Edge deep learning for neural implants: a case study of seizure detection and prediction

Abstract

Objective. Implanted devices providing real-time neural activity classification and control are increasingly used to treat neurological disorders, such as epilepsy and Parkinson’s disease. Classification performance is critical to identifying brain states appropriate for the therapeutic action (e.g. neural stimulation). However, advanced algorithms that have shown promise in offline studies, in particular deep learning (DL) methods, have not been deployed on resource-restrained neural implants. Here, we designed and optimized three DL models or edge deployment and evaluated their inference performance in a case study of seizure detection. Approach. A deep neural network (DNN), a convolutional neural network (CNN), and a long short-term memory (LSTM) network were designed and trained with TensorFlow to classify ictal, preictal, and interictal phases from the CHB-MIT scalp EEG database. A sliding window based weighted majority voting algorithm was developed to detect seizure events based on each DL model’s classification results. After iterative model compression and coefficient quantization, the algorithms were deployed on a general-purpose, off-the-shelf microcontroller for real-time testing. Inference sensitivity, false positive rate (FPR), execution time, memory size, and power consumption were quantified. Main results. For seizure event detection, the sensitivity and FPR for the DNN, CNN, and LSTM models were 87.36%/0.169 h⁻¹, 96.70%/0.102 h⁻¹, and 97.61%/0.071 h⁻¹, respectively. Predicting seizures for early warnings was also feasible. The LSTM model achieved the best overall performance at the expense of the highest power. The DNN model achieved the shortest execution time. The CNN model showed advantages in balanced performance and power with minimum memory requirement. The implemented model compression and quantization achieved a significant saving of power and memory with an accuracy degradation of less than 0.5%. Significance. Inference with embedded DL models achieved performance comparable to many prior implementations that had no time or computational resource limitations. Generic microcontrollers can provide the required memory and computational resources, while model designs can be migrated to application-specific integrated circuits for further optimization and power saving. The results suggest that edge DL inference is a feasible option for future neural implants to improve classification performance and therapeutic outcomes.

Full text

Accepted for publication in the Journal of Neural Engineering
Citation: Xilin Liu and Andrew G Richardson 2021 J. Neural Eng. 18 046034
Edge Deep Learning for Neural Implants: A Case
Study of Seizure Detection and Prediction
Xilin Liu and Andrew G. Richardson
University of Pennsylvania, Philadelphia, PA, USA
E-mail: xilinliu@seas.upenn.edu
Abstract. Objective. Implanted devices providing real-time neural activity
classification and control are increasingly used to treat neurological disorders, such as
epilepsy and Parkinson’s disease. Classification performance is critical to identifying
brain states appropriate for the therapeutic action (e.g. neural stimulation). However,
advanced algorithms that have shown promise in offline studies, in particular deep
learning (DL) methods, have not been deployed on resource-restrained neural implants.
Here, we designed and optimized three DL models or edge deployment and evaluated
their inference performance in a case study of seizure detection. Approach. A deep
neural network (DNN), a convolutional neural network (CNN), and a long short-term
memory (LSTM) network were designed and trained with TensorFlow to classify ictal,
preictal, and interictal phases from the CHB-MIT scalp EEG database. A sliding
window based weighted majority voting (WMV) algorithm was developed to detect
seizure events based on each DL model’s classification results. After iterative model
compression and coefficient quantization, the algorithms were deployed on a general-
purpose, off-the-shelf microcontroller for real-time testing. Inference sensitivity, false
positive rate (FPR), execution time, memory size, and power consumption were
quantified. Main results. For seizure event detection, the sensitivity and FPR for
the DNN, CNN, and LSTM models were 87.36%/0.169 h−1, 96.70%/0.102 h−1, and
97.61%/0.071 h−1, respectively. Predicting seizures for early warnings was also feasible.
The LSTM model achieved the best overall performance at the expense of the highest
power. The DNN model achieved the shortest execution time. The CNN model showed
advantages in balanced performance and power with minimum memory requirement.
The implemented model compression and quantization achieved a significant saving
of power and memory with an accuracy degradation of less than 0.5%. Significance.
Inference with embedded DL models achieved performance comparable to many prior
implementations that had no time or computational resource limitations. Generic
microcontrollers can provide the required memory and computational resources, while
model designs can be migrated to application-specific integrated circuits (ASICs) for
further optimization and power saving. The results suggest that edge DL inference is
a feasible option for future neural implants to improve classification performance and
therapeutic outcomes.
Keywords: neural interface, machine learning, deep learning, edge computing, DNN,
CNN, LSTM, seizure detection, epilepsy, EEG.

2
1. Introduction
Many brain injuries and diseases may be treated by implanted devices that provide real-
time classification of neural activity to produce a suitable control output. For example,
closed-loop neuromodulatory devices control neural activity classified as pathological
with electrical stimulation to treat epilepsy and Parkinson’s disease [1]. Brain-machine
interface (BMI) devices classify volitional intent to control external communication or
movement devices for paralyzed individuals [2]. In both cases, the neural implants record
brain activity, select and extract relevant activity features, and perform classification
and control on the basis of these features.
Performance of these devices is largely dependent on the feature selection and
classification algorithms. Feature selection aims to transform the often noisy, correlated
signals from many recording channels into a few non-redundant, informative inputs
to the machine learning classifier. Feature selection is often a manually-specified,
time-consuming process that requires substantial domain expertise [3]. Furthermore,
current implantable devices such as the Neuropace RNS and Medtronic Activa PC+S
closed-loop neurostimulators have restrained computation resources. Thus, only simple
classification algorithms, such as feature thresholding or linear discriminant analysis,
have been implemented [4].
Deep learning (DL) is a machine learning algorithm that has recently been applied
to research-grade, non-implantable neural interface devices to improve performance
[5]. Specifically, DL combines feature extraction, feature selection, and classification
into a single framework, jointly optimizing the end-to-end process [6]. When sufficient
training data is available [7], DL can achieve superior performance compared to more
conventional algorithms [8], especially in distinguishing hidden features critical for
classification [9]. Furthermore, the DL approach is robust and more generalizable
across different applications [3]. Although numerous DL approaches for neural interface
devices have been studied, nearly all have done so offline [7]. While DL training is
computationally intensive and likely to remain offline, DL inference must be performed
online for real-time control. This is a particularly challenging problem for clinical-grade
devices operating on battery-powered microprocessors or integrated circuits with limited
computational resources and energy budget.
To implement real-time DL inference for clinical neural implants, three paradigms
have been proposed (Figure 1, left). First, inference can be done remotely through
cloud computing (Paradigm A) [4]. In this paradigm, the neural implant transmits
the recorded data to a wearable device via local wireless communication, and the
wearable device in turn uploads the data to a cloud-based workstation via internet or
telecommunication. The cloud-based inference result is downloaded wirelessly back to
the wearable device and then to the neural implant for producing the control output (e.g.
neural stimulation). Second, inference can be done on the wearable device without the
need for data transfer to a cloud-based platform (Paradigm B) [10]. Third, DL inference
can be performed directly on the neural implant itself, requiring no data transfer to

3
Figure 1. Left, Illustration of three paradigms for executing deep learning inference.
Paradigm A: DL inference on the cloud via internet or telecommunication. Paradigm
B: DL inference on the wearable device via local wireless communication. Paradigm
C: DL inference directly on the neural implants without any data transfer. Right, The
qualitative strengths and weaknesses of each paradigm.
another device (Paradigm C) [11]. Each of these three inference paradigms has its
strengths and weaknesses (Figure 1, right). Although the cloud computing in Paradigm
A offers the best possible inference accuracy, the latency and the robustness are the
worst, which can adversely effect the timing-critical closed-loop therapeutic intervention.
There are also concerns about cybersecurity and data privacy due to the required data
transfer via internet or telecommunication [12]. Paradigm B eliminates the dependence
on remote data transfer and the associated concerns, but it still relies on the robustness
and security of the local wireless communication between the wearable devices and
the neural implants. Moreover, the inference model complexity will be limited by
the computational and power resources of the wearable devices. Paradigm C, using
edge computing [13], avoids the disadvantages and concerns of wireless communication
and can therefore potentially achieve the best robustness, the lowest latency, and the
minimum security risks. Even if occasional wireless data transfer is performed in
Paradigm C for offline analysis and performance assessments, the therapy itself would
be robust to data transfer disruption and the opportunity for malicious attack would
be the smallest of the three paradigms due to infrequency of transmission. However,
a major question regarding Paradigm C is whether an edge DL model can achieve
desired inference performance. Although machine learning-enabled processors for neural
implants have been reported [14, 15], DL models have only rarely been implemented
[11, 16, 17]. Therefore, the objective of the present work was to evaluate the inference
performance limitations and resource constraints of edge DL designs through a case
study.
For a DL model to be successfully deployed on a neural implant for real-time
inference, the model design needs to meet three requirements. First, the model’s

4
inference performance (e.g. accuracy, sensitivity, specificity, etc.) should meet the
requirements of the target application, including potential degradation from model
compression and quantization. Second, the model’s inference time should meet the
latency requirement, especially in closed-loop applications. This not only requires a
sufficient computational speed, but also limits the maximum length of the buffered input
data segment, resulting in a constraint on the model’s architecture. Third, the power
and resource costs of the DL model should meet the budget of the neural implant. For
instance, the available memory resources, both non-volatile storage memory and random
access memory (RAM), limit the total number of trainable model parameters.
Potential hardware platforms for neural implants with edge DL include
dedicated AI accelerators and general-purpose low-power processors. Dedicated AI
accelerators provide higher energy efficiency than general-purpose processors. However,
commercially available AI accelerators, such as IBM’s TrueNorth [18], Google’s Edge
TPU [19], and Intel’s Loihi [20], are too power hungry for neural implants. Also, large
volume mobile devices-oriented AI intellectual property cores are not easily accessible
for research-grade prototyping [21]. Ultra low-power AI accelerator chips have mainly
been developed in research labs [22–24]. On the other hand, general-purpose low-
power processors, such as ARM Cortex R
, Texas Instruments MSP430 R
, and open-
source reduced instruction set computer (RISC)-based microcontrollers (MCUs) provide
a solution for low-cost, rapid prototyping for medical research and pre-clinical trials. If
a further reduction in power or device footprint is desired, the MCU-based design can
be migrated to an application-specific integrated circuit (ASIC) by integrating the MCU
core together with analog neural recording and stimulation circuits, with an optional
wireless communication module [14, 25]. Thus, in this work, we focused on edge DL
design using a general-purpose MCU.
To investigate DL design and optimization methods fulfilling practical requirements
of neural implants, we conducted a case study of epileptic seizure detection. Real-time
seizure detection and intervention through closed-loop neural stimulation has proven to
be an effective treatment for medically-refractory epilepsy [26]. Due to the importance
of accurate, low-latency detection by a battery-powered neural implant, this application
could benefit from edge computing. We adopted three commonly used DL architectures
and customized the design for seizure detection on a MCU. We validated the methods
using a publicly available annotated epilepsy database. The models were trained offline
using Tensorflow. Compression and quantization methods were investigated for reducing
the computational cost, memory, and power consumption for real-time inference. The
inference results were compared to prior publications with no time and computational
resource limitations. The strengths and weaknesses of each model were analyzed.
Finally, we discuss the remaining challenges and future efforts toward implementing
this novel paradigm in clinical neural implants.

5
2. Methods
2.1. Data Preparation
This study used the Boston Children’s Hospital (CHB)-MIT scalp electroencephalogra-
phy (EEG) database [27], which is publicly available at PhysioNet.org [28]. Although
scalp EEG is inconsistent with signals utilized by real neural implants, the CHB-MIT
database is one of the most popular epilepsy data sets used for benchmarking, allow-
ing a fair performance comparison between our edge DL approach and prior methods.
The database contains recordings collected from 23 pediatric patients with intractable
epilepsy. The placement of the surface electrodes followed the international 10-20 system
[29]. The original recordings were sampled at 256 Hz and 16-bit resolution, with a 60
Hz notch filter applied for removing the mains interference. The seizure start and stop
time points for each patient were manually annotated by clinical experts. Channels that
were not constantly available throughout the entire duration in each case were excluded
in developing the algorithms, as suggested in [9]. No signal processing was used before
training and testing the models.
Data segments during ictal, preictal, and interictal phases were selected for training
the DL models. The ictal phases were well defined by the expert annotated seizure start
and stop time points. There is no consistent definition of preictal phases in the literature,
and the occurrence of preictal characteristics may differ across patients. In this work,
we used the 3 min of recordings that are 30 sec prior to each seizure’s start time as
the preictal phases. It should be noted that preictal characteristics could happen much
earlier than this period [30]. Interictal segments were selected at least 2 hours before
and after any ictal phase to avoid potential signal contamination [31]. An even longer
interval between ictal and interictal segments up to 4 hours may further improve the
classification performance [9, 32]. Unfortunately, the recordings of those phases are not
always available in the chosen database. Postictal recordings were not used in this study.
The recordings from the interictal and preictal phases were much longer than the
ictal phase. If all data segments were used in training, the severe imbalanced class
distribution would cause a bias in the model’s prediction [33]. We adopted two methods
to address this issue. First, we generated more ictal segments by sliding the window
with an overlap of 50% [31]. Second, we applied a class weight to the loss function [3].
Interictal segments were randomly selected during day and night to avoid overfitting to
irrelevant activities.
2.2. Deep Learning Models
We investigated three DL models for the seizure detection task: 1) a deep neural network
(DNN) model, 2) a convolutional neural network (CNN) model, and 3) a bidirectional
long short-term memory (LSTM) network model. The architectures of the three models
are illustrated in Figure 2. Although the model training was patient-specific, each
model’s architecture and parameter settings were kept the same for all patients in the

6
database. The design details are presented in the following sections.
Figure 2. Simplified diagrams of the three DL models used in this study: (a) the deep
neural network (DNN) model, (b) the convolutional neural network (CNN) model, and
(c) the bidirectional long short-term memory (LSTM) model. Parameters of each layer
are highlighted in blue.
2.2.1. Deep Neural Network (DNN) DNN is a type of feedforward artificial neural
network that consists of multiple layers. Each layer contains many processing units,
namely artificial neurons. The processing function of each neuron is given in the
Appendix. By connecting these neurons, DNNs emulate the way the brain processes
information. Each neuron’s output is processed by an activation function. In this work,
the rectified linear unit (ReLU) function is used in all models [34]. Conventional artificial
neural networks often use nonlinear activation functions, such as hyperbolic tangent
and sigmoid. However, these functions make gradient-based training challenging as the
number of network layers increases, an issue known as the gradient vanishing problem
[35]. The introduction of the ReLU function overcomes this limitation by preserving the
linear properties of positive values for gradient-based optimization, while still providing
nonlinearity by setting all negative values to zero [34]. Moreover, the ReLU function is

7
computational friendly, which is important for the purposes of this work.
Our DNN model consists of a scaling layer, a flat and concatenation layer, followed
by five layers of fully-connected (FC) neurons. The scaling layer removes the DC offset
from each channel and linearly scales the input data based on the dynamic range. The
FC layers form a pyramid shape with Nneurons in the first layer, where Nis the length
of the input segment of one channel. A 50% dropout layer is inserted before each of
the first two FC layers. The class (ictal, preictal, or interictal) that has the highest
activation function value is the final classification result.
2.2.2. Convolutional Neural Network (CNN) CNNs have achieved extensive success
in visual image recognition and natural language processing [36, 37]. Using CNNs in
seizure detection tasks has also been reported [31, 38, 39]. Convolutional (Conv) layers
can be trained to automatically extract underlying features that best represent the data
without human intervention. Given their superior ability in extracting features from
images, past studies have converted time-domain EEG segments into spectrograms to
be used as the CNN inputs [31]. However, the computational costs of the Fourier or
wavelet transform of multiple EEG channels are prohibitive for real-time applications
using low-power MCUs. In this work, we used Conv layers to extract features directly
from time-domain EEG signals.
Instead of using 2-D standard Conv filters, we used 1-D Conv filters to process each
EEG channel. This allowed us to reduce the model size and computational cost, while
still preserving the most important features [3]. Our CNN model consists of three Conv
layers, all using ’same’ padding and stride=1. Each of the first two Conv layers consists
of 4 kernels with a length of fs/2, where f s is the sampling rate of the input EEG signal.
The third Conv layer consists of 2 kernels with a length of fs/4. A max-pooling layer
with a length of 4 was added to each of the Conv layers. The max pooling was used to
prevent overfitting while reducing the computational costs. A 25% dropout layer was
inserted after each of the first two Conv layers to regularize the model. The outputs of
the Conv layers were concatenated and processed by three FC layers.
2.2.3. Long Short-term Memory Network (LSTM) LSTM is a type of recurrent neural
network (RNN). In contrast to the feedforward neural networks described above, a RNN
has recurrent connections that are suitable for capturing sequential information in the
data. In particular, gating functions are used in each cell of the LSTM layers to control
precisely what information is to be kept in the network and what is to be removed.
Thus, LSTM has an inherent advantage in extracting certain temporal characteristics
in time-domain signals [40], which is crucial in tasks such as seizure detection.
In this work, we adopted a bidirectional LSTM network. Bidirectional LSTM
networks process sequential information from two opposite directions simultaneously,
which has proven useful in applications such as speech processing [41]. For the gate
activation function, we used hard-sigmoid instead of sigmoid to avoid the exponential
operation, and we used softsign as the state activation function. The processing and

8
activation functions of the LSTM model are given in the Appendix.
We adopted a topology that combines an input Conv layer with the bidirectional
LSTM layers. One kernel with a length of fs/2 was used in the Conv layer, followed by
a max-pooling layer with a length of 4. A 50% dropout layer was inserted between the
input Conv layer and the bidirectional LSTM layers. Each of the bidirectional LSTM
layers consists of 128 cells. Finally, FC layers convert the outputs for classification.
2.2.4. Sliding Window Based Weighted Majority Voting Short data segment based
classification often suffers from a trade-off between sensitivity and FPR [4]. Since
not every single data segment in the preictal phase exhibits signal characteristics that
are related to an oncoming seizure, there is usually a limitation of the achievable
classification accuracy before model overfitting. To achieve a high sensitivity while
minimizing the FPR, we propose a novel sliding window based weighted majority
voting (WMV) algorithm. The algorithm was implemented for real-time operation,
as described in the pseudo-code of Algorithm 1.
The algorithm uses the DL model’s segment-based classification results as inputs.
The evaluation is based on a sliding window of Msegments. A score for the present
window to be in an ictal phase is calculated based on two weighted terms: (i) how
many individual segments generate a classification result of ictal phase, weighted by
a coefficient of αI, and (ii) how many times the ictal classification repeats in a row,
weighted by a coefficient of βI. The score for the present window to be in a preictal
phase is estimated in an equivalent way, based on the two terms corresponding to preictal
phases, weighted by coefficients αPand βP. Finally, two pre-defined thresholds θI
and θPare used to determine the final classification result from the sliding window.
Once a score crosses the threshold, the algorithm will mark the event and break from
evaluating the current sliding window. This is mainly to avoid delays during real-time
seizure detection. A new evaluating sliding window starts immediately after the previous
window terminates. Compared with traditional majority voting or moving average based
algorithms, the proposed algorithm favors a prediction if the same classification results
appear in a continuous manner. This significantly reduces the FPR in the classification.
Moreover, the detection latency is not limited by the sliding window length as the
algorithm terminates once the thresholds are crossed.
2.3. Model Compression and Quantization
For the models to be successfully deployed in a MCU, we applied compression and
quantization techniques to reduce the computational cost. There are existing channel
pruning techniques for reducing model dimension without retraining with data [42].
However, since our model architectures are compact, the most effective method of
compression is through iterations of retraining.
To select the most informative channels as the model inputs, we ranked all available
recording channels based on the line length feature of these channels during the ictal

9
Algorithm 1: Sliding window based WMV algorithm
Inputs :P red Seg [ 1 : M]∈ {I ctal, P reictal, I nterictal}, ;
M: total segments in the win,
P arameters :αI,P , βI,P , and θI,P
Output:P red Ev ent ∈ {Ictal, P reictal, Interictal}
1Initialization:
2Score [Ictal ]←0,Score [P reictal ]←0,Score [I nterictal ]←0
3Acc [Ictal ]←0,Acc [P reictal ]←0
4P red Ev ent ←Interictal ;
5for i←1to Mdo
6switch based on Pred Seg [ i ] do
7case Ictal
8Score [Ictal ]←+= αI+βI·Acc [Ictal ]
9Acc [Ictal ]←+= 1, Acc [P reictal ]←0
10 endsw
11 case Preictal
12 Score [P reictal ]←+= αP+βP·Acc [P reictal ]
13 Acc [P reictal ]←+= 1, Acc [Ictal ]←0
14 endsw
15 case Interictal
16 Acc [Ictal ]←0, Acc [P reictal ]←0
17 endsw
18 endsw
19 if Score [Ictal ]> θIthen
20 P red Ev ent ←Ictal
21 Break;
22 else if S core [P reictal ]> θPthen
23 P red Ev ent ←P reictal
24 Break;
25 else
26 Continue;
27 end
28 end
phase. Line length is a measure of both high amplitude and high-frequency content of
a time-domain signal, and is proven to be among the most effective features of seizures
[43, 44]. It is computed as:
fL(xi) = 1
N
N−1
X
t=1
|xi(t−1) −xi(t)|(1)
where xi(t) is the EEG signal of channel iat time t,Nis the sample count of the
ictal segment. The line length feature was only computed offline for channel ranking.
It was not used in real-time seizure detection. Based on the channel ranking, the top
Kchannels were used as the inputs for the models. The classification performances

10
using different Kwere compared for determining the optimal channel set. Similarly, we
compared the performance using different data segment length N.
Computing a DL model using high arithmetic precision doesn’t necessarily improve
performance [45]. Quantization of high precision coefficients not only saves the
computational cost, but also reduces the memory cost for storing the coefficients. In fact,
loading the coefficients from memory to the arithmetic unit can dominate the energy
consumption in a microprocessor [46]. In this work, we quantized all coefficients to 8-bit
fixed-point numbers. To evaluate the effects of model quantization, the performance of
8-bit quantized models was compared to that of 16-bit models (referred to hereafter as
unquantized models). Together with the computationally efficient non-linear activation
functions (Appendix Eqs. 7, 14 and 15) used in the model, the computational cost and
latency were reduced.
2.4. Training and Testing Methods
The DL models were trained by supervised learning using the annotated data segments.
Adam optimizer was used in training all of the models [47]. The experiments for model
compression and quantization were performed in a 10-fold cross validation (CV). We
used stratified CV where each fold has equal number of instances for each class. The
number of segments depends on the available ictal onset time in each subject’s recording,
and the data segment length. The total data segments used during the 10-fold CV were
less than 1% of the whole recording per subject.
After finalizing the model architectures, we evaluated the performances using a
variant of the leave-one-out cross validation (LOOCV). For a subject’s recording that
contains Kseizure events, the recording is divided into Ksections with one seizure
event in each of them. One section is used for validation and the rest K−1 sections are
used for training. The process is repeated Ktimes so that all data is used exhaustively
for validation. LOOCV suffers from large variation when Kis small [48]. In this work,
we evaluate cases in the CHB-MIT that have at least 5 seizures in the recording, which
lead to 15 cases including chb01, 03, 05, 06, 08, 10, 12-16, 18, 20, 23 and 24. Unlike
the segment based 10-fold CV, the LOOCV is performed in a continuous manner in
real-time using the sliding WMV algorithm.
Standard classification metrics were used, including accuracy, sensitivity, specificity,
and FPR:
Sensitivity =T P
T P +F N (2)
Specif icity =T N
T N +F P (3)
Accuracy =T P +T N
T P +T N +F P +F N (4)
FPR (h−1) = F P
Total recording length (in hours) (5)

11
where TP, TN, FP, and FN are true positive, true negative, false positive, and false
negative detection, respectively. During the segment-based evaluation, results were
directly compared with the annotated labels. For event-based seizure detection, we
made the following definitions:
(i) If a detection event happens within 5 seconds of the seizure start time, it is
considered as a TP. There can be at most one TP per genuine seizure.
(ii) If no detection event happens within 5 seconds of the seizure start time, it is
considered as a FN. There can be at most one FN per genuine seizure.
(iii) If a detection event happens earlier than 5 seconds of the seizure start time, or 5
seconds after the seizure end time, it is considered to be a FP. The FP count is not
limited by the number of seizures.
Similarly, we made the definitions for event-based seizure prediction as follows:
(i) If a warning alert happens within 40 min before the seizure start time, it is
considered as a TP. There can be at most one TP per genuine seizure.
(ii) If no warning alert happens within 40 min before the seizure start time, it is
considered as a FN. There can be at most one FN per genuine seizure.
(iii) If a warning alert happens earlier than 40 min before or after the seizure start time,
it is considered to be a FP. The FP count is not limited by the number of seizures.
MATLAB R
was used for data handling. The DL models were implemented in
Python with Tensorflow, which is an open-source machine learning library developed by
the Google Brain team [49]. The DL training was performed on Google Cloud clusters
using tensor processing units.
2.5. Hardware Implementation
After training, the compressed and quantized models were deployed on a low-power
MCU. We used a 32-bit ARM R
Cortex-M4 based MCU nRF52840 from Nordic
Semiconductor [50]. The nRF52840 features a flash memory size of 1 MB and a RAM
size of 256 kB. The MCU runs at a clock rate of 64MHz. The Cortex-M4 core supports
multiple types of hardware multiplication in one clock cycle, including 16-bit signed
multiplication with 32-bit results [51]. The integrated floating-point unit was not used
in this work.
The trained DL models were implemented in C/C++ for programming the MCU.
The code was developed using Keil R
MDK [52]. Open-source and commercial tools can
be used to assist the code conversion from trained DL models to embedded systems
[53, 54]. The neural signal acquisition process was not implemented in this work.
Instead, the integrated USB 2.0 full speed module was used for transferring the recorded
EEG data from the computer to the MCU. The input data segments were buffered in
the RAM. The direct memory access module was used for transferring the data so that
the CPU was not interrupted [55]. The inference results were returned to the computer
via the same USB interface.

12
3. Results
3.1. Model Optimization
Three DL models were trained on all 24 cases (from 23 patients) in the CHB-MIT
database. Models were not quantized during the training and optimization phase. The
input channels were selected among all available EEG channels for each patient using
the proposed line-length based ranking (Eq. 1). Figure 3 (a) shows the classification
performance of each model using 1, 5, 9, 13, or 18 channels as the inputs. The segment
length was chosen to be 1 sec (256 samples) in this analysis. The experimental result
suggests that the optimal channel count is different for each model. Given the existing
model size and experimental setup, the DNN model can take no more than 5 channels
before it fails to extract critical information. More channels were generally helpful for
the CNN model, but the performance improved marginally beyond 9 channels. The
LSTM model reached peak performance at 9 and 13 channels, while more channels only
caused additional variance.
Segment length is another key parameter that impacts the overall performance as
well as the computational cost. Figure 3 (b) shows the classification performance of
each model using a segment length of 0.25 sec, 0.5 sec, 1 sec, 2 sec, or 4 sec. The ictal
segments were generated with 50% overlap in all cases. In the DNN model, using 5
channels as the input, the optimal performance was obtained with a segment length of
0.5 sec. In the CNN model, using 9 channels as the input, the optimal performance was
obtained with a segment length of 1 sec. In the LSTM model, also using 9 channels as
the input, the optimal performance was obtained with a segment length of 2 sec.
We targeted the optimal performance of each model that the hardware resources
permit. In certain model configurations, using more channels as inputs may reduce the
segment size for achieving similar performance. Depending on the system’s requirement
and the available hardware resources, one may prefer to use more input channels or a
larger segment size. In practice, more input channels require more wearable/implantable
electrodes and corresponding recording hardware, such as low-noise neural amplifiers.
On the other hand, a larger window size requires more model coefficients thus more
memory and RAM resources, and may cause a systematic latency. It should also be
noted that increasing the segment length reduces the number of training and testing
data segments, especially for patients with short ictal phases (eg. 6-9 sec in chb16).
This could be a limiting factor in training DL models using a limited database. The
hyperparameters used for training should be carefully tuned to minimize the impact.
Figure 4 shows the segment-based classification accuracy of each model. The
DNN uq, CNN uq, and LSTM uq indicated in the figure are the 16-bit unquantized
models, while DNN, CNN, and LSTM are the 8-bit quantized versions. The average
performance degradation due to quantization was less than 0.5%. The LSTM model
achieved the best overall performance (90.94%), followed by the CNN model (89.21%).
The performance of the DNN model was relatively poor (64.55%).
To compare the classification performance with non-DL algorithms, we constructed

13
Figure 3. Model optimization during 10-fold CV. Each model’s segment-based
classification accuracy is plotted with error bars showing 95% confidence interval. (a)
Using 1, 5, 9, 13, or 18 EEG channels as the inputs. The channel selection was based
on the line-length ranking algorithm. The segment length was 1 sec. (b) Using segment
lengths of 0.25s, 0.5s, 1s, 2s, or 4s. The number of input channels for the DNN, CNN,
and LSTM model was 5, 9, and 9, respectively.
a linear discriminant analysis (LDA) classifier. Spectral amplitudes in selected frequency
bands were used as the input features. The LDA classifier was chosen because it has
been used in the Medtronic Activa PC+S device [4]. The selected frequency bands
were 0-2.7Hz, 2.7-5.4Hz, 5.4-10.8Hz, 10.8-21.7Hz, 21.7-43.4Hz, and 43.4-86.8Hz [56].
Table 1 shows the segment-based classification accuracy, sensitivity, and specificity of
the LDA classifier and the three DL models for 10-fold CV. The DL models achieved
better performance than the LDA classifier. Among the three DL models, the CNN
and LSTM models showed superior overall performance, while the DNN model had a
limited ability in discriminating the preictal and ictal phases.

14
Figure 4. Performance of segment-based stage classification accuracy of each model.
DNN uq, CNN uq, LSTM uq are the unquantized models. DNN, CNN, and LSTM are
the quantized models. The box tops indicate 75th percentiles, box bottoms indicate
25th percentiles, solid lines indicate medians, whiskers indicate the span of the data,
and dots show data points (from all 24 cases).
Table 1. Segment-based Classification in 10-fold CV using Quantized Models
LDA DNN CNN LSTM
Overall Accuracy 62.07% 64.55% 89.21% 90.94%
Sensitivity
Ictal 83.67% 82.52% 96.59% 97.30%
Preictal 49.24% 51.83% 88.22% 91.46%
Interictal 55.53% 60.66% 83.67% 85.33%
Avg. 62.81% 65.00% 89.50% 91.53%
Specificity
Ictal 83.05% 88.89% 97.79% 98.68%
Preictal 73.55% 75.29% 90.47% 91.58%
Interictal 74.94% 83.75% 96.32% 97.00%
Avg. 77.18% 82.64% 94.86% 95.75%
3.2. Model Performance Evaluation
Real-time seizure event detection and prediction using the trained, quantized DL
models were simulated by streaming the selected EEG time series to the MCU in a
continuous manner. Figure 5 shows an illustrative example of the inferred scores of
ictal (Score [Ictal] in Algorithm 1) and preical (S core [P reictal] in Algorithm 1) phases
calculated by each model at each instance in time. The selected input EEG channels
for the CNN and LSTM models are shown at the top of Figure 5. Only the upper 5
channels were used as inputs to the DNN model. All three models successfully detected

15
the seizure event within a 5 sec window around the marked start time (Figure 5 (a)).
Furthermore, all three models successfully predicted the seizure within a 40 min horizon
prior to the actual onset time (Figure 5 (b)). In this example, the CNN model showed
the best robustness against false positives in predicting seizures.
Figure 5. The detection and prediction of one seizure event of patient chb01 using
the three DL models and sliding WMV algorithm. The genuine seizure occurred at
2:13 pm as annotated by clinical experts. Selected EEG channels used by the models
are shown at the top (the DNN model used only the upper 5 channels). The scores of
ictal (a) and preictal (b) phases calculated by the WMV algorithm are plotted. The
normalized score is coded by color, with dark blue being the lowest and dark red being
the highest. The earliest detection of each model is marked with a red arrow.
The performance of event-based seizure detection and prediction was tested with
LOOCV. The performance of the WMV algorithm was compared with a traditional
moving average. The WMV algorithm improved the seizure detection FPR from
0.745h−1to 0.169h−1and the seizure prediction FPR from 2.341h−1to 0.710h−1. Figure
6 shows the box chart of each model’s performance in seizure detection and prediction
tasks in LOOCV. For seizure detection, the LSTM model achieved the highest average
sensitivity (97.61%) and the lowest FPR (0.071 h−1), which corresponds to one false
alarm in every 14.1 hours. For seizure prediction, the LSTM and CNN models achieved
a comparable sensitivity above 90%. The CNN model had a better average FPR (0.204
h−1), which corresponds to one false alarm in every 4.9 hours. The performance of the
DNN model was relatively poor in seizure prediction tasks. The average and median
sensitivities and FPRs of each model are summarized in Table 2.
Finally, each model’s memory size, inference execution time, and power
consumption are shown in Figure 7. The results before and after quantization were
plotted for comparison. The memory size reflects the actual hardware implementation

16
Figure 6. Performance of event-based seizure detection (a-1 & a-2) and seizure
prediction (b-1 & b-2) in LOOCV.
Table 2. Performance of Event-based Seizure Detection and Prediction in LOOCV
Detection Prediction
Sensitivity FPR (h−1) Sensitivity FPR (h−1)
Avg. Median Avg. Median Avg. Median Avg. Median
DNN 87.36% 85.71% 0.169 0.140 76.66% 75.00% 0.710 0.474
CNN 96.70% 100% 0.102 0.084 90.66% 90.00% 0.204 0.168
LSTM 97.61% 100% 0.071 0.063 90.72% 90.00% 0.241 0.227
including the code overhead for data handling. The data transfer time via the USB port
was excluded from the inference time since this would not be present in autonomous
neural implants. The CPU core was put in sleep mode after executing the inference,
and the power consumption was measured directly from the power supply. The reported
power consumption is an average within the segment period. The DNN model required
the largest memory size, while the CNN model required the least. This is mainly
because of the limited convolutional kernel size used in the CNN model. The LSTM

17
model had the longest inference time, but it was still within its segment period. From
the perspective of power consumption, the DNN and CNN models were comparable,
while the LSTM model consumed the most.
Figure 7. Each deep learning model’s (a) memory size, (b) inference execution time,
and (c) power consumption. DNN uq, CNN uq, LSTM uq are the unquantized models.
DNN, CNN, and LSTM are the quantized models.
4. Discussion
Each of the three DL models developed in this work has its own strengths and
weaknesses. The DNN model achieved the shortest inference time with minimum power
consumption. The CNN model achieved a balanced performance with moderate power
consumption and the smallest memory cost. The LSTM model achieved the best overall
performance (e.g. highest seizure detection sensitivity and lowest FPR) at the expense
of relatively long inference time and high power consumption. The optimal choice of
model mainly depends on the specific application as well as the available hardware
resources of the neural implant, including the batteries.
Model quantization improved inference time, memory size, and power consumption
without sacrificing inference performance. On the other hand, model compression and
channel pruning may require iterations of retraining. Layer size should be scaled with
the sampling rate for capturing temporal features. Cascading DL models with a second
stage sliding window-based algorithm, such as the sliding window based WMV algorithm
implemented in this work, may compensate for the limitation of small input data buffer
size of the edge DL models. Optimizing the design jointly for available hardware
resources and real-time operational requirements is the key to achieving a satisfactory
overall performance.
For event-based seizure detection, the CNN and LSTM models achieved an average
high sensitivity of 96.7% and 97.61%, and low FPR at 0.102 and 0.071, respectively. The

18
results suggest that they are potential candidates for real-time, closed-loop therapeutic
systems. For seizure prediction, the performance of the DNN model was limited. We
compare the seizure prediction performance of this work with prior studies (Table 3).
Importantly, most of the listed studies assumed no time and computational resource
limitations. The highest performances were achieved at the cost of sophisticated,
hand-crafted, patient-specific feature engineering [9]. However, such complicated
processing typically prevents its application in embedded devices for real-time, closed-
loop inference. The relatively good hardware-based seizure prediction performance
of our CNN and LSTM models compared to prior software implementations provides
support for their use in future therapies.
One caveat to seizure prediction is that more advanced notice is not always better,
unless a seizure prediction horizon (SPH), which is a seizure-free warning period between
the alarm and the actual seizure onset, can be guaranteed by the algorithm [57].
Otherwise, an early warning may increase the anxiety of the patient given that a seizure
may or may not happen any time within a long period after the warning. Including an
accurate SPH significantly increases the difficulty in prediction [57]. It is not applicable
for the light-weight DL models developed in this work. It should be noted, however,
that the proposed edge DL inference paradigm doesn’t preclude the functionality of
uploading data to an external system including cloud-based computing resources for
further analysis. Periodical diagnosis and model updating may be necessary for clinical
adoption. But since these operations are not continuous during everyday use, data
encryption can be reinforced and the peaking power dissipation during these short
periods is not a big concern.
Another caveat to the results is that the EEG signals in the selected data set
were recorded noninvasively using scalp electrodes. Therapeutic neural implant devices
typically acquire EEG signals intracranially with a much higher signal-to-noise ratio
(SNR) [58]. The performance of DL algorithms typically improves with higher SNR
signals, since more subtle neural features can be unveiled in these recordings [31]. We
expect the performance of the models can be further improved, or the size and power
consumption can be reduced if intracranial EEG recordings are used as the input.
Although this work uses a general-purpose, off-the-shelf MCU as the edge hardware
platform, the design and optimization methods are applicable to ASIC development for
clinical neural implants. The MCU core can be directly integrated into an ASIC design
with the required memory or the models can be directly synthesized in the register-
transfer level for minimizing the design overhead. Furthermore, in-memory or near
memory computational techniques can be used to further reduce the power consumption
of repeated multiply-accumulate (MAC) operations [22, 23]. An ASIC system-on-chip
(SoC) that integrates analog neural interfaces (e.g. neural recorders and stimulators),
digital DL inference module, power management and wireless communication modules
can achieve the optimal power consumption and device footprint for chronic neural
implants.
Finally, neural implants with different clinical purposes use various types of input

19
neural signals, inference model complexities, and control strategies. Our work may serve
as a reference for related studies on cognitive monitoring, sleep interventions, BMI, and
other applications with closed-loop neuromodulation or neuroprosthetic control. The
generalizability of our edge DL approach depends on the model complexity required to
achieve high performance in these different applications. While DL model complexity
has been defined in different ways, it is specifically the speed of inference, along with
memory requirements, that is critical for resource-restrained edge devices [59]. To assess
generalizability, we sought to compare our models’ inference speed with that of a sample
of prior DL studies in different neural application domains. Elapsed inference time (Fig.
7 (b)) is rarely reported and is hardware dependent. A hardware-independent measure is
the number of required computations (i.e. MAC operations) based on model architecture
[60]. Restricting the comparison to CNN models, which are used in the majority
of neural applications [5], the total number of operations in the Conv and FC layers
provide a reasonable estimate of complexity (see Appendix). Our CNN model required
approximately 2.4M MACs to process the 256 ×9 input matrix. This complexity is
higher than many prior CNN models used for BMI applications like P300 detection [61],
steady-state evoked potential classification [62], and attentive state detection [63], which
ranged from about 0.1M to 0.5M MACs. Although CNN models for other tasks like sleep
scoring [64] can be significantly more complex, on the order of 10M to 100M MACs,
we suggest that edge DL systems are a realistic option for many BMI applications.
Since DL eliminates the domain-specific manual feature selection used in conventional
algorithms, research progress and technological advancement in one specific application
area should readily generalize to other applications.

20
Table 3. Comparison with Prior Seizure Prediction Studies
Year Publication database Features Algorithm Sensitivity FPR (h−1)Prediction
Duration
2013 Li et al. [65] Freiburg (21 cases) spike rate Threshold 72.7% 0.11 50 min
2016 Zhang et al. [66] CHB-MIT (17 cases) PSD ratio SVM 98.68% 0.047 50 min
2017 Chu et al. [67] CHB-MIT (13 cases) Fourier Transform, PSD Threshold 83.33% 0.392 86 min
2017 Alotaiby et al. [68] CHB-MIT (24 cases) Spatial pattern statistics LDA 81% 0.47 60 min
2017 Arabi et al. [69] Freiburg (10 cases) Uni-/bivariate features Rule-based 86.7% 0.126 30 min
2018 Khan et al. [38] CHB-MIT (15 cases) Wavelet Transform CNN 87.80% 0.142 10 min
2018 Truong et al. [31] CHB-MIT (13 cases) Short-time Fourier Transform CNN 81.20% 0.16 30 min
2018 Tsiouris et al. [9] CHB-MIT (24 cases) Wavelet Transform, PSD, statistics, etc. LSTM 99.60% 0.006 30 min
2018 Shahbazi et al. [70] CHB-MIT (14 cases) Short-time Fourier Transform LSTM 98.2 % 0.13 45 min
2019 Affes et al. [71] CHB-MIT (24 cases) Fourier Transform C-GRNN 89.0% 1.6 35 min
2020 Abiyev et al. [39] CHB-MIT (7 cases) Raw EEG CNN 97.67% 1.97 N/A
This work CHB-MIT (15 cases) Raw EEG
DNN
CNN
LSTM
76.66%
90.66%
90.72%
0.710
0.204
0.241
40 min
40 min
40 min

21
5. Conclusion
In this work, we developed edge DL models and investigated their potential utility for
future clinical applications involving neural implants. We adopted three commonly
used DL architectures (DNN, CNN, and LSTM) and optimized the models for
deployment in resource-restrained hardware. Using the CHB-MIT database, we
show that edge DL inference can achieve comparable performance in epileptic seizure
detection to many prior implementations that had no time and computational resource
limitations. Our results suggest that edge DL inference is a promising option for closed-
loop neuromodulation, with superior robustness and security compared to wireless
communication-based solutions. While clinical studies are needed to confirm the efficacy
of this paradigm, this work demonstrates the feasibility and potential advantages. We
envision that the next generation of clinical neural implants could leverage edge DL
inference to greatly improve their therapeutic benefit.
6. APPENDIX
6.1. Processing and activation functions
In the FC layers, the processing function of each neuron jis given by:
yj=fR(
n
X
i=1
wj,i ·xi+bj) (6)
where xiis from the previous layer, wj,i is the weight factor, bjis the bias term, nis
the number of neurons in the previous layer, and fR(x) is the activation function. The
fR(x) used in this work is the rectified linear unit (ReLU) function. ReLU is a piecewise
linear function as given by:
fR(x) = max{0, x}(7)
The processing functions of each LSTM cell jare given by:
fj=fG(Wh,f,j ·hj−1+Wx,f,j ·xj+bf,j ) (8)
ij=fG(Wh,i,j ·hj−1+Wx,i,j ·xj+bi,j ) (9)
oj=fG(Wh,o,j ·hj−1+Wx,o,j ·xj+bo,j ) (10)
˜cj=fS(Wh,c,j ·hj−1+Wx,c,j ·xj+bc,j ) (11)
cj=fj·cj−1+ij·˜cj(12)
hj=oj·fS(cj) (13)
where xjis the output from the previous layer, hjis the hidden state, cjis the memory
state ( ˜cjis the candidate), ijis the input gate, ojis the output gate, and fjis the forget
gate. W’s and b’s are the weight and bias terms for each neuron in the LSTM cell.

22
The gate activation function fG(x) is given by:
fG(x) = max{0, min{1,x
5+1
2}} (14)
The state activation function fS(x) is given by:
fS(x) = x
1 + |x|(15)
6.2. Multiply-accumulate operation estimates
The number of MAC operations for a CNN can be estimated as follows. For FC layers:
#MACf c =Cin ∗Cout (16)
where Cin is the number of input channels (neurons) and Cout is the number of output
channels. For Conv layers:
#MACconv =Cin ∗mout ∗nout ∗Cout ∗h∗v(17)
where mout ∗nout is the size of the output feature map, Cout is again the number of
output channels (equivalent to the number of kernels), and h∗vis the size of each
kernel. Max pooling and activation operations are ignored as they typically contribute
very little to the total MAC count.
References
[1] Sun F T and Morrell M J 2014 Neurotherapeutics 11 553–563
[2] Mak J N and Wolpaw J R 2009 IEEE Reviews in Biomedical Engineering 2187–199
[3] Lawhern V J, Solon A J, Waytowich N R, Gordon S M, Hung C P and Lance B J 2018 Journal of
neural engineering 15 056013
[4] Baldassano S, Zhao X, Brinkmann B, Kremen V, Bernabei J, Cook M, Denison T, Worrell G and
Litt B 2019 Journal of Neural Engineering 16
[5] Roy Y, Banville H, Albuquerque I, Gramfort A, Falk T H and Faubert J 2019 Journal of Neural
Engineering 16 051001
[6] LeCun Y, Bengio Y and Hinton G 2015 Nature 521 436–444
[7] Lotte F, Bougrain L, Cichocki A, Clerc M, Congedo M, Rakotomamonjy A and Yger F 2018
Journal of Neural Engineering 15 031005
[8] Schirrmeister R T, Springenberg J T, Fiederer L D J, Glasstetter M, Eggensperger K, Tangermann
M, Hutter F, Burgard W and Ball T 2017 Human Brain Mapping 38 5391–5420
[9] Tsiouris K M, Pezoulas V C, Zervakis M, Konitsiotis S, Koutsouris D D and Fotiadis D I 2018
Computers in biology and medicine 99 24–37
[10] Mahmood M, Mzurikwao D, Kim Y S, Lee Y, Mishra S, Herbert R, Duarte A, Ang C S and Yeo
W H 2019 Nature Machine Intelligence 1412–422
[11] Kiral-Kornek I, Roy S, Nurse E, Mashford B, Karoly P, Carroll T, Payne D, Saha S, Baldassano
S, O’Brien T et al. 2018 EBioMedicine 27 103–111
[12] Naufel S, Knaack G L, Miranda R, Best T K, Fitzpatrick K, Emondi A A, Van Gieson E and
McClure-Begley T 2020 Journal of Neuroscience Methods 332 108539
[13] Hartmann M, Hashmi U S and Imran A Transactions on Emerging Telecommunications
Technologies e3710
[14] O’Leary G, Groppe D M, Valiante T A, Verma N and Genov R 2018 IEEE Journal of Solid-State
Circuits 53 3150–3162

23
[15] Zhu B, Shin U and Shoaran M 2020 Closed-loop neural interfaces with embedded machine learning
(Preprint 2010.09457)
[16] Heller S, H¨ugle M, Nematollahi I, Manzouri F, D¨umpelmann M, Schulze-Bonhage A, Boedecker
J and Woias P 2018 Hardware implementation of a performance and energy-optimized
convolutional neural network for seizure detection 2018 40th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC) (IEEE) pp 2268–2271
[17] H¨ugle M, Heller S, Watter M, Blum M, Manzouri F, Dumpelmann M, Schulze-Bonhage A, Woias
P and Boedecker J 2018 Early seizure detection with an energy-efficient convolutional neural
network on an implantable microcontroller 2018 International Joint Conference on Neural
Networks (IJCNN) (IEEE) pp 1–7
[18] Akopyan F, Sawada J, Cassidy A, Alvarez-Icaza R, Arthur J, Merolla P, Imam N, Nakamura Y,
Datta P, Nam G J et al. 2015 IEEE transactions on computer-aided design of integrated circuits
and systems 34 1537–1557
[19] Cass S 2019 IEEE Spectrum 56 16–17
[20] Davies M, Srinivasa N, Lin T H, Chinya G, Cao Y, Choday S H, Dimou G, Joshi P, Imam N, Jain
Set al. 2018 IEEE Micro 38 82–99
[21] Wang X, Han Y, Wang C, Zhao Q, Chen X and Chen M 2019 IEEE Network 33 156–165
[22] Si X, Chen J J, Tu Y N, Huang W H, Wang J H, Chiu Y C, Wei W C, Wu S Y, Sun X, Liu R
et al. 2019 IEEE Journal of Solid-State Circuits 55 189–202
[23] Jia H, Valavi H, Tang Y, Zhang J and Verma N 2020 IEEE Journal of Solid-State Circuits
[24] Chen J and Ran X 2019 Proceedings of the IEEE 107 1655–1674
[25] Liu X, Zhang M, Richardson A G, Lucas T H and Van der Spiegel J 2016 IEEE transactions on
biomedical circuits and systems 11 729–742
[26] Heck C N, King-Stephens D, Massey A D, Nair D R, Jobst B C, Barkley G L, Salanova V, Cole
A J, Smith M C, Gwinn R P, Skidmore C, Van Ness P C, Bergey G K, Park Y D et al. 2014
Epilepsia 55 432–441
[27] Shoeb A H 2009 Application of machine learning to epileptic seizure onset detection and treatment
Ph.D. thesis Massachusetts Institute of Technology
[28] Goldberger A L, Amaral L A, Glass L, Hausdorff J M, Ivanov P C, Mark R G, Mietus J E, Moody
G B, Peng C K and Stanley H E 2000 circulation 101 e215–e220
[29] Homan R W, Herman J and Purdy P 1987 Electroencephalography and clinical neurophysiology
66 376–382
[30] Litt B and Echauz J 2002 The Lancet 27 421–424
[31] Truong N D, Nguyen A D, Kuhlmann L, Bonyadi M R, Yang J, Ippolito S and Kavehei O 2018
Neural Networks 105 104–111
[32] Daoud H and Bayoumi M A 2019 IEEE transactions on biomedical circuits and systems 13 804–813
[33] Japkowicz N and Stephen S 2002 Intelligent data analysis 6429–449
[34] Nair V and Hinton G E 2010 Rectified linear units improve restricted boltzmann machines ICML
[35] Goodfellow I, Bengio Y, Courville A and Bengio Y 2016 Deep learning vol 1 (MIT press Cambridge)
[36] Krizhevsky A, Sutskever I and Hinton G E 2012 Imagenet classification with deep convolutional
neural networks Advances in neural information processing systems pp 1097–1105
[37] Kim Y 2014 arXiv preprint arXiv:1408.5882
[38] Khan H, Marcuse L, Fields M, Swann K and Yener B 2018 IEEE Transactions on Biomedical
Engineering 65 2109–2118
[39] Abiyev R, Arslan M, Idoko J B, Sekeroglu B and Ilhan A 2020 Applied Sciences 10 4089
[40] Gers F A, Schmidhuber J and Cummins F 1999 Neural Computation
[41] Graves A, Jaitly N and Mohamed A r 2013 Hybrid speech recognition with deep bidirectional lstm
2013 IEEE workshop on automatic speech recognition and understanding (IEEE) pp 273–278
[42] He Y, Zhang X and Sun J 2017 Channel pruning for accelerating very deep neural networks
Proceedings of the IEEE International Conference on Computer Vision pp 1389–1397
[43] Esteller R, Echauz J and Tcheng T 2004 Comparison of line length feature before and after brain

24
electrical stimulation in epileptic patients The 26th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society vol 2 (IEEE) pp 4710–4713
[44] Logesparan L, Casson A J and Rodriguez-Villegas E 2012 Medical & biological engineering &
computing 50 659–669
[45] Han S, Mao H and Dally W J 2015 arXiv preprint arXiv:1510.00149
[46] Le Gallo M, Sebastian A, Mathis R, Manica M, Giefers H, Tuma T, Bekas C, Curioni A and
Eleftheriou E 2018 Nature Electronics 1246–253
[47] Kingma D P and Ba J 2014 arXiv preprint arXiv:1412.6980
[48] Kearns M and Ron D 1999 Neural computation 11 1427–1453
[49] Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, Devin M, Ghemawat S, Irving G, Isard
Met al. 2016 Tensorflow: A system for large-scale machine learning 12th US EN I X symposium
on operating systems design and implementation (OSDI 16) pp 265–283
[50] Semiconductor N 2019
[51] Inc A 2009 URL https://developer.arm.com/documentation/ddi0439/b/
[52] Keil URL https://www2.keil.com/mdk5
[53] MathWorks 2020
[54] Inc G 2019 URL https://www.tensorflow.org/lite/inference with metadata/codegen
[55] Liu X, Zhang M, Subei B, Richardson A G, Lucas T H and Van der Spiegel J 2015 IEEE
transactions on biomedical circuits and systems 9248–258
[56] Subasi A and Gursoy M I 2010 Expert systems with applications 37 8659–8666
[57] Winterhalder M, Maiwald T, Voss H, Aschenbrenner-Scheibe R, Timmer J and Schulze-Bonhage
A 2003 Epilepsy & Behavior 4318–325
[58] Youngerman B E, Khan F A and McKhann G M 2019 Neuropsychiatric Disease and Treatment
15 1701–1716
[59] Sze V, Chen Y, Yang T and Emer J S 2017 Proceedings of the IEEE 105 2295–2329
[60] Taghavi M and Shoaran M 2019 Hardware complexity analysis of deep neural networks and decision
tree ensembles for real-time neural data classification 2019 9th International IEEE/EMBS
Conference on Neural Engineering (NER) pp 407–410
[61] Cecotti H and Graser A 2011 IEEE Transactions on Pattern Analysis and Machine Intelligence
33 433–445
[62] Kwak N, Muller K and Lee S 2017 PLoS One 12 e0172578
[63] Fahimi F, Zhang Z, Goh W B, Lee T S, Ang K K and Guan C 2019 Journal of Neural Engineering
16 026007
[64] Tsinalis O, Matthews P M, Guo Y and Zafeiriou S 2016 arXiv (Preprint 1610.01683)
[65] Li S, Zhou W, Yuan Q and Liu Y 2013 IEEE transactions on neural systems and rehabilitation
engineering 21 880–886
[66] Zhang Z and Parhi K K 2016 IEEE transactions on biomedical circuits and systems 10 693–706
[67] Chu H, Chung C K, Jeong W and Cho K H 2017 Computer methods and programs in biomedicine
143 75–87
[68] Alotaiby T N, Alshebeili S A, Alotaibi F M and Alrshoud S R 2017 Computational intelligence
and neuroscience 2017
[69] Aarabi A and He B 2017 Clinical Neurophysiology 128 1299–1307
[70] Shahbazi M and Aghajan H 2018 A generalizable model for seizure prediction based on deep
learning using cnn-lstm architecture 2018 IEEE Global Conference on Signal and Information
Processing (GlobalSIP) (IEEE) pp 469–473
[71] Affes A, Mdhaffar A, Triki C, Jmaiel M and Freisleben B 2019 A Convolutional Gated Recurrent
Neural Network for Epileptic Seizure Prediction Lecture Notes in Computer Science (including
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) vol 11862
LNCS pp 85–96