common form of medically refractory, or drug-
resistant epilepsy, and surgical resection of epileptogenic tissue results in a high rate of seizure
freedom [13]. However, resection of effected tissue from areas such as the hippocampus can result
in considerable side effects including reduced verbal memory [14]. Thus, the ability to prevent or
abort seizures without resection or ablation (partial tissue removal) would represent a major
advance in the surgical treatment of epilepsy.
6
1.2.2.1.1 The Epilepsy Monitoring Unit
Determining a treatment for those with refractory epilepsy often requires an observation period in
an Epilepsy Monitoring Unit (EMU) where neural recordings are analyzed to localize the seizure
onset zone. The focus of a seizure can be identified from electrophysiological testing (e.g. EEG).
functional testing (e.g. MEG, fMRI) and structural testing (e.g. MRI). Intracranial EEG, or iEEG,
is currently the most accurate method for seizure localization in patients with intractable epilepsy
who are candidates for resective surgery [15]. Intracranial EEG recording can be performed either
using electrodes placed on the surface of the brain (subdural grid and strip electrodes) or by
electrodes inserted into the brain (depth electrodes).
Recording from the cortical surface using subdural electrodes is referred to as electrocorticography
(ECoG), whereas EEG recording using multiple depth electrodes is referred to as
stereoencephalography (sEEG). ECoG recordings provide a better signal-to-noise ratio when
compared to scalp surface EEG, along with capturing a higher signal bandwidth and an increased
spatial resolution. Unlike surface EEG and ECoG, the depth electrodes used in sEEG can be used
to record activity from deeper brain structures such as the hippocampus.
The implanted electrodes are connected to clinically approved electrophysiological neural
recording systems. These systems are monitored in conjunction with visual recordings on a 24-
hour basis by clinical technicians to identify segments of interest to epileptologists and to alert
clinical staff early in the development of a seizure. This is critical to enhance patient safety, and
for the timely clinical assessment necessary for semiological classification of seizures [16]. During
the observation period, some patients discontinue, or reduce their dosages of anti-seizure
medications to allow the generation of ictal activity for localization purposes. Clinicians may also
electrically stimulate the implanted electrodes to analyze the response and isolate the seizure
source. The number of seizures needed to determine seizure localization reliably depends on the
reproducibility of the ictal pattern. In general, 3 to 9 seizures are recorded to establish a consistent
site of origin [17]. When sufficient information is gathered, courses of treatment can include the
resection of the suspected epileptogenic zone or the implantation of a seizure control device which
will be described in the coming sections. Another disorder, Parkinson’s Disease (PD), has greatly
benefited from related treatments and is introduced in the following section.
7
1.2.2.2 Parkinson’s Disease
More than 180 years ago, James Parkinson first described the disorder that bears his name. The
disease is a chronic progressive neurodegenerative disorder of the central nervous system that
mainly affects the motor system. Early symptoms include shaking, rigidity, slowness of
movement, and difficulty with walking [18]. The disorder is characterized by the death of
dopaminergic neurons in the basal ganglia, a brain structure involved in voluntary motor
movements [19].
Typically, patients with Parkinson's disease have a robust response to one or more medications
such as Levadopa. As the disease progresses and neurons continue to be lost, these medications
become less effective. Furthermore, typically after 5 years of therapy, medication-related
complications develop in many patients. Such complications include “on–off” fluctuations, in
which a sudden, sometimes unpredictable loss of benefit from medication occurs [20]. Alternative
therapies are actively sought, and the use of electrical stimulation as outlined in Section 1.4 has
shown excellent treatment efficacy in those who no longer respond to existing medications [21].
1.3 Brain State Classification
The process of classifying brain activity starts with the understanding that a given brain state has
a pattern in measurable neural signal characteristics. In the case of epilepsy, a seizure state can be
characterized by abnormal excessive or synchronous EEG activity. Given a data set with recorded
neural activity, the following section will introduce the concept of machine learning which allows
us to classify brain states within this data.
1.3.1 Machine learning
Machine learning evolved from the study of pattern recognition and computational learning theory
in artificial intelligence. Machine learning algorithms are constructed to learn from and make
predictions on data, through building a model from sample inputs. This is in contrast to following
strictly static program instructions by making data-driven predictions or decisions [22].
Two broad categories of machine-learning frameworks exist: discriminative and generative. The
key difference is that a generative algorithm models how the data was generated to categorize a
signal. It asks the question: based on the generation assumptions, which category is most likely to
8
generate this signal? A discriminative algorithm does not care about how the data was generated,
it simply categorizes a given signal.
The challenge with generative frameworks is that substantial training data can be needed to create
adequate models [23]; this is of particular concern in the clinical applications considered in this
work where data for pathological events such as seizures are rare. Discriminative models, however,
focus on specific variables of interest which is potentially more robust [24]. The work presented
in this thesis focuses on a discriminative approach to brain state classification.
Machine learning models fall under supervised, unsupervised, and reinforcement learning
paradigms. In supervised learning, the algorithm is presented with example inputs and their desired
outputs, given by a "teacher", and the goal is to learn a general rule that maps inputs to outputs.
The second general category is unsupervised learning. In this paradigm, no labels are given to the
learning algorithm, leaving it on its own to find structure in its input. Reinforcement learning
enables model formation based on rewards and punishments. These cues are given as feedback in
response to the algorithms actions in a dynamic environment, such as playing a game against an
opponent [22]. The work in this thesis involves the use of supervised methods to learn the
distinction in EEG between normal physiological activity and seizure-related activity (see section
3.1), and utilizes unsupervised learning to identify generally anomalous activity (see section 4.3).
While recent work in the machine learning field has focused on high-performance approaches
without specific computational constraints [25], this thesis will focus on hardware-efficient
methods for implantable devices. An overview of one such method, the support vector machine,
is outlined in the following section.
1.3.2 Support Vector Machines
The support vector machine (SVM) is a supervised machine-learning framework for discriminative
classification that has gained popularity due to its computational efficiency and robustness [26].
Recent work has shown comparable performance to more complex approaches such as
convolutional neural networks (CNNs) [27].
In the case of support vector machines, a data point is viewed as an N-dimensional vector. The
fundamental idea behind the SVM is to find an N-dimensional hyperplane that can separate two
groups of input data points, which should also be mapped to the same high-dimensional space as
9
the hyperplane. There are many hyperplanes that might classify the data as shown in Figure 3.
While H2 separates both colors, this might not be the case if further samples were added to the
class distributions. This implies that the hyperplane is sensitive to noisy data points and would
result in erroneous classifications. The most reasonable choice is the one that represents the largest
separation, or margin, between the two classes, in this case, H3.
Figure 3: SVM hyperplane selection. Class data can be separated using many hyperplanes. The most reasonable
choice is the one that has the largest margin between both classes. Image from [28].
This linear boundary can be formulated as:
𝑓(�⃗�) = �⃗⃗⃗�𝑇�⃗� + b = 𝑤1𝑥1 + 𝑤2𝑥2 + 𝑤𝑛𝑥𝑛 + 𝑏 = 0 (1)
Hence, “learning” in this case involves optimizing �⃗⃗⃗� and 𝑏 to find the most suitable hyperplane,
where �⃗⃗⃗� is a normal vector perpendicular to the hyperplane, and 𝑏 is used to define the hyperplane
offset from the origin. To achieve this, we can select two parallel hyperplanes that separate the two
classes of data, so that the distance between them is as large as possible as shown in Figure 4.
Figure 4: Maximum Hyperplane Margin. Learning, in this case, involves optimizing �⃗⃗⃗� and 𝑏 to find the most
suitable hyperplane. Image from [28].
10
The region between these hyperplanes is known as the "margin", and the maximum-margin
hyperplane is the hyperplane that lies halfway between them. Geometrically, the distance between
these two hyperplanes is 2
‖�⃗⃗⃗�‖. Therefore, the optimization problem can be defined as:
Minimize ‖�⃗⃗⃗�‖ subject to 𝑦𝑖(�⃗⃗⃗� ∙ �⃗�𝑖 − 𝑏) ≥ 1, for 1 < 𝑖 < 𝑛 (2)
Where 𝑦𝑖 is the class of training vector �⃗�𝑖, and 𝑛 is the number of training points. By solving for
the Lagrangian dual of this optimization problem as outlined in [29], we can reach a solution more
efficiently using quadradic programming algorithms [30], resulting in the following decision
function for the model:
𝑓(�⃗�) = ∑(𝑎𝑖𝑦𝑖(�⃗�𝑖 ∙ �⃗� − 𝑏))
𝑁
𝑖=1
(3)
Here, 𝑎𝑖 are Lagrangian multipliers, where non-zero values correspond to a subset of training
vectors, �⃗�𝑖, which define the hyperplane. These points define our support vectors. This model can
be used to assign new examples to one category or another depending on which side of the
hyperplane they fall. Given 𝑓(�⃗�), the classification is obtained as:
𝑦 = sign(𝑓(�⃗�)) = {+1 𝑓(�⃗�) > 0
−1 𝑓(�⃗�) < 0 (4)
In the ideal case shown in Figure 3, both classes can be easily separated by hyperplane H3.
However, in practice the data to be classified can include noise, the features used may not be
informative enough to discern differences between the two classes and the data may not be linearly
separable. In such cases, we would like to find a hyperplane which separates the majority of the
data points to get the best generalized performance at the expense of misclassification of some
outliers. This can be achieved by modifying our optimization problem to allow a tolerance for
misclassification with a regularization parameter, C, as follows:
Minimize ‖�⃗⃗⃗�‖ + C‖𝜉‖ subject to 𝑦𝑖(�⃗⃗⃗� ∙ �⃗�𝑖 − 𝑏) ≥ 1 − 𝜉𝑖, for 1 < 𝑖 < 𝑛 (5)
Where 𝜉𝑖 defines how far the training example is from the correct class hyperplane as shown in
Figure 5, and C is a tunable parameter to set the tolerance of this error. For large values of C, the
optimization will choose a smaller-margin hyperplane if that hyperplane is more successful in
11
classifying the training points correctly. Conversely, a small value of C will cause the optimizer to
look for a larger-margin separating hyperplane, even if that hyperplane misclassifies more points.
Figure 5: SVM Slack Variable. A regularization parameter, C, can be used to control the error tolerance with respect
to 𝜉𝑖. Image from [31].
In addition to performing linear classification, SVMs can efficiently perform a non-linear
classification using what is called the kernel trick, implicitly mapping their inputs into high-
dimensional feature spaces. The mapping from the original space to the higher-dimensional space
is commonly implemented through a kernel function such as the Radial Basis Function (RBF),
which measures the distance between two vectors in the higher-dimensional space:
K(�⃗�, 𝑠𝑣⃗⃗⃗⃗⃗) = 𝑒−𝛾‖𝑠𝑣⃗⃗⃗⃗⃗ − �⃗�‖2 (6)
Where 𝑠𝑣𝑖 are the support vectors used to construct the hyperplane, K is implemented here as the
Radial Basis Function (RBF) kernel and 𝛾 is the inverse of the standard deviation of the RBF, or
Gaussian function. Intuitively, the gamma parameter defines how far the influence of a single
support vector reaches.
1.3.3 Quantifying Classifier Performance
The performance of a classifier can be quantified in terms of sensitivity and specificity. To
understand these terms, let’s take the example of seizure classification of a neural recording
12
segment in epilepsy, where a seizure is defined as ictal activity, and normal physiological activity
is defined as interictal activity.
When a classification is performed, and the test result is positive, the segment is classified as ictal
activity. If the test is negative, the segment is classified as interictal activity. However, the
classification may yield a wrong result by either falsely classifying an interictal segment if it is in
fact ictal, or by classifying a segment as ictal when it is in fact interictal. This can be summarized
in a two-by-two table:
Table 1: Understanding a classifier’s output
Ictal Segment Interictal Segment
Output Ictal True positive (TP) False positive (FP)
Output Interictal False negative (FN) True negative (TN)
The specificity or true negative rate (TNR) is defined as the percentage of segments which are
correctly identified as being interictal:
Specificity =TN
TN + FP (7)
The false positive rate is defined as 1 - specificity and represents the percentage of segments that
are incorrectly identified as ictal. The sensitivity, or true positive rate (TPR), is defined as the
percentage of segments which are correctly identified as ictal:
Sensitivity =TP
TP + FN (8)
1.4 Electrical Neural Stimulation
1.4.1 History
The first recorded clinical use of electrical stimulation was in 46 AD, involving the use of torpedo
fish for the relief of headaches [32]. In the late 18th century, Galvani discovered “a great secret of
life” when his assistant touched the exposed sciatic nerve of a frog with a metal scalpel that had
accumulated an electric charge. It was first observed that the muscles of the legs twitched when
struck by direct current on the nervous system [33]. The electrical excitability of the brain was
demonstrated in 1870 using stimulation of the motor cortex in dogs, resulting in limb movement
13
[34]. Since this time, the field has developed greatly with the introduction of sensory prosthetic
devices, such as cochlear implants, visual implants and spinal cord stimulators.
1.4.2 Mechanisms
When a metal electrode is placed inside a physiological medium, an artificial interface is formed.
In the metal electrode and in the attached circuits, charge is carried by electrons. In the electrolyte,
charge is carried by ions, including sodium, potassium, and chloride. The central process that
occurs at the electrode/electrolyte interface is a transduction of charge carriers from electrons in
the metal electrode to ions in the electrolyte [35]. Electrical stimulation can initiate a functional
response by depolarizing the membranes of excitable cells. Depolarization is achieved by the flow
of ionic current between two or more electrodes, at least one of which is in close proximity to the
target tissue [36]. The use of high-frequency electrical stimulation can induce continuous firing to
put cells into a refractory state [37], thus blocking their activity and effectively creating a
functional lesion [38].
These mechanisms can be used to selectively activate and deactivate inhibitory and excitatory
networks in the brain in a process known as neuromodulation. This process has led to the
development of deep brain stimulation (DBS), a surgical procedure now considered safe and well
tolerated to alleviate many neurological disorders [14] [21].
Over the last few decades DBS has been shown to provide remarkable therapeutic effect on
carefully selected Parkinson’s disease (PD) patients [21]. There remains considerable debate
concerning the mechanisms underlying its beneficial effect. The effect of stimulation on the motor
signs associated with PD are similar to ablation in the subthalamic nucleus (STN). This has led
many groups to conclude that DBS acts to suppress neuronal activity, decreasing output from the
stimulated site [39]. However, recent studies indicate that the affect may be influenced by the
activation of fibers inadvertently recruited near the stimulation target [21].
A recent study using DBS targeting the hippocampus in TLE patients was effective in significantly
reducing seizure frequency in patients with fifty-percent of the patients rendered seizure-free [40].
As in PD, the exact mechanism responsible for the efficacy is unknown, but studies suggest that
stimulation directly inhibits the seizure focus [41]. The mechanism in any given study likely
14
depends on several stimulus features including timing and waveform shape as will be discussed in
section 3.3.
1.4.3 Side-effects
An improperly designed electrical stimulation system may cause damage to the tissue being
stimulated, or to the electrode itself [42]. The mechanisms for stimulation induced tissue damage
are not well understood, but two major classes have been proposed. The first is that tissue damage
is caused by intrinsic biological processes as excitable tissue is overstimulated. This could be due
to depletion of oxygen or glucose, or changes in ionic concentrations both intracellularly and
extracellularly [42]. The second proposed mechanism for tissue damage is the creation of toxic
electrochemical reaction products at the electrode surface during cathodic stimulation at a rate
greater than that which can be tolerated by the physiological system. This will be discussed further
in the following section.
In terms of psychiatric side-effects, a wide range of behavioral changes may be seen following
DBS which depend greatly on the stimulation target. Depression was found to be the most common
side-effect in the treatment of PD, although this could be due to the disease itself [43]. In the case
of epilepsy, there have been reports of memory deficits in patients subject to stimulation of the
anterior nucleus of thalamus (ANT) [44].
1.4.4 Charge Balancing
In most applications, electrical stimulation is applied as a series of biphasic current pulses. Each
pulse has cathodal and anodal phases, with current amplitudes and durations that result in an
overall zero net charge deposition for the pulse. A cathodal current is reducing at the stimulation
electrode, with the direction of electron flow being from the electrode to the tissue. Anodal
indicates an oxidizing current with electron flow in the opposite direction [36]. The charge
delivered is the integral of the stimulating current over time.
It has been demonstrated the use of stimulus waveforms with a net direct current component
increases the probability of tissue damage [45]. Furthermore, being imbalanced in terms of charge
yields formation of undesirable chemical reactions at the electrode contact surface [46]. Even if a
stimulus pulse pair is charge-balanced, an electrode may be polarized during delivery of the pulse
15
to a point that tissue or electrode-damaging irreversibility occurs. Charge-balancing must also be
maintained to prevent the diffusion of ions away from the electrodes range of influence [35].
In summary, the objective of charge-balancing is to maintain the electrode potential within a range
that does not induce irreversible reactions that degrade the electrode, damage tissue, or limit the
efficacy of the stimulation pulse.
1.5 Implantable Medical Devices for Neuromodulation
1.5.1 Open-Loop Devices
Neural stimulation devices can be broadly categorized by open-loop, and closed-loop control.
Open-loop stimulation involves the use of regular pulses and is sometimes referred to as a
“pacemaker for the brain”. One example is the Medtronic Kinetra [47], an implantable pulse
generator (IPG). The stimulator device is implanted under the skin in the chest and a wire from the
device is connected to depth electrodes in the brain as shown in Figure 6. The Kinetra allows DBS
through two leads programmed independently with one pulse generator. A clinician can configure
ON and OFF periods, the amplitude (0 to 10.5 V) and pulse width (60 to 450 μs). The stimulation
frequency (2 to 250 Hz) is the same for both channels.
Figure 6: Medtronic open-loop deep brain stimulator. The Medtronic Kinetra is implanted under the skin in the
chest to generate pulses which are delivered to electrodes implanted in the brain. Image from Medtronic.
1.5.2 Closed-Loop Devices
Closed-loop neurostimulation aims to suppress and stop seizure activity by delivering a stimulus
to the site of seizure onset in response to detected electrographic activity [48]. Thus the stimulation
is provided “as needed” and not on a continual basis, essentially only during the event of a seizure
occurrence[49]–[52]. This is unlike open-loop stimulation which involves the continuous or
16
prescheduled stimulation of the brain, respectively [48]. Furthermore, there is evidence that
suggests intermittent stimulation poses fewer risks to neural tissue than continuous
stimulation[53], [54]. Thus contingent neurostimulation provides individualized treatment [49]
and it has the advantage of specificity because stimulation can be targeted to a specific brain region
involved in the seizure as opposed to the typical systematic administration of pharmaceutical drugs
[48].
1.5.2.1 Neuropace Responsive Neurostimulator
Currently, in the US, the only approved device for direct electrical brain stimulation for epilepsy
is the Neuropace responsive neurostimulator system (RNS) shown in Figure 7 [55]. The RNS
delivers electrical stimulation in response to real-time ECoG based abnormalities in up to 4 bi-
polar channels recording at a sample rate of 250 Hz. The device relies on three basic signal
characteristics; line length [56], area [57] and bandpass extraction [58]. The parameters of these
indicators must be manually determined by the clinician on a patient-specific basis.
Figure 7: Neuropace RNS. The RNS delivers electrical stimulation in response to real-time ECoG based
abnormalities. Image from [55].
The line length detection feature [56] is used to identify changes in both amplitude and frequency.
This average line length is calculated as the sum of the sample-to-sample differences within a
window divided by the number of samples in the window. The line length is compared between
two sliding windows, one short-term (128 ms to 4 seconds), and one long-term (4 seconds to
16 minutes) [55]. A sudden change in the signal characteristics is detected when the short-term
measurement exceeds a threshold relative to the long-term measurement. As both amplitude and
17
frequency changes will affect the average line length, it is not a reliable indicator to be used in
isolation.
The area detection feature [57] is used to identify changes in the overall signal energy. Area is
defined as the average absolute value of the signal within a window. As with the line length feature,
a short-term average is compared to a long-term average, and detection occurs when a positive or
negative threshold is exceeded.
The bandpass feature is a common method first described in [58] to detect rhythmic and spiking
activity. The EEG signal is segmented at local minima and maxima resulting in ‘half-waves,’ the
amplitude and duration of which are representative of the amplitude and frequency components of
the ECoG. A seizure is said to be detected if the number of half-waves occurring within a given
window length exceeds a certain threshold. The bandpass feature is intended to act as an
approximation of conventional DSP filtering and can be used to detect activity within specific
frequency bands (e.g., theta, alpha, beta and gamma).
An electrical stimulus is generated in response to a combination of these three indicators.
Stimulation currents between 0.5 and 12 mA can be selected, pulse width can be set between 40
and 1000 μs, burst duration can be programmed between 10 and 5000 ms, and the frequency of
stimulation can be between 1 and 333 Hz.
It should be noted that the parameters for each biomarker extractor and the resulting stimulus
waveform must be manually tuned by a clinician to configure the seizure detection and stimulation
functionality of the device. For example the bandpass detector operation alone requires the manual
tuning of four parameters; the minimum and maximum frequencies that define the bandpass, the
minimum signal amplitude and the minimum signal duration [55].
1.5.2.2 Medtronic Activa PC+S
The Medtronic Activa PC + S neurostimulator (Figure 8), is a research device based on an existing
stimulator and telemetry system found in the open-loop Activa PC neurostimulator with additional
sensing, stimulation, and detection features [59]. Three additional components form the sensing
and data processing subsystem.
18
The first component is a brain activity sensing IC (BASIC) for detecting LFPs as described in [60].
This IC extracts specific bands of interest in the neural signal with a configurable center frequency.
A low-pass filter is used to reject stimulation interference allowing for simultaneous recording and
stimulation. The second component is a custom, three-axis accelerometer for detecting motion-
based biomarkers such posture, activity and tremors [61]. The third component is a general-
purpose microcontroller for custom algorithm processing, signal classification, telemetry
streaming and system control.
Figure 8: Medtronic Activa PC+S. The Activa PC+S is a closed-loop capable device currently available for research
purposes only. Image from [62].
The Activa PC+S is generalized bi-directional brain–machine interface in which closed-loop
functionality must be manually implemented on the general-purpose microcontroller.
With an understanding of the signal processing capabilities of existing neurostimulation devices,
the following chapter will begin to outline more advanced approaches to extracting pathological
biomarkers in a low-power manner.
19
Chapter 2
On-Device Neural Signal Processing
There exists a tradeoff in device signal processing between embedded computation, and wireless
transmission for remote processing. Devices must operate on a low power budget [63] to maximize
their battery life, and reduce the number of replacement surgeries which result in a risks of
infection and an additional clinical burden. A comparative study found local signal processing an
order of magnitude lower than typical wireless data transmission for a typical use case [64], but
this tradeoff should be carefully considered for more complex processing.
Furthermore, biomedical device designers must pay particular care to the security of any wireless
interfaces. A recent vulnerability can be seen in diabetic patients with implanted insulin pumps. It
has been shown that it’s possible for attackers to scan a public space from as much as 300 feet
away, find vulnerable pumps and force them to dispense fatal insulin doses [65]. In heart
pacemakers, an exploit has been demonstrated which allows attackers to shut off the
device, read and write to its memory. In the case of implantable cardioverter-defibrillators (ICDs),
attackers can deliver a high voltage shock of up to 830 volts [65]. It is clear that wireless device
designers must carefully consider wireless security, and that basic functionality can still be
maintained with the use of on-device processing if the device is compromised.
2.1 Spectral Features
To detect anomalous activity in neural signals, spectral energy in conventional physiological signal
bands can be used to indicate electrographic events [66]. Clinical studies have determined that
seizure onset information is contained in the spectral energy distribution of the patient’s EEG [67].
An example of this is shown in the analysis of a typical patient in the EU epilepsy database shown
in Figure 9, where elevated band energy is seen during seizure events. Signal bands of interest
include Delta (<4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz) and Gamma (30-60 Hz)
bands. These can be extracted using conventional spectral estimation methods such as the Fast
Fourier Transform or more efficiently by passing recorded samples through parallel bandpass
20
filters [64]. A mechanism to extract these bands in a power-efficient manner will be outlined in
section 3.4.
Figure 9: Signal band energy. Conventional physiological signal bands can be used to indicate electrographic such
as the elevated energy levels shown here at the onset of a seizure.
2.2 Phase Locking Value
Neural signals can be analyzed not just in terms of their frequency components and magnitudes,
but also in terms of their phase. Taking two extracted oscillations, 𝑓0 and 𝑓1, when the difference
between their instantaneous phases is constant, synchronization is present between the two signals.
Recent work has led to the discovery of a “preictal state” characterized by a desynchronization of
the neuronal populations related to the epileptogenic focus before a seizure onset [68] as illustrated
in Figure 10. In contrast, high levels of synchronization observed towards the end of seizures may
facilitate termination [69]. Synchronization (or desynchronization) can be determined by a
measure exceeding (or falling below, respectively) a defined threshold.
Figure 10: Phase locking value. A desynchronization of the neuronal populations related to the epileptogenic focus
can be used as an indicator for seizure onset.
21
Estimation of this property using the Phase Locking Value (PLV) has emerged as a popular leading
method of quantifying neural synchronization [70] [71]. Quantifying the phase locking between
two neural signals requires the computation of the phase difference followed by the computation
of PLV. First, the Hilbert transform is applied to both signals 𝑓0 and 𝑓1 to obtain an analytic signal:
𝐴0 = 𝑓0 + 𝑗𝑓0̃ 𝐴1 = 𝑓1 + 𝑗𝑓1̃ (9)
where 𝑓�̃� is the Hilbert transform of 𝑓𝑖. 𝑓0 and 𝑓1 are two sinusoidal continuous or discrete-time
signals. The Hilbert transform is conventionally performed over the full band of frequencies in the
neural spectrum, and thus, a narrow-band bandpass filter is applied before the Hilbert transform to
isolate the signal band of interest [72]. The magnitude in the extracted frequency band can be
computed as:
𝑀𝐴𝐺(𝐴𝑖) = √[𝑅𝑒(𝐴𝑖)]2 + [𝐼𝑚(𝐴𝑖)]2 (10)
Next, the instantaneous phases are computed for each channel:
ϕi = 𝑎𝑟𝑐𝑡𝑎𝑛 (𝐼𝑚(𝐴𝑖)
𝑅𝑒(𝐴𝑖) ) (11)
If phase-synchronization exists between the two channels in the same frequency band then the
difference in phase is equal to a constant:
∆ϕ(𝑡) = ϕ0(t) − ϕ1(t) (12)
This angle is used to create an instantaneous complex vector normalized to the unit circle that is
averaged over N samples (Equation 13). The magnitude of this average vector is used as a measure
of phase locking as described in [70].
𝑃𝐿𝑉(𝑓0(t),𝑓1(t)) =1
𝑁√(∑ (𝑐𝑜𝑠(∆ϕ(𝑡)))
𝑁−1
𝑡=0
2
+ (∑ (𝑠𝑖𝑛(∆ϕ(𝑡)))𝑁−1
𝑡=0
2
(14)
where N defines the number of vectors in the moving-average. If the average ∆ϕ(𝑡) is constant,
the magnitude of the average vector will be 1. If ∆ϕ(𝑡) is changing, the vectors will span the
complex plane and the average will approach the origin, or 0, as shown in Figure 11.
22
Figure 11: Phase locking value mean complex vector. If the phase difference between input signals varies, average
vector shown in red will be low. If synchrony exists, the vectors are more tightly grouped, resulting in a larger
magnitude.
2.2.1 Hardware Architecture
In summary, the PLV algorithm requires the hardware computation of the Hilbert transform,
arctan, addition, sine and cosine, moving-average filtering, square-root, and lastly, the complex
magnitude. The arctan, sine/cosine and magnitude operators will be implemented in hardware
using the CORDIC algorithm as outlined in the following section, while the moving-average
filtering will be computed using digital filtering.
2.2.2 CORDIC
CORDIC (for COordinate Rotation DIgital Computer), also known as Volder's algorithm [73], is
an efficient algorithm to calculate hyperbolic and trigonometric functions, typically converging
with one bit per iteration. The CORDIC algorithm operates on a vector of complex numbers by
multiplying it by powers of two, removing the requirement of complex multipliers and utilizing
only adders, shifters and memory retrieval operations. Using an iterative approach, CORDIC
provides a high-accuracy, low-power and low area VLSI implementation at the cost of reduced
speed. To compute PLV, two CORDIC modes were implemented; the rotational mode which is
used for computing sine and cosine, and the vectoring mode which is used to compute the
magnitude and the phase. The two modes only differ in the directions of rotation. The CORDIC
algorithm involves iterations on three difference equations as shown in (15).
23
𝑓𝑜𝑟 𝑛 = 1: 16
𝑥[𝑛 + 1] = 𝑥[𝑛] + 𝑦[𝑛]2−(𝑛−1)
𝑦[𝑛 + 1] = 𝑦[𝑛] − 𝑠𝑖𝑔𝑛(𝑦[𝑛])𝑥[𝑛]2−(𝑛−1)
𝑧[𝑛 + 1] = 𝑧[𝑛] + 𝑠𝑖𝑔𝑛(𝑦[𝑛])arctan (2−(𝑛−1))
(15)
To compute the magnitude and the phase using CORDIC in vectoring mode, the initial x and y
values are set to represent the real and imaginary components of the signal, respectively, with y
set to 0. A look-up table is used to store 16 arctan values which are used in (15). Over the
proceeding 16 clock cycles, iterations of the algorithm are repeated while y converges to 0. The
final value of x represents a scaled magnitude and z represents the phase.
In the case of computing sine and cosine, CORDIC is used in rotation mode where x is initialized
to the CORDIC aggregate constant K, y is set to 0 and z is set to the desired angle to compute. In
this case z converges to 0, the final value of x represents the cosine of the angle and y represents
the sine of the angle.
2.2.3 Design Optimizations
This implementation of the PLV algorithm offers several advantages over existing solutions [71].
FIR based moving average filters have been replaced by an IIR approximation, resulting in a 60%
decrease in latency. Furthermore, the number of CORDIC cores has been reduced from five to two
by adding system arbitration to support resource reuse (where only two CORDIC blocks are used
at a given time). This optimization results in a silicon area reduction of over 8x as shown in Figure
12, with no impact on latency.
Figure 12: Phase locking value area reduction. Cadence Encounter view showing an 8x reduction in area when
compared to existing work [71].
24
2.3 Cross-frequency Coupling
Growing evidence suggests that cross-frequency coupling (CFC) is a key mechanism in neuronal
computation, communication, and learning in the brain. Abnormal CFC has been implicated in
pathological brain states such as epilepsy and Parkinson’s disease. A reduction in excessive
coupling has been shown in effective neuromodulation treatments, suggesting that CFC may be a
useful feedback measure in closed-loop neural stimulation devices. However, processing latency
limits the responsiveness of such systems. A VLSI architecture is presented which implements
three selectable measures of CFC to enable the application specific trade-off between low-latency
and high-accuracy processing. The architecture is demonstrated using in-vitro human neocortical
slice recordings, with a latency of 48ms.
2.3.1 Introduction
Recent studies have found that cross-frequency coupling (CFC) may play a functional role in
biological information processing in the brain [74]. In particular, phase-amplitude coupling (PAC)
has been hypothesized to provide an effective means to integrate functional systems, transferring
information from large-scale brain networks operating at behavioral timescales, to rapid cortical
processing [75].
It has been theorized that global brain rhythms modulate the excitability of local neural populations
through fluctuations in membrane potentials, increasing the probability of neuronal spiking at a
specific phase of slower rhythms [76]. In electroencephalography (EEG), this mechanism
manifests itself in local field potentials that resemble amplitude modulation in electronic
communication (Figure 12).
Abnormal CFC has been implicated in pathological brain states such as epilepsy and Parkinson’s
disease. CFC between pathological high frequency oscillations (pHFOs) and lower frequency
This section is based on the “Low-latency VLSI architecture for neural cross-frequency coupling analysis,”
manuscript published in the proceedings of IEEE Engineering in Medicine and Biology Conference, 2017. The
majority of this work (90%) was contributed by Gerard O’Leary, under the supervision of Prof. Taufik A.
Valiante2,3 and Prof. Roman Genov1.
1 Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada 2 Krembil Neuroscience Center, Toronto, M5T 2S8, Canada 3 Division of Neurosurgery, Department of Surgery, University of Toronto, Toronto, ON M5T 2S8, Canada
25
rhythms are elevated in the seizure-onset zone compared to healthy regions [77], and exaggerated
PAC has been observed in the primary motor cortex of Parkinson’s Disease patients [78]. A
reduction in excessive coupling has been shown in effective neuromodulation treatments [79]. This
suggests that CFC may be a useful feedback measure for use in closed-loop neural stimulation
devices.
Figure 13: Cross-frequency coupling. Neuronal sub-population activity (red) is modulated by global low-frequency
oscillations (blue).
Depending on the application, a tradeoff must be made between precision and computational
efficiency. In applications utilizing closed-loop processing based on CFC, a low-latency is
required to quickly determine parameters for responsive neural stimulation. Such experiments
performed in-vivo may require the use of implantable microsystems, which must operate under
highly power-constrained conditions. This configurable architecture supports three metrics to
enable the application-specific tradeoff between low-power, low-latency and high-precision.
Furthermore, a key computational overhead routinely involves determining the statistical
significance of a given measure of CFC. This architecture supports this functionality by efficiently
performing surrogate statistical analysis at the expense of increased latency and power-
consumption.
2.3.2 Measuring CFC
Three measures of CFC have been implemented to enable an application specific tradeoff between
sample latency (τs) and precision. The Cross-Frequency Phase Locking Value (CF-PLV), has been
26
proposed to detect the cross-frequency synchrony between a low frequency phase and a high
frequency envelope phase [80]. The CF-PLV does not quantify the relative amplitude of
modulation, and so the Heights Ratio (HR) has been implemented to compliment this measurement
[75]. The Mean Vector Length Modulation Index (MVL-MI) computes the magnitude of an
averaged complex-valued time series, where each sample is comprised of the envelope amplitude
of a modulated high-frequency signal, and the phase of a low-frequency modulating signal [81].
The VLSI implementation of these methods will be discussed in detail in Section V.
2.3.3 System Architecture
The system architecture is comprised of modulation signal extraction, CFC processing, and
surrogate analysis cores as outlined in Figure 14. A VLSI implementation of the architecture is
also shown. All processing blocks share common resources which are time-multiplexed between
processing stages. These include a configurable 512-tap Finite Impulse Response (FIR) filter
supporting both linear-phase bandpass and Hilbert transform functionality, a dual-core CORDIC
block that supports both rotational and vectoring modes in a circular configuration, and a 32-bit
low-power optimized multiply and accumulate block (MAC).
Figure 14: CFC processing system architecture. System uses shared resources in different configurations to enable
the calculation of three selectable metrics. Shown right is the VLSI implementation in Cadence Encounter.
2.3.4 Modulation Signal Extraction
The first stage in calculating the implemented cross-frequency metrics involves extracting the
“phase-modulating” and “amplitude-modulated” signals. These will be referred to as the phase
27
frequency (𝑓𝑝) and amplitude frequency envelope (𝑓𝐴), respectively. Theta modulated high-gamma
oscillations have been observed where fp is in the range 4-8 Hz and the fA carrier is in the range
50-200 Hz [81].
The raw EEG signal, X(t), is bandpass filtered using a linear-phase FIR filter to extract both 𝑓𝑝(𝑡)
and the high-frequency modulated signal. Once the modulated high-gamma component has been
isolated, its amplitude envelope time series, 𝑓𝐴(𝑡) can be extracted (Figure 15). An analytic signal
is first formed by passing the signal through a Hilbert filter, creating both real and imaginary
components. The amplitude of the envelope, 𝑓𝐴(𝑡), can then be extracted by taking the magnitude
of analytic high-gamma signal.
MAG()
CORDIC
MODULATION SIGNAL EXTRACTOR
fA(t)
THETA
FILTER
X(t)
fp(t)
HIGHGAMMA
FILTER
HILBERTFILTER
Z1 -N
I
R
Figure 15: Modulation signal extractor. System block is responsible for extracting the high-gamma envelope (𝑓𝐴)
and the theta modulating signal (𝑓𝑝). Delay matching is required to compensate for filter latency.
2.3.5 CFC Processor Implementation
2.3.5.1 Cross-Frequency Phase Locking Value (CF-PLV)
The goal of the CF-PLV algorithm is to detect the cross-frequency synchrony between the phase
of the low frequency modulating signal (Φ𝑓𝑝(t)), and the phase of the envelope extracted from the
high frequency modulated signal (Φ𝑓𝐴(t)). The phase difference between both signals is calculated
as:
∆Φ(𝑡) = Φ𝑓𝐴(t) − Φ𝑓𝑝(t) (16)
28
This angle is used to create an instantaneous complex vector that is averaged over N samples (17).
The magnitude of this average vector is used as a measure of phase locking as outlined in Section
2.3. If the average ∆Φ𝑖 is 0, both 𝑓𝑝(t) and 𝑓𝐴(t) are phase-locked.
𝑃𝐿𝑉 =1
𝑁√(∑(𝑐𝑜𝑠(∆Φ(𝑡)))
𝑁−1
𝑡=0
2
+ (∑(𝑠𝑖𝑛(∆Φ(𝑡)))
𝑁−1
𝑡=0
2
(17)
The VLSI architecture to compute this measure is shown in Figure 16. The envelope time series is
first filtered using the same parameters as used for 𝑓𝑝. Following this, analytic signals for both
𝑓𝑝(𝑡) and 𝑓𝐴(𝑡) are created using a Hilbert transform. The PLV is then determined for the phase
modulating theta signal and the extracted envelope of the amplitude modulated high-gamma
signal.
ARCTAN SIN(Δϕ)
CORDIC PHASE 1
CORDIC PHASE 2
COS(Δϕ)
Δϕ
1N
∑SIN(Δϕ)
MAG()
CORDIC PHASE 3
PLV (t-τs)
Cross Frequency Phase Locking Value
fP(t) HILBERTFILTER
1N ∑COS(Δϕ)
I 0
0R
ARCTAN fA(t) HILBERT
FILTERI 1
1R
ϕ1
ϕ0
I
R
I
R
FRO
M
MO
DU
LATI
ON
SIG
NA
L EX
TRA
CTO
R
Figure 16: Cross-frequency phase locking value (CF-PLV) architecture. Filters and CORDIC blocks are time
multiplexed to compute the PLV between 𝑓𝑝(𝑡) and 𝑓𝐴(𝑡).
2.3.5.2 The Heights Ratio
A limitation of the CF-PLV measure is highlighted in Figure 17, where both Figure 17 (a) and
Figure 17 (b) result in the same coupling measure independent of the relative amplitude of
modulation.
29
Figure 17: Limitation of CF-PLV. Phase alone does not reflect the intensity of high-frequency modulation, as ΦfA
is the same for (a) and (b).
One method to address this issue involves the use of the Heights Ratio [75] to compliment the CF-
PLV value (Figure 18), where ℎ𝑚𝑎𝑥 and ℎ𝑚𝑖𝑛 are the maximum and minimum heights taken from
the extracted 𝑓𝐴(t) envelope. When enabled, these values are updated for each new sample.
Figure 18: The heights ratio. The heights ratio (HR) is used to quantify the modulation ratio by taking a measure of
the maximum and minimum values of the high frequency envelope.
2.3.5.3 Mean Vector Length Modulation Index (MVL-MI)
The mean vector length modulation index (MVL-MI) [81] computes the magnitude of the average
value of a complex-valued time series (18). Each sample point is comprised of the amplitude of
the modulated high-frequency envelope, and the phase of the low-frequency modulating signal.
𝑧(𝑡) = 𝐴𝐻𝐹𝑒𝑖𝜑𝐿𝐹(𝑡) (18)
30
The time series defined in the complex plane is used to extract a phase-amplitude coupling
measure. Each instantaneous amplitude point is represented by the length of the complex vector,
whereas the modulating signal phase of the time point is represented by the vector angle. The
magnitude of the average complex vector of this time series reflects the raw modulation index.
𝑚𝑟𝑎𝑤 = |𝑧(𝑡)| (19)
In the case of an absence of phase-amplitude coupling, the plot of the time series in the complex
plane is characterized by a roughly uniform circular density of vector points, symmetric around
zero (Figure 19).
Figure 19: MVL-MI complex vector series. A bump at 0 degrees in the complex plane indicates PAC (Left). A
uniform circular density indicates no relationship between phase and amplitude (Right).
If there is modulation of the high-frequency amplitude by the 𝑓𝑝(t) phase, the 𝑓𝐴(t) envelope is
higher at certain phases than others. This higher amplitude for certain angles will lead to a “bump”
in the complex plane plot, leading to loss of symmetry around zero. This loss of symmetry can be
inferred by measuring the length of the average vector of all points in the complex plane. As a lack
of coupling results in a symmetric distribution around zero, the resulting mean vector length is
small. The existence of coupling leads to a non-uniform circular distribution, resulting in a larger
31
mean vector length. The MVL-MI is computed as shown in Figure 20 for the phase modulating
signal and extracted envelope.
Figure 20: Mean vector length modulation index extractor. The MVL-MI has similar processing requirements to
the PLV, and is time multiplexed with a multiplier to scale the vector before being passed to the moving average filter.
2.3.6 Surrogate Analysis
To consider the level of statistical significance of the estimated CFC metrics, a surrogate
distribution is formed by repeating the computation of either MVL-MI or CF-PLV with shuffled
versions of the amplitude signal, 𝑓𝐴(t). This process is accelerated using a random-access circular
sample buffer that is used for inter-sample memory storage.
The random selection of the start point in each shuffle iteration is performed using a linear
feedback shift register (LFSR) as a pseudorandom number generator. This acts as a memory
address generator to access a random start points in the circular buffer, from which the surrogate
is generated.
The mean, μ, and variance (Mean Absolute Deviation) are calculated across the set of values from
each surrogate iteration, which are then used to assess the significance of the CFC metric. 50
permutations can be sufficient [80]; however, this can be increased dynamically, depending on the
required accuracy.
32
2.3.7 Experimental Methods
The functionality of the architecture was verified using data obtained at The Toronto Western
Hospital [82] under Research Ethic's Board Approval. Local field potentials (LFPs) were recorded
simultaneously in superficial (layers II/III) and deep (layers V–VI) layers in 500um thick human
temporal cortical slices using a single glass electrode in each layer filled with a solution containing
150 mM NaCl or a standard artificial cerebrospinal fluid (ACSF) (Figure 21).
Signals were acquired at 10 kHz (low pass at 5 kHz). Recordings were obtained at 36°C in standard
ACSF perfusion (baseline), during kainate (50 nM) applications, and during kainate plus carbachol
(50 μM) conditions.
Figure 21: Slice electrode arrangement for signal acquisition. LFPs were recorded simultaneously in superficial
(layers II/III) and deep (layers V–VI) layers. Image from [82].
For each slice, a single 30 s region displaying the largest power increase during kainate plus
carbachol conditions was analyzed. A sub segment of PAC detected in [82] was used to
demonstrate the functionality of the architecture (Figure 22), and was downsampled to a 1kHz
sample rate.
2.3.8 Results
The accuracy of each metric is evaluated using 16-bit fixed point implementation (Figure 22). Both
CF-PLV and the MVL-MI are sensitive to the intensity of PAC. CF-PLV has a range of 0 to 1 and
MVL-MI is unconstrained.
33
Figure 22: Computed measures of PAC using MVL-MI and CF-PLV. While it is difficult to verify the
instantaneous physiological presence of CFC for validation, simulations of the CF-PLV and MVL-MI show sensitivity
to segments of data which have visually strong coupling between theta and high-gamma oscillations.
The output latency for each reported metric is shown in Figure 23. HR offers the lowest processing
latency with a sample latency equivalent to that of MVL-MI. While the introduction of surrogate
analysis does not impact the sample latency, the processing latency increases almost linearly with
each iteration as the metric is computed for each surrogate. A greater number of CPU cycles
directly increases the power required for a given CFC metric.
34
Figure 23. CFC measurement latency. Processing and sample latency comparison (Left). And the overhead to
perform surrogate analysis (n=50) for a metric (Right).
2.3.9 Discussion
In this section, a VLSI architecture is presented to detect cross-frequency coupling in neural signals
and is validated using in-vitro human slice recordings. The processing latency has been minimized
to enable highly-responsive interaction with neural systems under investigation.
For better time resolution and precision, CF-PLV has been demonstrated. However, this
measurement comes at the expense of increased latency. The MVL-MI offers a good tradeoff
between power and latency, but relies more heavily on surrogate analysis for de-biasing [76]. The
Heights Ratio comes at the lowest computational expense, but is sensitive to noise and certain
amplitude distributions [75]. The outlined VLSI architecture in this section implements these
measures with a power efficiency suitable for implantable devices, enabling responsive closed-
loop in-vivo experiments involving CFC.
35
Chapter 3
Intelligent Closed-Loop Neuromodulation
Devices
3.1 Machine-Learning-Based Seizure Detection
It should be noted that the devices outlined in section 1.5 require sensitive manual tuning of
detection parameter thresholds by a clinician to optimize the accuracy of seizure detection. The
premise for automated brain state classification has been outlined in section 1.3 with the use of
accumulated patient data to train patient-specific seizure detection models which reside on an
implanted device. In this chapter, we integrate the feature extractors outlined in Chapter 2 to inform
a trained model. Biomarker extraction, classification, and chronically-safe neurostimulation is
integrated in an implantable integrated circuit, the Neural Interface Processor (NURIP). Before
describing this device, the nuances of using Support Vector Machines for time series
classifications must be understood, and the solutions to some key challenges will be outlined.
3.2 Neural Signal Time Series Classification
To capture the temporal evolution of machine learning features such as signal energy, conventional
methods use a windowing approach where contiguous time epochs are concatenated to form a
feature vector to be classified. Using this approach, it is possible to learn temporal differences
between windows for events such as seizure onset [83].
Window-based approaches have several limitations in performance and hardware efficiency.
Firstly, processing larger windows requires proportionally large accumulation logic. Secondly, if
classification is performed at every epoch, re-ordering logic may be necessary to remove old
windows and add new windows. Thirdly, the minimum detection latency is the time required to
generate a window (typically multiple seconds). Lastly, as EEG recordings are patient specific,
one window size may give sufficient temporal resolution in one case, but may not be optimal for
36
another. This suggests the need to learn feature timescales in a patient-specific manner to maximize
classification performance.
3.2.1 Exponentially Decaying Memory
Exponentially Decaying Memory (EDM) is an approach which addresses the challenges outlined.
Rather than accumulating and concatenating fixed windows, a continuous sampling recursive
window can be defined by:
𝐸𝐷𝑀(𝑡) = 𝐸𝐷𝑀(𝑡 − 1) − 𝛼 [𝐸𝐷𝑀(𝑡 − 1) − 𝑥(𝑡)] (20)
This approach incorporates new inputs, or degrades existing memory of a feature according to the
decay rate, 𝛼. Where:
𝛼 = 1
𝑁 , 𝑁 = 2𝑖 , 1 < 𝑖 < 16 (21)
The EDM minimizes latency as the output is continuous and can be classified at every sample,
rather than every window. Furthermore, temporal resolution is maximized as accumulation over
an epoch is not required. Most interestingly, the EDM can be implemented efficiently in hardware
using shift and add operations if N is limited to powers-of-two. This efficiency allows multiple
EDMs to be used in parallel, enabling multiple timescales to be processed simultaneously (Figure
24).
Figure 24: EDM Array. Using EDMs in parallel represents the input feature across multiple timescales to form a test
vector for classification.
37
3.3 Arbitrary Waveform Neurostimulation
Three main challenges arise when implementing stimulation strategies for neuromodulation
devices; which stimulation parameters are necessary to induce the desired neural activity, and
hence produce the desired effects? How can we minimize the power required to achieve a given
effect? And how can we ensure our stimulation parameters are safe for chronic use?
It has been shown that arbitrary waveforms can induce complex neural activity. Indeed, this is the
mechanism behind modern cochlear implants [84], where non-biphasic electrical stimulation has
been used to control auditory neurons to convey meaningful information to the brain.
The ability to activate one population of neurons without activating neighboring, but unwanted
populations, is referred to as stimulation selectivity. Certain waveforms have been developed that
allow selectivity during electrical excitation of tissue [35]. Furthermore, the use of combinations
of electrodes can be used to shape the field of the stimulus and increase selectivity [85].
Knowing that control of the stimulus in time and space can affect neural activity, it is desirable to
increase the resolution of these parameters to enable the investigation of more optimal stimulation
strategies. More fine-grained control could enable a tradeoff between increased selectivity,
reduced power consumption and waveform safety [86]. This can be achieved with the use of
arbitrary waveforms which can have any value at a given point in time and are not generated based
on a periodic function.
Arbitrary stimulation waveforms can be produced by manually setting the intensity of the stimulus
on a sample by sample basis, enabling more flexibility for future investigations. However, the
generation of such waveforms can easily violate the charge balancing constraints outlined in
Section 1.4.4. The device in the following section outlines an approach to enable the functionality
outlined above, while ensuring the safety of the stimulation protocol.
38
3.4 NURIP: The Neural Interface Processor
The advancement of closed-loop neuromodulation for treating neurological disorders demands:
(1) analog circuits monitoring the brain activity uninterruptedly even during neurostimulation, (2)
energy-efficient high-efficacy processors for responsive, adaptive, personalized neurostimulation,
and (3) safe neurostimulation paradigms with rich spatio-temporal stimuli for controlling the
brain’s complex dynamics. This section presents an implantable neural interface processor
(NURIP) that addresses these requirements - it performs brain state classification for reliable
seizure prediction and contingent seizure abortion.
3.4.1 Architecture Overview
As shown in Figure 25 NURIP includes 32 bidirectional channels, each with an arbitrary-
waveform neurostimulator and an input-tracking ΔΣ-based analog-to-digital converter. The
recording channel automatically detects any sharp transitions in the intracranial
electroencephalogram (iEEG), such as those due to a stimulation artifact, and shifts its high-
resolution input range to zoom to the input signal, anywhere within the power rails, as depicted in
Figure 25 (top, left). Compared to conventional amplifiers, it experiences no blind intervals caused
by sharp input transitions. The input digital stage features an autoencoder neural network for both
iEEG spatial filtering and dimensionality reduction. Dedicated feature extraction blocks are
implemented for univariate (signal-band energy, SE) and multivariate (phase locking value, PLV,
and cross frequency coupling, CFC) neural signal processing. The proceeding support vector
machine (SVM) accelerator employs these features for brain state classification. A general-
purpose CPU facilitates additional custom feature extraction and system control. In response to
This section is based on “A Recursive-Memory Brain State Classifier with 32-Channel Track-and-Zoom Δ2Σ
ADCs and Charge-Balanced Programmable-Waveform Neurostimulators,“ a manuscript to be presented at the
IEEE International Solid State Circuits Conference, 2018. This work was submitted with an 80% contribution by
Gerard O’Leary, in collaboration with M. Reza Pazhouhandeh1, (responsible for the design of the track-and-zoom
channel array), Michael Chang2, Dr. David Groppe2, Prof. Taufik Valiante2,3, Prof. Naveen Verma4 (SVM kernel
accelerator) and Prof. Roman Genov1. 1 Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada 2 Krembil Neuroscience Center, Toronto, M5T 2S8, Canada 3 Division of Neurosurgery, Department of Surgery, University of Toronto, Toronto, ON M5T 2S8, Canada 4 Princeton University, Princeton, NJ 08544, USA
39
the detection of a pathological brain state, an appropriate modulation waveform is generated to
control the operation of the current-mode neurostimulator.
Figure 25: NURIP architecture for responsive neuromodulation. The processing flow record, preprocess, analyze,
classify, and respond to neural signals is illustrated (the 32-channel array was designed by M.R. Pazhouhandeh).
3.4.2 Data Preprocessing
Figure 26 illustrates both spatial and spectral filtering of the input neural signals. An auto-encoder
neural network performs spatial filtering and dimensionality reduction from 32 recording channels
to 4 weighted combinations (such as in principal component analysis), reducing the processing
requirements by 8x. Sixteen hardware-based circular buffers (including the shown 4) enable online
processing of neural recording streams, with a 256-sample window. They are mapped to 8 kB of
address space within 64 kB of global SRAM. Incoming samples are mapped to varying physical
addresses whereas the corresponding virtual addresses used by the system are fixed. The output
stream is band-pass filtered using a global configurable FIR filter which utilizes coefficients
symmetry to halve the number of MACs.
40
Figure 26: Preprocessor and signal band filtering. The autoencoder neural network enables spatial filtering and
data dimensionality reduction. The circular buffer enables the real-time data management of incoming neural signal
streams. The configurable order FIR filtering block reuses system resources to isolate signal bands of interest.
3.4.3 Feature Extraction
A subsequent array of three configurable neural signal feature extractors, shown in Figure 27 (top),
enables custom patient-specific processing to maximize classifier performance. The absolute
output value of each bandpass filter is taken as a measure of signal energy. As shown in Figure 27
(top, left), a specific power signature in the δ, θ, α, β and γ iEEG bands is indicative of a seizure.
The phase locking value (PLV) extractor shown in Figure 27 (top, middle) detects phase difference
precursors of an upcoming seizure onset. An analytic signal is obtained using a global Hilbert FIR
filter along with the dual-core CORDIC block to extract the phase difference between two input
channels. Efficient resource reuse results in an overall area reduction of 9x versus the state of the
art [87]. Cross-frequency coupling (CFC) is a key mechanism in neuronal computation and its
abnormal appearance can serve as a key spatial biomarker for seizure detection [88]. Low-
frequency brain rhythms modulate high-frequency activity and the resulting envelope is extracted
with re-use of PLV hardware. CFC is then computed as the synchrony between the extracted
envelope and a low-frequency modulating signal. The ensemble of these three biomarkers yields
a uniquely high-dimensional feature space for the classifier.
41
Figure 27: Feature extraction and EDM-SVM classification. The feature extraction cores outlined in Chapter 2 are
integrated with the exponentially decaying memory SVM (EDM-SVM) to enable high-dimensional classification of
brain states.
3.4.4 Classification
In the case of seizure prediction, onset biomarkers are subtle and can occur minutes before seizure
onset. This presents a challenge in processing and memory requirements for implantable devices.
NURIP introduces the Exponentially Decaying-Memory Support Vector Machine (EDM-SVM)
accelerator for efficient classification of long-term temporal patterns. The EDM-SVM input stage
(Figure 27 (bottom, right)) recursively captures a feature’s history across multiple timescales, up
to multiple minutes, using a combination of memory decay rates to enable the learning of temporal
relationships. An efficient implementation using shift and add operations is implemented by
constraining decay coefficients to powers-of-two. The proceeding SVM accelerator core allows
the selection of linear, polynomial and radial-basis function (RBF) kernels to tradeoff between
performance, energy and memory usage [24]. As the EDM is updated every sample, classification
can be performed continuously to minimize detection latency.
42
3.4.5 Responsive Neurostimulation
Upon the detection of a seizure, the integrated digitally charge-balanced neurostimulation
waveform generator responds as demonstrated in Figure 28. To prevent electrode and tissue
damage due to stimulus charge imbalance, binary exponential charge recovery (BECR) ensures
the net stimulus integral is zero after arbitrary waveform stimulation or when safe limits have been
exceeded. Function generation with a sample rate of 15KHz is efficiently implemented with the
reuse of the dual-core CORDIC and MAC blocks, enabling the generation of sums or products of
sinusoids. Arbitrary waveform replay at 3MHz is supported by streaming samples from on-chip
SRAM.
Figure 28: Charge-balanced neuromodulation architecture. Experimentally measured arbitrary waveform
generation reuses system CORDIC blocks. Digital charge balancing is enabled with the continuous monitoring of
stimulus amplitudes sent to the current-mode stimulators. The stimulus waveform generator output was recorded with
the assistance of Shahryar Rajabzadeh.
43
3.4.6 Results
Figure 29 outlines NURIP’s classifier performance using the EU intracranial EEG database [89].
The extracted feature space used for classification consists of 125 dimensions derived through
offline feature selection and is constrained to <200 support vectors by the on-chip SRAM. A
sensitivity of 100% and a false positive rate (FPR) of 0.81 per hour have been achieved. A
classification rate of 4 Hz requires a power consumption of 674.4 µW with a nominal voltage of
1.2v and an operational frequency of 10 MHz. The SoC micrograph and the channel floorplan are
shown in Figure 29 (top). The chip is compared with the state of the art both in terms of the channel
performance and digital processing performance in Figure 29 (bottom).
Figure 29: EDM-SVM output and NURIP power consumption. Experimentally computed feature space and EDM-
SVM classifier output in offline human iEEG recordings is shown. A 125-dimensional feature space is extracted from
8 iEEG channels. The classifier output closely aligns with expert labels. Power consumption is measured during
continuous feature extraction and 4 Hz classification. The power measurements were recorded with the assistance of
Shahryar Rajabzadeh.
44
3.4.7 Chip Micrograph and Comparison
Figure 30 shows the device floorplan and a comparison table for the digital processing system.
NURIP offers a large array of feature extraction capabilities, along with effective EDM-SVM
based classification and flexible, charge-balanced neuromodulation waveform generation.
JSSC’13 [24] JSSC’14 [90] JSSC’15 [3] ISSCC’17 [91] THIS WORK
TECHNOLOGY (µm) - 0.18 0.18 0.13 0.13
FEATURE
EXTRACTION CPU
FFT,
Entropy SE PLV
PLV, CFC,
SE, CPU
CLASSIFIER SVM LLS D2A-LSVM Threshold EDM-SVM
SAMPLE MEMORY 6s 96 samples 3s 64 samples ∞ (EDM)
LATENCY (s) 2 0.8 1 - <0.1
MEMORY (kB) 64 0 64 0 96
ENERGY/CLASS. (µJ) 273 77.91 2.73 - 168.6
WAVEFORM GEN. - Bi-phasic Bi-phasic - AWG
CHARGE BALANCING - - PVTES - BECR
Figure 30: NURIP SoC micrograph and comparison table. The neural recording and stimulator array was
implemented by M.R. Pazhouhandeh.
45
Chapter 4
Continuous Lifelong Device Personalization
4.1 The Need for Chronic Neural Recordings
Intelligent neurostimulation devices require recorded patient data to configure seizure detection
algorithms and stimulation parameters for a given patient. The initial training of such devices relies
on neural recordings collected over a short period during a patient’s stay in an EMU as outlined in
Section 1.2.2. However, the recordings collected during the time is limited as just 3 to 9 seizures
are typically sufficient for the purpose of seizure localization [92].
There is evidence to suggest that EMU recordings are not optimal for the purpose of training
patient-specific machine learning models. Firstly, it has been shown that chronic implants
experience considerable recording signal variability over time [93]. Furthermore, some evidence
suggests that the recordings obtained in the EMU may not be representative of the real-world
measurements of chronic recording systems. For example, in the context of cognitive science, there
is a growing body of evidence to suggest that the use of laboratory-based experiments are less
ecologically valid than real-world behavior [94]. It is clear that devices must adapt to
environmental changes in the electrode-tissue interface, and the dynamic neural activity in the
patient’s real-world environment. The continuous refinement of device parameters and machine
learning models over the lifetime of an individual is essential to maximize classification accuracy
and treatment efficacy.
The challenge of continuous treatment refinement has been recognized in the development of
existing devices, and a number of solutions have been implemented as will be shown in the
following section. In the case of machine-learning based closed-loop systems such as NURIP,
more complex systems are necessary for model adaptation and the final section in this chapter will
outline the development of a patient-localized device training system to achieve this goal.
46
4.2 Existing Long-Term Monitoring Solutions
The following devices have been developed as support tools for commercially available
neurostimulation devices. They allow for basic communication with the implant to allow the
wireless configuration of stimulation parameters, battery level monitoring, and batch-mode data
acquisition. It should be noted that no existing clinically approved devices support continuous
long-term wireless streaming of recorded neural signals. Their main utility is to enable the manual
adjustment of the device to optimize its efficacy during the post-operative care of the patient.
4.2.1 NeuroPace Remote Monitor
The remote monitor (Figure 31) is a home-use monitoring device used to retrieve data stored on
the NeuroPace device. The patient data management system (PDMS) is a centralized database,
which contains data uploaded from the programmer and remote monitor. Neurostimulator data and
detection settings can also be transferred from PDMS to the programmer [55] .
Figure 31: Neuropace remote monitor. The NeuroPace remote monitor enables external access to the implanted
device. It is used to upload device settings and wirelessly receive device data.
4.2.2 Medtronic Patient Programmer
The Medtronic patient programmer is used to control and monitor the implanted neurostimulator
[95]. It allows the user to Check the neurostimulator battery status and provide alerts, and allows
the configuration of stimulation settings. However, it does not support the ability to stream data
remotely.
47
Figure 32: Medtronic DBS device programmer. The patient programmer is Medtronic’s solution to interacting with
implanted neurostimulation devices. Image from [95].
4.2.3 Medtronic Nexus
The Medtronic Nexus Systems are intended to serve as an investigational algorithmic development
tool for first principled approaches to closed loop systems [96]. The Nexus is a bi-directional data
port that interacts with devices such as the Activa PC+S outlined in Chapter 1. It transmits real-
time sensing data and allows a host computer to update the neurostimulator’s stimulation
parameters based on real-time analysis of the sensor data by the host computer. All decisions
regarding stimulation updates are made by a host computer using algorithm prototyping
environments such as Matlab or C#.
This functionality provides the ability for research sites to explore potential closed-loop therapy
algorithms in a flexible manner to assess the feasibility of closed loop therapy adjustment. The
Nexus system allows for firmware upgrades to the system for algorithms that are shown promising
using the Nexus. A closed-loop application which utilizes this system is outlined in [97].
This platform could theoretically support the remote monitoring and configuration of an implanted
closed-loop device, but there have been no reports of the system being used for clinical purposes.
48
4.3 Machine Learning Microserver for Neuromodulation Device
Training
The following section addresses the need for lifelong learning in personalized biomedical devices.
A patient-localized microserver is presented which enables continuous model adaptation for
intelligent neuromodulation devices aided by unsupervised machine learning. The system employs
a one-class support vector machine (OC-SVM) to identify irregular neural activity for remote
clinical assessment and device re-training. The system performance is demonstrated using 500
hours of human intracranial EEG (iEEG) where a clinical seizure detection rate of 97.05% is
achieved. As chronic recording implants become more prevalent, the accumulation of larger
volumes of ictal data using this method could enhance the performance of supervised learning
classifiers, improving the treatment efficacy and the quality of life for patients.
4.3.1 Introduction
The limitations of data collected in the EMU has been highlighted in section 4.1 The acquisition
of further data in the patient’s home environment necessitates the ability to highlight electrographic
events remotely and autonomously to be used for analysis and device training.
As outlined in Section 3.1 distinguishing seizure activity from normal brain activity can be a
difficult task because of potentially great variation in seizure morphology [98]. Machine learning
enables the utilization of large volumes of patient recordings to accurately distinguish pathological
from physiological activity. This approach has led to the introduction of responsive closed-loop
neuromodulation devices which proactively detect and inhibit the onset of seizures [44][91]. Such
systems generally utilize supervised learning models to maintain low false-detection rates for
improved power efficiency and reduced side-effects. However, the use of supervised learning
classifiers for seizure detection exposes a class imbalance problem which arises from a lack of
ictal recordings compared to large volumes of inter-ictal data. Furthermore, supervised
The following section is based on the “Machine Learning Microserver for Neuromodulation Device Training”
manuscript from the proceedings of the IEEE Biomedical Circuits and Systems Conference, 2017. This work was
submitted with co-authors Gerard O’Leary (80% contribution), Asish O. Abraham1, Akshay K. Kamath1, Dr. David
Groppe2, Prof. Taufik A. Valiante2,3 and Prof. Roman Genov1. 1 Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S 3G4, Canada 3 Division of Neurosurgery, Department of Surgery, University of Toronto, Toronto, ON M5T 2S8, Canada 2 Krembil Neuroscience Center, Toronto, M5T 2S8, Canada
49
classification systems require accurate data labeling, and so are vulnerable to the human error in
annotating complex EEG recordings [99].
As outlined in section 1.2.2, EEG recordings are typically used by epileptologists to categorize
irregular neural recordings for seizure localization purposes. Such events include electrographic
seizure onsets and inter-ictal discharges (IIDs). This assessment is generally based on temporo-
spectral changes such as low-voltage fast activity in iEEG seizure onset [100]. To capture these
changes, the Exponentially Decaying Memory (EDM) approach outlined in Section 3.2.1 can serve
as a hardware efficient mechanism to represent temporal feature characteristics for machine
learning.
The proposed system employs an unsupervised learning-based One-class Support Vector Machine
(OC-SVM) [101]. This approach navigates the class imbalance and data labeling challenges by
learning to distinguish normal neural activity from segments of clinical interest. The irregular
recording periods indicated by the OC-SVM can be reviewed remotely by an epileptologist,
enabling ictal data to be labelled and accumulated over time. With increasing volumes of data,
specialized supervised learning classifiers can be trained more effectively for closed-loop
applications.
Chronic neural recording implants experience considerable signal variability over time [93],
leading to a gradual degradation of classifier performance. Thus, continuous model re-training is
necessary to adapt to changing physiological recording conditions and maximize the treatment
efficacy. This is impractical to perform on an implantable device as power-consumption is a
primary consideration to reduce both heat dissipation and the risks associated with battery
replacement surgery.
The proposed solution is a patient-localized microserver which communicates with an implanted
device to enable incremental training (Figure 33). Data recorded by the device is sent to the server
and processed by an FPGA-accelerated OC-SVM. iEEG segments which are considered irregular
are archived and sent to a epileptologist for remote review. Once an assessment is made, the
microserver re-trains the model to be uploaded to the implanted device.
50
Figure 33: Training microserver concept. The microserver system sends irregular EEG events for clinical
assessment and re-trains implanted device.
4.3.2 System Overview
The system shown in Figure 34 is implemented using a Xilinx Zynq SoC. A dedicated dual-core
CPU hosts an on-chip Linux operating system (OS) which runs in parallel with the FPGA fabric.
SVM training is performed on the microserver with encrypted patient data maintained on USB
storage. A TCP/IP implementation allows data to be streamed wirelessly from a compatible
neuromodulation device. To reduce the implanted device’s power requirements for data
transmission, raw samples are sent rather than higher dimensional features. This requires feature
extraction to be replicated on the microserver. Communication with a remote EEG analyst is
supported via an ethernet network interface. Feature extraction and machine learning accelerators
are implemented on the FPGA fabric.
Figure 34: Training microserver implementation overview. A Linux OS enables external communication, SVM
training and access to the FPGA fabric’s feature extraction and classification accelerators.
51
4.3.3 Feature Extraction
As outlined in Section 2.2, spectral energy in conventional physiological signal bands can be used
to label electrographic events [67]. Signal bands of interest are extracted by passing recorded
samples through parallel bandpass filters for Delta (<4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta
(13-30 Hz) and Gamma (30-60 Hz) bands. A 256-tap Type-1 FIR filter is implemented for each
band with a symmetric impulse response, allowing coefficient multiplications to be shared. Each
iEEG channel is processed sequentially and filter states are stored in FPGA block RAM (BRAM)
between sample processing. For each band, the absolute value of each output sample is taken as a
measure of signal energy. This approximation of instantaneous energy is accumulated over a time
window to generate a temporo-spectral measure of the signal.
After the signal energy for a given EEG band is extracted, its value is passed to a corresponding
bank of EDMs. Each EDM implements a different decay rate, 𝛼, complementing one another by
offering a different temporal perspective of the input feature to be used for classification (Figure
35). Small values of 𝛼 result in longer-term memory, while larger values capture finer time
resolutions.
Figure 35: EDM decay rate. The decay rate, 𝛼, controls the retention of input activity. N shown here is the number
of EDM binary shifts, and therefore 𝛼 = 1/N. Parallel EDMs give different time perspectives of the input, enabling a
classifier to learn temporal relationships.
52
4.3.4 One-Class Support Vector Machine
The support vector machine (SVM) is most commonly used as supervised learning model for
classification tasks of two or more classes, however, unsupervised variants have been proposed
[8]. The original algorithm requires a similar number of examples in each class to prevent classifier
bias. In the case of seizure detection, ictal activity is rare and accurate classification is critical to
prevent the onset of a seizure.
The one-class SVM [8] has been proposed as a method for datasets with such class imbalances. It
can be viewed as a regular two-class SVM where the training data is taken as one class, and the
origin is taken as the only member of the second class. Training is performed without labels, where
data is mapped to the kernel space and separated from the origin by a hyperplane with maximum
margin. To classify an input feature vector, a decision function is evaluated to distinguish an inlier
(𝑓(𝑥) > 0), from an outlier (𝑓(𝑥) < 0):
𝑓(�⃗�) = sgn (∑(𝑎𝑖K(𝑠𝑣⃗⃗⃗⃗⃗𝑖, �⃗�) − 𝑏)
𝑁
𝑖=1
) (22)
Where 𝑠𝑣𝑖 are the support vectors used to construct the hyperplane, 𝑎𝑖 are the corresponding
weights, b is the classifier bias term, and K is implemented here as the Radial Basis Function
(RBF) kernel, defined as:
K(𝑠𝑣⃗⃗⃗⃗⃗, �⃗�) = 𝑒−𝛾‖𝑠𝑣⃗⃗⃗⃗⃗ − �⃗�‖2 (23)
This concept is illustrated in Figure 36, where the model is trained using normal physiological
activity (blue). The extracted test vectors classified as outliers (red) could indicate anomalous
activity such as inter-ictal discharges (IIDs) or subclinical seizures.
53
Figure 36: One-class support vector machine. The OC-SVM enables the detection of outlier activity (red) following
training using inter-ictal data (blue).
4.3.5 Hardware Implementation
The microserver prototype (Figure 37) is implemented on a Xilinx Zynq XC7Z045, which contains
dual-core ARM Cortex-A9 CPUs alongside programmable logic (PL) FPGA fabric. A Petalinux
OS is used to host a TCP/IP server for remote communication with a clinician, and local
communication with the implanted device for data retrieval and model uploading. The OS is also
used to host a LibSVM implementation [30] to perform SVM training, to encrypt neural
recordings, and for external data storage.
Figure 37: Prototype training microserver. The prototype shown was developed with the Xilinx ZC-706. The
Petalinux OS was implemented with the assistance of Akshay Kamath and Asish Abraham.
54
Feature extraction, real-time feature normalization, and RBF SVM classification are implemented
on the FPGA fabric using custom accelerators which are accessible from the OS via an AXI-4
interface. Dual-port BRAMs are shared between the OS and the PL to enable online access to
FPGA memories for SVM model coefficients, feature extraction coefficients, and test vectors.
Samples are streamed from the TCP/IP server to a PL BRAM which are processed in parallel by
the feature extractors and classified by the SVM accelerator. If a sample is classified as an outlier,
an interrupt is sent to the OS to encrypt and store the surrounding 10 minutes of data. This enables
a remote clinician to access and annotate the irregular segments and re-train the implant SVM. The
OC-SVM can be retrained and uploaded to the PL to refine the outlier detection segments and
reduce the volume of data to be sent to the analyst for review.
The key design constraints for the microserver include the completion of feature extraction and
anomaly classification with a speed at least equal to the devices recording sample rate, while
minimizing power consumption. The processing latency for feature extraction and classification is
1.259 ms and 0.189 ms respectively, enabling a maximum sample rate of 690 Hz for 16 recording
channels. The dynamic power consumption for the system was estimated with Xilinx Vivado as
0.698 W at a sample rate of 256 Hz. However, due to the low FPGA fabric utilization in the system
prototype (Table 2) a smaller, lower-power device could be utilized for a portable system
deployment.
Table 2: FPGA Resource Utilization
XC7Z045
Resource
Feature
Extractor RBF SVM Used* Available %
BRAMs 42 18 60 1090 5.5
DSPs 16 46 62 900 6.8
Flip-
Flops 2262 3863 6805 437200 1.6
LUTs 2960 7400 10360 218600 4.7
* for 16 Channels, 5 frequency bands and 5 EDM decay coefficients.
4.3.6 Training Method
System functionality is demonstrated using the EU intracranial EEG epilepsy database with expert-
annotated clinical and subclinical seizure events [89]. Patients were selected based on a
postoperative outcome of Engel class I, indicating that intracranial electrodes were within the
55
epileptogenic zone. After the first 24 hours of neural recordings are accumulated, feature extraction
is performed to generate the initial training set. Expert-labelled subclinical and clinical seizure
events are then removed along with the surrounding 10 minutes of recordings. The OC-SVM
model is trained and stored on the FPGA fabric along with feature normalization coefficients used
for the training data.
Minimizing the misclassification of normal physiological neural activity while ensuring that all
pathological activity is captured is a key consideration. To enable this tradeoff, the classifier output
is smoothed using a moving average window which can be increased at the expense of detection
latency.
Once highlighted activity has been annotated by a remote clinician, a refined supervised model
can be trained on the microserver to be uploaded to the implanted device. SVM training is
performed on the Zynq SoC’s dual-core CPUs using a LibSVM implementation. The required
training time scales linearly with the number of features used on the implanted device, and the
FPGA fabric (Figure 38). The number of training vectors is constrained by external memory to
50,000 with a dimensionality of 400. Incremental training can be performed to enable the use of
larger volumes of data.
Figure 38: SVM training time. Performance was measured on dual-core ARM Cortex-A9 CPUs. The training time
was measured with the assistance of Asish Abraham.
56
4.3.7 Results
The performance of the system is validated using 500 hours of iEEG data across four subjects in
the expert-labelled EU epilepsy database. A combination of 16 depth and surface electrodes was
determined on a per patient basis based on proximity to the seizure onset zone. The feature
extraction implementation uses five spectral bands per channel, each with 𝛼 decay coefficients, of
4, 6, 8, 10, 12, 14 and 16. The resulting feature vector has a dimensionality of 560. An illustration
of the feature space for an electrode placed in the seizure onset zone is shown in Figure 39 (b).
The resulting OC-SVM output is shown in Figure 39 (c) where the detected outlier activity
corresponds to an expert labelled clinical seizure onset time segment.
Figure 39: Microserver processing example. Recorded neural signals (top), extracted features (middle) and the
classifier output (bottom), where red indicates the expert-labels and blue is the output calculated on the FPGA fabric.
The OC-SVM output was measured with the assistance of Asish Abraham.
The system performance is outlined in Table 3, where a seizure detection rate of 97.05% is
achieved. In the context of iEEG anomaly detection, the system performance has been optimized
57
to capture seizure-like events at the expense of a high detection rate. The alarms per hour is
reported using forward chaining validation with a moving average smoothing window length
determined on a per-patient basis. These alarms could indicate unlabeled ictal activity and so are
not considered false positives as in the case of supervised learning performance analysis.
Table 3: EU Database System Performance
Patient EU1096 EU442 EU548 EU1125
Hours Analyzed 147 118 129 108
Clinical Seizures 7 8 17 12
Seizures Detected 7 8 15 12
Alarms Per Hour 0.81 1.21 1.47 1.33
Detection Rate 100% 100% 88.2% 100%
58
Chapter 5
Conclusion and Future Directions
This thesis presented the design and implementation of a device-based framework for chronically-
implanted, intelligent, closed-loop neural signal recording and responsive stimulation. The process
of extracting biomarkers from recorded neural signals in low-power devices was outlined, and
three effective approaches were implemented: signal band energy, phase locking, and cross-
frequency coupling. This array of biomarkers supports the classification of a wide array of normal
physiological and pathological brain states. An effective and efficient solution to performing this
time series classification in an implanted device was presented in the form of the exponentially
decaying memory support vector machine. The use of responsive neuromodulation upon the
detection of a pathological state was outlined, and a safe approach to more precise stimulation
using arbitrary waveforms was presented. These concepts were integrated on a single SoC in the
form of NURIP, the neural interface processor. The path towards adapting machine learning based
devices over the lifetime of a patient was outlined and a device-localized microserver solution was
presented. Such developments are necessary to improve the efficacy of existing closed-loop
devices to maximize the quality of treatment for patients.
5.1 Contributions and Relevant Publications
In Chapter 2, the signal processing required to extract pathological biomarkers was described,
improvements upon state of the art phase-based feature extractors were outlined, and a novel
implementation of the cross-frequency coupling biomarker was described. This work was
presented at the 2017 IEEE Engineering in Medicine and Biology Conference (EMBC).
Chapter 3 outlined the process of classifying brain states based on the above biomarkers, and a
new time-series classification approach, the Exponentially Decaying Memory Support Vector
Machine (EDM-SVM), was introduced. Biomarker extraction, classification, and chronically-safe
neurostimulation was integrated in an implantable integrated circuit, the Neural Interface
59
Processor (NURIP). This work has been accepted to be presented at the 2018 IEEE International
Solid-State Circuits Conference (ISSCC).
In Chapter 4, a machine learning microserver is presented to enable continuous post-implantation
adaptation for personalized seizure-control neuromodulation devices. The system demonstrates
the efficacy of OC-SVMs to assist in the labelling of complex iEEG recordings for training
supervised learning models. The concept of the patient-localized microserver addresses the need
for life-long learning in personalized biomedical devices. As chronic recording implants become
more prevalent, the accumulation of larger volumes of ictal data using this method could enhance
the performance of supervised learning classifiers, improving the treatment efficacy and the quality
of life for patients. This work was presented at the 2017 IEEE Biomedical Circuits and Systems
Conference (BIOCAS).
5.2 Future Directions
The neuromodulation devices currently used for the treatment of neurological disorders are still in
very early stages of development. With the recent introduction of closed-loop systems, the move
towards more intelligent devices seems imminent. Future directions include the clinical validation
of existing devices such as NURIP to encourage the mainstream adoption of data-driven,
personalized devices. Further research in fundamental neuroscience is needed to find new
biomarkers which may more informative for device-based classifiers. The mechanisms behind
neuromodulation are still unfolding, with many promising technologies emerging which could be
supported with implanted devices.
5.2.1 Clinical Trial
A clinical trial of NURIP has received research ethics board approval at the Toronto Western
Hospital. This trial will involve connecting the digital processing core of the SoC to clinically
approved neural recording systems (Neuralynx, inc.) and stimulation systems (Tucker-Davis
Technologies) as shown in Figure 40 and Figure 41.
60
Figure 40: Proposed clinical trial setup. NURIP’s planned integration with clinically approved recording and
stimulation systems.
The objective of this study is to investigate the abortion of seizure generation using minimal brain
stimulation on 50 epilepsy patients undergoing standard intracranial EEG recordings to determine
surgical candidacy for resective surgery (see Section 1.2.2) over the course of 2 years. To evaluate
the feasibility of this intervention, the outcome to be measured will be the reduction in clinically
and electrographically manifested seizures upon closed-loop neurostimulation. Considering the
results of this clinical trial, the NURIP design will be optimized to improve treatment efficacy and
efficiency with the aim of commercializing the device for clinical use.
Figure 41: Implementation of NURIP. Shown here is an early stage setup of NURIP with clinical neural recording
systems (Neuralynx, inc.) and stimulation systems (Tucker-Davis Technologies).
61
5.2.2 Adaptive Stimulation Parameterization
When a conventional DBS system is implanted, the stimulation parameters must be manually
defined. The procedure in which the optimum stimulation parameters are defined is a trial-and-
error-based programming task that is carried out by clinicians on a per-patient basis. Using a set
of guidelines, the stimulation parameters are changed based on the observable behavioral
responses [102]. This is error prone, and the expanding configurability of neural stimulators such
as the one outlined in Chapter 3 means that manual exploration of the parameter space by a
clinician is unlikely to result in an optimal stimulation protocol. Furthermore, relying on the
manual setting of stimulation parameters consumes valuable and limited clinician time.
The next generation of DBS devices should be automatically programmable, compatible with
biomarker variations, and flexible in stimulation type and pattern, to yield greater benefits and
fewer side effects [103]. Initial in-vitro work in the field of reinforcement learning has provided
promising results, and a successful methodology could be implemented in an implanted device.
5.2.3 Additional Biomarkers
Additional biomarkers other than those outlined in Chapter 2 could lead to better classification
performance, and earlier detection of seizures. The investigation of the use of interictal discharges
for classification purposes could yield more accurate performance [104]. Furthermore, Results
from studies in animals with epilepsy and presurgical patients have consistently found a strong
association between high-frequency oscillations (HFOs) and epileptogenic brain tissue that
suggest they could be a potential biomarker of seizure genesis [105]. A careful tradeoff must be
made between the energy required to compute these features on chip, and the performance
improvement they provide.
5.2.4 Towards Seizure Prediction
The work outlined in this thesis mainly focuses on the early detection of seizures. The approach
involves learning based on data labels defined by clinicians to detect a seizure as early as possible
after the defined onset. However, recent work has shown that seizures may be detected with a
horizon of many hours [106]. The applicability of this work to implantable devices in practice
should be the subject of future investigation. Furthermore, initial reports have emphasized the need
for chronic recordings for this purpose as outlined in Chapter 4 [107].
62
5.2.5 Alternative Stimulation Modalities
Although this work has focused on the use of electrical stimulation, other promising brain
stimulation methods are emerging such as optogenetics which could revolutionize the level of
control available for neuromodulation. Optogenetics is a relatively new neuromodulation
technology in which light-sensitive proteins (“opsins”) are genetically inserted into cell
membranes to enable their activation using light [108].
Room still exists to greatly advance the utility of electrical stimulation, with a recent report
indicating that DBS may be achieved using noninvasive techniques [109]. Transcranial stimulation
using multiple electric fields at frequencies too high to recruit neural firing, can utilize wave
interference to produce a prominent electric field envelope modulated at the difference frequency.
The use of transcranial methods would remove the constraints associated with implanted devices,
and greatly improve the quality of treatment for those with epilepsy.
63
References
[1] R. Caton, The Electric Currents of the Brain. Taylor & Francis, 1970.
[2] E. D. Adrian, “The Basis of Sensation,” Br. Med. J., vol. 1, no. 4857, pp. 287–290, Feb. 1954.
[3] F. Scholz, Electroanalytical Methods: Guide to Experiments and Applications. Springer Science &
Business Media, 2009.
[4] K. A. Ng, E. Greenwald, Y. P. Xu, and N. V. Thakor, “Implantable neurotechnologies: a review of
integrated circuit neural amplifiers,” Med. Biol. Eng. Comput., vol. 54, no. 1, pp. 45–62, Jan. 2016.
[5] C. Pollo, M. Shoaran, Y. Leblebici, A. Mercanzini, C. Dehollain, and A. Schmid, “The Future of
Intracranial EEG Recording in Epilepsy: a Technological Issue?,” Epileptologie, vol. 29, no. EPFL-
ARTICLE-182900, pp. 114–119, 2012.
[6] G. Buzsáki, C. A. Anastassiou, and C. Koch, “The origin of extracellular fields and currents — EEG,
ECoG, LFP and spikes,” Nat. Rev. Neurosci., vol. 13, no. 6, pp. 407–420, May 2012.
[7] M. Tudor, L. Tudor, and K. I. Tudor, “[Hans Berger (1873-1941)--the history of
electroencephalography],” Acta Medica Croat. Cas. Hravatske Akad. Med. Znan., vol. 59, no. 4, pp. 307–
313, 2005.
[8] G. Buzsáki, Rhythms of the brain. Oxford ; New York: Oxford University Press, 2006.
[9] C. B. Saper, P. M. Fuller, N. P. Pedersen, J. Lu, and T. E. Scammell, “Sleep State Switching,”
Neuron, vol. 68, no. 6, pp. 1023–1042, Dec. 2010.
[10] Institute of Medicine (US) Committee on the Public Health Dimensions of the Epilepsies, Epilepsy
Across the Spectrum: Promoting Health and Understanding. Washington (DC): National Academies Press
(US), 2012.
[11] R. S. Fisher et al., “Epileptic seizures and epilepsy: definitions proposed by the International League
Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE),” Epilepsia, vol. 46, no. 4, pp.
470–472, Apr. 2005.
[12] E. B. Bromfield, J. E. Cavazos, and J. I. Sirven, Eds., An Introduction to Epilepsy. West Hartford
(CT): American Epilepsy Society, 2006.
64
[13] S. Wiebe, W. T. Blume, J. P. Girvin, M. Eliasziw, and Effectiveness and Efficiency of Surgery for
Temporal Lobe Epilepsy Study Group, “A randomized, controlled trial of surgery for temporal-lobe
epilepsy,” N. Engl. J. Med., vol. 345, no. 5, pp. 311–318, Aug. 2001.
[14] N. G. Laxpati, W. S. Kasoff, and R. E. Gross, “Deep Brain Stimulation for the Treatment of
Epilepsy: Circuits, Targets, and Trials,” Neurotherapeutics, vol. 11, no. 3, p. 508, Jul. 2014.
[15] A. K. Shah and S. Mittal, “Invasive electroencephalography monitoring: Indications and presurgical
planning,” Ann. Indian Acad. Neurol., vol. 17, no. Suppl 1, pp. S89–S94, Mar. 2014.
[16] H. Lüders et al., “Semiological seizure classification,” Epilepsia, vol. 39, no. 9, pp. 1006–1013, Sep.
1998.
[17] J. M. Sum and M. J. Morrell, “Predictive value of the first ictal recording in determining localization
of the epileptogenic region by scalp/sphenoidal EEG,” Epilepsia, vol. 36, no. 10, pp. 1033–1040, Oct.
1995.
[18] J. Jankovic, “Parkinson’s disease: clinical features and diagnosis,” J. Neurol. Neurosurg. Psychiatry,
vol. 79, no. 4, pp. 368–376, Apr. 2008.
[19] A. E. Lang and A. M. Lozano, “Parkinson’s disease,” N. Engl. J. Med., vol. 339, no. 15, pp. 1044–
1053, 1998.
[20] M. S. Okun, “Deep-Brain Stimulation for Parkinson’s Disease,” N. Engl. J. Med., vol. 367, no. 16,
pp. 1529–1538, Oct. 2012.
[21] D. Martinez-Ramirez, W. Hu, A. R. Bona, M. S. Okun, and A. W. Shukla, “Update on deep brain
stimulation in Parkinson’s disease,” Transl. Neurodegener., vol. 4, p. 12, Jun. 2015.
[22] Pattern Recognition and Machine Learning | Christopher Bishop | Springer. .
[23] N. Verma, K. H. Lee, and A. Shoeb, “Data-Driven Approaches for Computation in Intelligent
Biomedical Devices: A Case Study of EEG Monitoring for Chronic Seizure Detection,” J. Low Power
Electron. Appl., vol. 1, no. 1, pp. 150–174, Apr. 2011.
[24] K. H. Lee and N. Verma, “A Low-Power Processor With Configurable Embedded Machine-Learning
Accelerators for High-Order and Adaptive Analysis of Medical-Sensor Signals,” IEEE J. Solid-State
Circuits, vol. 48, no. 7, pp. 1625–1637, Jul. 2013.
65
[25] A. Barachant, A. Temko, F. Li, and G. Titericz, “Kaggle Melbourne University
AES/MathWorks/NIH Seizure Prediction challenge - Winning Solution.” .
[26] The Nature of Statistical Learning Theory | Vladimir Vapnik | Springer. .
[27] P. W. Mirowski, Y. LeCun, D. Madhavan, and R. Kuzniecky, “Comparing SVM and convolutional
networks for epileptic seizure prediction from intracranial EEG,” in Machine Learning for Signal
Processing, 2008. MLSP 2008. IEEE Workshop on, 2008, pp. 244–249.
[28] “Support vector machine,” Wikipedia. 30-Dec-2017.
[29] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, no. 3, pp. 273–297,
1995.
[30] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM Trans. Intell.
Syst. Technol., vol. 2, no. 3, pp. 1–27, Apr. 2011.
[31] A. Zisserman, “The SVM Classifier, University of Oxford.” [Online]. Available:
http://www.robots.ox.ac.uk/~az/lectures/ml/lect2.pdf. [Accessed: 03-Jan-2018].
[32] G. Tsoucalas, M. Karamanou, M. Lymperi, V. Gennimata, and G. Androutsos, “The ‘torpedo’ effect
in medicine,” Int. Marit. Health, vol. 65, no. 2, pp. 65–67, 2014.
[33] J. Garson, “The birth of information in the brain: Edgar Adrian and the vacuum tube,” Sci. Context,
vol. 28, no. 1, pp. 31–52, Mar. 2015.
[34] Fritsch and Hitzig, “Uber die elektrische Errebarkeit des Grosshirns,” Arch. Für Anat. Physiol. Wiss.
Med., vol. 37, Nov. 1869.
[35] D. R. Merrill, M. Bikson, and J. G. R. Jefferys, “Electrical stimulation of excitable tissue: design of
efficacious and safe protocols,” J. Neurosci. Methods, vol. 141, no. 2, pp. 171–198, Feb. 2005.
[36] S. F. Cogan, “Neural Stimulation and Recording Electrodes,” Annu. Rev. Biomed. Eng., vol. 10, no.
1, pp. 275–309, 2008.
[37] Z. Feng, Y. Yu, Z. Guo, J. Cao, and D. M. Durand, “High frequency stimulation extends the
refractory period and generates axonal block in the rat hippocampus,” Brain Stimulat., vol. 7, no. 5, pp.
680–689, 2014.
66
[38] C. Magarinos-Ascone, J. H. Pazo, O. Macadar, and W. Buno, “High-frequency stimulation of the
subthalamic nucleus silences subthalamic neurons: a possible cellular mechanism in Parkinson’s disease,”
Neuroscience, vol. 115, no. 4, pp. 1109–1117, 2002.
[39] J. L. Vitek, “Mechanisms of deep brain stimulation: excitation or inhibition,” Mov. Disord. Off. J.
Mov. Disord. Soc., vol. 17 Suppl 3, pp. S69-72, 2002.
[40] A. Cukiert, C. M. Cukiert, J. A. Burattini, P. P. Mariani, and D. F. Bezerra, “Seizure outcome after
hippocampal deep brain stimulation in patients with refractory temporal lobe epilepsy: A prospective,
controlled, randomized, double-blind study,” Epilepsia, vol. 58, no. 10, pp. 1728–1733, Oct. 2017.
[41] T. Tykocki, T. Mandat, A. Kornakiewicz, H. Koziara, and P. Nauman, “Deep brain stimulation for
refractory epilepsy,” Arch. Med. Sci. AMS, vol. 8, no. 5, pp. 805–816, Nov. 2012.
[42] D. Zhou and E. Greenbaum, Implantable Neural Prostheses 2: Techniques and Engineering
Approaches. Springer Science & Business Media, 2010.
[43] M. Pinsker, F. Amtage, M. Berger, G. Nikkhah, and L. T. van Elst, “Psychiatric side-effects of
bilateral deep brain stimulation for movement disorders,” Acta Neurochir. Suppl., vol. 117, pp. 47–51,
2013.
[44] R. Fisher et al., “Electrical stimulation of the anterior nucleus of thalamus for treatment of refractory
epilepsy: Deep Brain Stimulation of Anterior Thalamus for Epilepsy,” Epilepsia, vol. 51, no. 5, pp. 899–
908, May 2010.
[45] S. B. Brummer and M. J. Turner, “Electrochemical Considerations for Safe Electrical Stimulation of
the Nervous System with Platinum Electrodes,” IEEE Trans. Biomed. Eng., vol. BME-24, no. 1, pp. 59–
63, Jan. 1977.
[46] B. Piallat et al., “Monophasic But Not Biphasic Pulses Induce Brain Tissue Damage During
Monopolar High-Frequency Deep Brain Stimulation,” Neurosurgery, vol. 64, no. 1, pp. 156–163, Jan.
2009.
[47] J. Vesper, S. Chabardes, V. Fraix, N. Sunde, and K. Østergaard, “Dual channel deep brain
stimulation system (Kinetra) for Parkinson’s disease and essential tremor: a prospective multicentre open
label clinical study,” J. Neurol. Neurosurg. Psychiatry, vol. 73, no. 3, pp. 275–280, Sep. 2002.
[48] S. F.T., M. M.J., and W. J. R.E., “Responsive Cortical Stimulation for the Treatment of Epilepsy,”
Neurotherapeutics, vol. 5, no. 1, pp. 68–74, 2008.
67
[49] M. J. Morrell, “Responsive cortical stimulation for the treatment of medically intractable partial
epilepsy,” Neurology, vol. 77, no. 13, pp. 1295–1304, 2011.
[50] C. N. Heck et al., “Two-year seizure reduction in adults with medically intractable partial onset
epilepsy treated with responsive neurostimulation: Final results of the RNS System Pivotal trial,”
Epilepsia, vol. 55, no. 3, pp. 432–441, 2014.
[51] J. Yamamoto et al., “Low-frequency electric cortical stimulation has an inhibitory effect on epileptic
focus in mesial temporal lobe epilepsy,” Epilepsia, vol. 43, no. 5, pp. 491–495, 2002.
[52] M. Z. Koubeissi, E. Kahriman, T. U. Syed, J. Miller, and D. M. Durand, “Low-frequency electrical
stimulation of a fiber tract in temporal lobe epilepsy,” Ann Neurol, vol. 74, no. 2, pp. 223–231, 2013.
[53] R. Lesser, B. Gordon, and S. Uematsu, “Electrical stimulation and language,” J. Clin. Neurophysiol.,
vol. 11, no. 2, pp. 191–204, 1994.
[54] B. Gordon et al., “Parameters for direct cortical electrical stimulation in the human: histopathologic
confirmation.,” Electroencephalogr. Clin. Neurophysiol., vol. 75, no. 5, pp. 371–7, 1990.
[55] F. T. Sun and M. J. Morrell, “The RNS System: responsive cortical stimulation for the treatment of
refractory partial epilepsy,” Expert Rev. Med. Devices, vol. 11, no. 6, pp. 563–572, Nov. 2014.
[56] R. Esteller, J. Echauz, T. Tcheng, B. Litt, and B. Pless, “Line length: an efficient feature for seizure
onset detection,” in Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd Annual
International Conference of the IEEE, 2001, vol. 2, pp. 1707–1710.
[57] B. Litt et al., “Evolution of accumulated energy predicts seizures in mesial temporal lobe epilepsy,”
in Proceedings of the First Joint BMES/EMBS Conference. 1999 IEEE Engineering in Medicine and
Biology 21st Annual Conference and the 1999 Annual Fall Meeting of the Biomedical Engineering
Society (Cat. N, 1999, vol. 1, p. 440 vol.1-vol.
[58] J. Gotman, “Automatic recognition of epileptic seizures in the EEG,” Electroencephalogr. Clin.
Neurophysiol., vol. 54, no. 5, pp. 530–540, Nov. 1982.
[59] S. Stanslaski et al., “Design and Validation of a Fully Implantable, Chronic, Closed-Loop
Neuromodulation Device With Concurrent Sensing and Stimulation,” IEEE Trans. Neural Syst. Rehabil.
Eng., vol. 20, no. 4, pp. 410–421, Jul. 2012.
68
[60] A.-T. Avestruz et al., “A 5 $\mu$W/Channel Spectral Analysis IC for Chronic Bidirectional Brain–
Machine Interfaces,” IEEE J. Solid-State Circuits, vol. 43, no. 12, pp. 3006–3024, Dec. 2008.
[61] T. Denison, K. Consoer, W. Santa, M. Hutt, and K. Miesel, “A 2 #x003BC;W Three-Axis MEMS-
based Accelerometer,” in 2007 IEEE Instrumentation Measurement Technology Conference IMTC 2007,
2007, pp. 1–6.
[62] A. G. Rouse et al., “A chronic generalized bi-directional brain–machine interface,” J. Neural Eng.,
vol. 8, no. 3, p. 036018, Jun. 2011.
[63] N. Soltani, M. S. Aliroteh, and R. Genov, “Cellular inductive powering system for weakly-linked
resonant rodent implants,” in 2013 IEEE Biomedical Circuits and Systems Conference (BioCAS), 2013,
pp. 350–353.
[64] N. Verma, A. Shoeb, J. Bohorquez, J. Dawson, J. Guttag, and A. P. Chandrakasan, “A Micro-Power
EEG Acquisition SoC With Integrated Feature Extraction Processor for a Chronic Seizure Detection
System,” IEEE J. Solid-State Circuits, vol. 45, no. 4, pp. 804–816, Apr. 2010.
[65] D. G. Armstrong, D. N. Kleidermacher, D. C. Klonoff, and M. J. Slepian, “Cybersecurity Regulation
of Wireless Devices for Performance and Assurance in the Age of ‘Medjacking,’” J. Diabetes Sci.
Technol., vol. 10, no. 2, pp. 435–438, 2016.
[66] D. M. Groppe et al., “Dominant frequencies of resting human brain activity as measured by the
electrocorticogram,” NeuroImage, vol. 79, pp. 223–233, Oct. 2013.
[67] J. Gotman, J. R. Ives, and P. Gloor, “Frequency content of EEG and EMG at seizure onset:
possibility of removal of EMG artefact by digital filtering,” Electroencephalogr. Clin. Neurophysiol., vol.
52, no. 6, pp. 626–639, Dec. 1981.
[68] M. Le Van Quyen et al., “Preictal state identification by synchronization changes in long-term
intracranial EEG recordings,” Clin. Neurophysiol., vol. 116, no. 3, pp. 559–568, Mar. 2005.
[69] P. Jiruska, M. de Curtis, J. G. R. Jefferys, C. A. Schevon, S. J. Schiff, and K. Schindler,
“Synchronization and desynchronization in epilepsy: controversies and hypotheses,” J. Physiol., vol. 591,
no. Pt 4, pp. 787–797, Feb. 2013.
[70] J.-P. Lachaux, E. Rodriguez, J. Martinerie, F. J. Varela, and others, “Measuring phase synchrony in
brain signals,” Hum. Brain Mapp., vol. 8, no. 4, pp. 194–208, 1999.
69
[71] K. Abdelhalim, V. Smolyakov, and R. Genov, “Phase-synchronization early epileptic seizure
detector VLSI architecture,” IEEE Trans. Biomed. Circuits Syst., vol. 5, no. 5, pp. 430–438, 2011.
[72] D. Gupta and C. J. James, “Narrowband vs. Broadband Phase Synchronization Analysis Applied to
Independent Components of Ictal and Interictal EEG,” in 2007 29th Annual International Conference of
the IEEE Engineering in Medicine and Biology Society, 2007, pp. 3864–3867.
[73] J. Volder, “The CORDIC computing technique,” in afips, 1899, p. 257.
[74] B. Voloh, T. A. Valiante, S. Everling, and T. Womelsdorf, “Theta–gamma coordination between
anterior cingulate and prefrontal cortex indexes correct attention shifts,” Proc. Natl. Acad. Sci., vol. 112,
no. 27, pp. 8457–8462, Jul. 2015.
[75] A. B. L. Tort, R. Komorowski, H. Eichenbaum, and N. Kopell, “Measuring Phase-Amplitude
Coupling Between Neuronal Oscillations of Different Frequencies,” J. Neurophysiol., vol. 104, no. 2, pp.
1195–1210, Aug. 2010.
[76] R. T. Canolty and R. T. Knight, “The functional role of cross-frequency coupling,” Trends Cogn.
Sci., vol. 14, no. 11, pp. 506–515, Nov. 2010.
[77] M. Guirgis, Y. Chinvarun, P. L. Carlen, and B. L. Bardakjian, “The role of delta-modulated high
frequency oscillations in seizure state classification,” in 2013 35th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC), 2013, pp. 6595–6598.
[78] C. de Hemptinne et al., “Exaggerated phase–amplitude coupling in the primary motor cortex in
Parkinson disease,” Proc. Natl. Acad. Sci. U. S. A., vol. 110, no. 12, pp. 4780–4785, Mar. 2013.
[79] C. de Hemptinne et al., “Therapeutic deep brain stimulation reduces cortical phase-amplitude
coupling in Parkinson’s disease,” Nat. Neurosci., vol. 18, no. 5, pp. 779–786, Apr. 2015.
[80] W. D. Penny, E. Duzel, K. J. Miller, and J. G. Ojemann, “Testing for nested oscillation,” J. Neurosci.
Methods, vol. 174, no. 1, pp. 50–61, Sep. 2008.
[81] R. T. Canolty et al., “High gamma power is phase-locked to theta oscillations in human neocortex,”
science, vol. 313, no. 5793, pp. 1626–1628, 2006.
[82] C. M. Florez et al., “In Vitro Recordings of Human Neocortical Oscillations,” Cereb. Cortex, vol. 25,
no. 3, pp. 578–597, Mar. 2015.
70
[83] A. H. Shoeb, “Application of machine learning to epileptic seizure onset detection and treatment,”
Thesis, Massachusetts Institute of Technology, 2009.
[84] P. C. Loizou, “Introduction to cochlear implants,” IEEE Eng. Med. Biol. Mag., vol. 18, no. 1, pp.
32–42, Jan. 1999.
[85] C. R. Butson and C. C. McIntyre, “Current Steering to Control the Volume of Tissue Activated
During Deep Brain Stimulation,” Brain Stimulat., vol. 1, no. 1, pp. 7–15, Jan. 2008.
[86] W. M. Grill, “Model-Based Analysis and Design of Waveforms for Efficient Neural Stimulation,”
Prog. Brain Res., vol. 222, pp. 147–162, 2015.
[87] H. Kassiri et al., “27.3 All-wireless 64-channel 0.013 mm 2/ch closed-loop neurostimulator with rail-
to-rail DC offset removal,” in Solid-State Circuits Conference (ISSCC), 2017 IEEE International, 2017,
pp. 452–453.
[88] G. O’Leary, T. A. Valiante, and R. Genov, “Low-latency VLSI architecture for neural cross-
frequency coupling analysis,” in Engineering in Medicine and Biology Society (EMBC), 2017 39th
Annual International Conference of the IEEE, 2017, pp. 2247–2250.
[89] M. Ihle et al., “EPILEPSIAE – A European epilepsy database,” Comput. Methods Programs
Biomed., vol. 106, no. 3, pp. 127–138, Jun. 2012.
[90] W.-M. Chen et al., “A Fully Integrated 8-Channel Closed-Loop Neural-Prosthetic CMOS SoC for
Real-Time Epileptic Seizure Control,” IEEE J. Solid-State Circuits, vol. 49, no. 1, pp. 232–247, Jan.
2014.
[91] M. A. Bin Altaf, C. Zhang, and J. Yoo, “A 16-Channel Patient-Specific Seizure Onset and
Termination Detection SoC With Impedance-Adaptive Transcranial Electrical Stimulator,” IEEE J. Solid-
State Circuits, vol. 50, no. 11, pp. 2728–2740, Nov. 2015.
[92] A. F. Struck, A. J. Cole, S. S. Cash, and M. B. Westover, “The number of seizures needed in the
EMU,” Epilepsia, vol. 56, no. 11, pp. 1753–1759, Nov. 2015.
[93] H. Ung et al., “Temporal behavior of seizures and interictal bursts in prolonged intracranial
recordings from epileptic canines,” Epilepsia, vol. 57, no. 12, pp. 1949–1957, Dec. 2016.
[94] S. Ladouce, D. I. Donaldson, P. A. Dudchenko, and M. Ietswaart, “Understanding Minds in Real-
World Environments: Toward a Mobile Cognition Approach,” Front. Hum. Neurosci., vol. 10, Jan. 2017.
71
[95] “Medtronic DBS Programmer Manual.” [Online]. Available:
http://www.medtronic.com/content/dam/medtronic-com-m/mdt/neuro/documents/dbs-prog37642-
ptmanl.pdf. [Accessed: 28-Dec-2017].
[96] Medtronic, “Medtronic Concepts for Investigational Tools to Support the NIH Brain Initiative.”
[Online]. Available: https://braininitiative.nih.gov/pdf/Medtronic_descriptions_508C.pdf. [Accessed: 30-
Dec-2017].
[97] B. Isaacson, S. Dani, S. A. Desai, T. Denison, and P. Afshar, “A rapid algorithm prototyping tool for
bi-directional neural interfaces,” in Neural Engineering (NER), 2013 6th International IEEE/EMBS
Conference on, 2013, pp. 633–636.
[98] R. Wennberg, H. P. Zaveri, M. G. Frei, and S. Arthurs, Epilepsy: The Intersection of Neurosciences,
Biology, Mathematics, Engineering, and Physics. CRC Press, 2016.
[99] S. R. Haut et al., “Interrater reliability among epilepsy centers: multicenter study of epilepsy
surgery,” Epilepsia, vol. 43, no. 11, pp. 1396–1401, Nov. 2002.
[100] R. S. Fisher, W. R. Webber, R. P. Lesser, S. Arroyo, and S. Uematsu, “High-frequency EEG
activity at the start of seizures,” J. Clin. Neurophysiol. Off. Publ. Am. Electroencephalogr. Soc., vol. 9,
no. 3, pp. 441–448, Jul. 1992.
[101] B. Schölkopf, R. C. Williamson, A. J. Smola, J. Shawe-Taylor, and J. C. Platt, “Support vector
method for novelty detection,” in Advances in neural information processing systems, 2000, pp. 582–588.
[102] J. Volkmann, E. Moro, and R. Pahwa, “Basic algorithms for the programming of deep brain
stimulation in Parkinson’s disease,” Mov. Disord., vol. 21, no. S14, pp. S284–S289, Jun. 2006.
[103] M. Parastarfeizabadi and A. Z. Kouzani, “Advances in closed-loop deep brain stimulation devices,”
J. NeuroEngineering Rehabil., vol. 14, no. 1, Dec. 2017.
[104] R. J. Staba, M. Stead, and G. A. Worrell, “Electrophysiological Biomarkers of Epilepsy,”
Neurotherapeutics, vol. 11, no. 2, pp. 334–346, Apr. 2014.
[105] J. Jacobs et al., “High-frequency oscillations (HFOs) in clinical epilepsy,” Prog. Neurobiol., vol. 98,
no. 3, pp. 302–315, Sep. 2012.
[106] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz, “Seizure prediction: the long and
winding road,” Brain, vol. 130, no. 2, pp. 314–333, Feb. 2007.
72
[107] K. Gadhoumi, J.-M. Lina, F. Mormann, and J. Gotman, “Seizure prediction for therapeutic devices:
A review,” J. Neurosci. Methods, vol. 260, pp. 270–282, Feb. 2016.
[108] S. Luan, I. Williams, K. Nikolic, and T. G. Constandinou, “Neuromodulation: present and emerging
methods,” Front. Neuroengineering, vol. 7, Jul. 2014.
[109] N. Grossman et al., “Noninvasive Deep Brain Stimulation via Temporally Interfering Electric
Fields,” Cell, vol. 169, no. 6, p. 1029–1041.e16, Jun. 2017.