Using Sparse Representation of EEG Signal from a Shallow Sparse

Autoencoder for Epileptic Seizure Prediction

Gul Hameed Khan, Nadeem Ahmad Khan, Wala Saadeh and Muhammad Awais Bin Altaf

Lahore University of Management Sciences (LUMS), Lahore, Pakistan

Keywords:

Autoencoder, Electroencephalogram (EEG), Feature Selection, Seizure Prediction, Signal Sparsity.

Abstract:

Patients with epilepsy are affected with unexpected seizure events, which signiﬁcantly diminish their quality

of life. It is crucial to evaluate whether an epileptic patient’s brain state is indicative of a possible seizure

occurrence so that necessary therapy or alarm can be generated on time. If seizures could be predicted before

the onset, interventions may be applied to avoid further damage during seizure attack, and patients could

take medications or other treatments to prevent seizures from occurring. This research describes a patient-

speciﬁc technique for predicting epileptic seizures based on a hybrid model. Single layer sparse autoencoder

is trained to obtain a aparse representation of the scalp electroencephalogram (EEG) signals. SVM classiﬁer

is used to categorize the sparse signal as inter-ictal or pre-ictal. Individual EEG channel analysis for seizure

prediction are presented. In addition, various hidden sizes of the autoencoder for optimal sparse representation

are anlyzed.The proposed model evaluates 13 patients from the CHB-MIT dataset and obtains a sensitivity of

98% and an area under the curve (AUC) of 98%. We have evaluated the performance of our hybrid strategy

to both deep learning models and conventional procedures. The proposed method outperforms current seizure

prediction techniques, proving its efﬁcacy.

1 INTRODUCTION

Epilepsy is a prevalent neurological disorder classi-

ﬁed as the second most likely neural disease of the

brain, and the number of epileptic patients has in-

creased dramatically in recent years. A spontaneous

epileptic seizure is characterized by the transient and

instantaneous abnormal disruption of brain neurons

(Yang et al., 2021b). Seizures caused by epilepsy

can produce serious disruptions in a patient’s emo-

tions, behavior, movements, and awareness, and can

result in severe damage or even death. The prema-

ture death rate of these patients is two to three times

that of disease-free persons, placing a signiﬁcant bur-

den on the patients, their families, and the community

(Rasheed et al., 2021).

The hypothesis underlying seizure prediction is

that there exists a transition state (preictal) between

the interictal (normal state) and the ictal (seizure

state) (Truong et al., 2018). This notion is supported

by an abundance of clinical evidence. Consequently,

researchers have devoted a substantial amount of ef-

fort over the past few decades to attempting to antici-

pate epileptic seizures based on intracranial EEG and

scalp EEG signals, with the latter being more practi-

cal for clinical application (Zhang and Li, 2022).

Early prediction of epileptic seizures provides suf-

ﬁcient time before the seizure really occurs; this is

extremely important because the treatment can pre-

vent the attack. The problem of seizure prediction (or

forecasting) is detecting seizure symptoms and deter-

mining if the patient is on the verge of an attack (Ryu

and Joe, 2021). Detecting seizure symptoms is analo-

gous to identifying the inter-ictal and pre-ictal phases,

according to the phase distinction.

Therefore, similar to conventional signal classiﬁ-

cation, numerous methods for seizure prediction have

been proposed in the literature based on feature ex-

traction and classiﬁcation. For forecasting epileptic

seizures from EEG readings, machine learning ap-

proaches and computational algorithms are applied

with different feature extraction techniques. For ex-

ample, (Usman et al., 2017) used empirical mode de-

composition (EMD) for preprocessing and retrieved

time and frequency domain information for training

a prediction model. A patient-speciﬁc technique for

predicting epileptic seizures based on the common

spatial pattern (CSP) feature extraction of scalp EEG

signals is presented in (Alotaiby et al., 2017). Seizure

prediction framework using bag-of-wave (BoWav)

Khan, G., Khan, N., Saadeh, W. and Altaf, M.

Using Sparse Representation of EEG Signal from a Shallow Sparse Autoencoder for Epileptic Seizure Prediction.

DOI: 10.5220/0011813400003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 4: BIOSIGNALS, pages 125-132

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

125

feature extraction and synchronization patterns is pro-

posed by (Cui et al., 2018). However, conventional

features are generally chosen empirically after a brief

period of time, therefore the key characteristics may

be ignored (Cui et al., 2018).

In a relatively new trend, deep learning algo-

rithms are being used in medical image and signal

processing. Interest has been drawn to deep learn-

ing techniques due to their robust feature extraction

capacity. This is possible because of advances in

computing power and the availability of big data.

These algorithms have a lot of potential and can have

a big impact because, in most cases, their perfor-

mance is better than what was possible with tradi-

tional machine learning techniques. Researchers have

mostly been interested in convolutional neural net-

works (CNNs) for seizure prediction (Truong et al.,

2018). This is likely because CNNs have been ex-

tensively used in image processing and are therefore

better known and more established in the research

community. Moreover, Long Short-Term Memory

(LSTM), stacked autoencoders (SAE) and convolu-

tional autoencoders (CAE) are developed for classify-

ing EEG data resulting in high accuracy systems (Ryu

and Joe, 2021), (Tautan et al., 2019). These meth-

ods, however, have a signiﬁcant energy consumption

as well as a huge number of parameters and hardware

resources (Zhao et al., 2020). Therefore, these tech-

niques cannot be used with small, low-power wear-

able or implantable medical equipment. To continu-

ously update the epileptic patients, the devices should

run in real-time.

There have been signiﬁcant advancements in the

ﬁeld of epileptic seizure prediction, with promising

outcomes reported by various methodologies. How-

ever, these existing systems typically utilize mul-

tichannel EEG signals for pre-ictal and inter-ictal

recognition. These methods extract features from

multiple EEG channels or analyse all available chan-

nels collectively in order to categorize multi-channel

epochs of short time intervals as pre-ictal or inter-

ictal for seizure prediction. Existing approaches do

not analyze individual EEG channels for the identiﬁ-

cation of abnormalities or signal variations that lead

to these state shifts. This also holds true for epilep-

tic seizure detection techniques. EEG analysis at the

level of a single channel is important for various rea-

sons: Initially, a neurologist would like to construct

his analysis of multi-channel EEG input bottom up

(from individual signal to group level). The indi-

vidual signal-level evidence builds the multichannel

epoch-level determination. Second, analysis at the

level of individual signals can reveal the potential of

each EEG channel for seizure prediction depending

on the type of seizure. Exploring multi-channel EEG

analysis by combining single channel evaluations is

an effective approach. This technique also presents

the possibility of exploring the use of minimal EEG

channels for seizure prediction (or detection) rather

than using data from the entire scalp. In addition,

the advancement of such a system is advantageous for

the creation and usage of wearable devices and sensor

networks, resulting in greater wearing comfort and a

more compact form factor due to a reduction in pro-

cessing demand.

This paper presents an approach for epileptic

seizure prediction using a hybrid model comprising of

a shallow sparse autoencoder (AE) and support vec-

tor machine (SVM) classiﬁer. We therefore elaborate

these methods as follows:

1.1 Shallow Sparse Autoencoder

The basic components of an AE are: an encoder and a

decoder. To rebuild the original dataset x from the en-

coder’s representation y, the decoder is conﬁgured to

minimize the difference between x and ˆx, as shown in

Fig. 1. The encoder’s output y represents the reduced

representation of x, which consists of n samples. To

be more precise, an encoder is a function f (.) that

converts a given input x into some unknown represen-

tation y. The procedure is stated as follows:

A shallow AE with a single hidden layer compris-

ing of n neurons in the input/output layer and m hid-

den neurons is developed with m < n as shown in Fig.

1. Encoder layer is evaluated with sharing weights

∈ R

n×m

and biases vector b

∈ R

. The decoder

layer reconstructs the signal with weights W

∈ R

m×n

and biases vector b

∈ R

. These weights and biases

of the model are calculated over signal reconstruction

error instead of classiﬁcation results. The scaled con-

jugate gradient (SCG) was designated to update these

weights and bias values. Encoding process for input

EEG signal x ∈ R

is modeled as:

y = f (W

x + b

) (1)

Where f (

) represents the activation function in

the encoder neurons. Two inputs, a bias vector b and

a matrix W , are used to conﬁgure the decoder. In most

cases, the activation function is selected based on

the characteristics of the available data (Meng et al.,

2017). However, in contrast to non-linear functions,

linear activation helps in building a system with low

computational cost. For this reason, both the encod-

ing and decoding procedures will use linear activa-

tions. For the encoder’s activation function, the satu-

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

126

rating linear transfer function (Satlin) is described as:

satlin(x) =







−1 x < −1

0 −1 ⩽ x ⩽ 1

1 x > 1







(2)

Finding the values of the training parameters θ =

(W, b

, b

) that minimize the reconstruction loss for a

given dataset is the goal of AE training.

Θ = min

L(x, ˆx) = min

L(x, f (W

x + b)) (3)

The reconstruction loss L in AE’s training phase

is typically derived from the square of the error:

L(θ) =

∑

k=1

(x − ˆx)

(4)

Here, both the n and m represents the number of

samples in the signal. The decoder layer reconstructs

the sparse vector y ∈ R

to its original form as:

ˆx = W

y + b

(5)

AE’s objective function reconstructs the input to

its original form. High weights of hidden layers make

the generated features more dependent on the net-

work structure rather than the input data. Therefore,

to avoid this complexity, sparse AE imposes weight-

decay regularization so as to keep neuron weights

small.

Θ = αmin

L(x, ˆx)+ λ||W||

(6)

Here, α is the scale parameter to control the

weights of the data reconstruction loss. We used L

regularization ||W ||

to ensure weight matrix W hav-

ing small elements. Hyper parameter λ is incorpo-

rated to control the regularization strength. Transfer

function to compute decoder layer’s output is linear

function (Purelin) f (x) = x. These functions are ap-

plied to each signal sample using the Mean Square

Error (MSE) loss function, presented in equation 4.

MSE is used in the AE’s training phase to compare the

original and reconstructed signals. Encoding yields

a sparse representation of the signal, while decoding

generates a reconstruction of the signal to its original

form. The decoding is only added in training phase to

ensure the best sparse representation of the data pos-

sible.

Sparsity level of the AE for input signal of length

n and hidden size m is given as:

SparsityLevel =

(7)

1.2 Gaussian SVM

The Gaussian kernel SVM is frequently employed

due to its excellent performance, and is frequently re-

garded as one of the most effective techniques when

it comes to supervised learning (Yang et al., 2021a).

To ensure that models perform well, it is necessary

to precisely determine hyper-parameters such ker-

nel width and penalty factor. The performance of

the Gaussian kernel SVM is thoroughly examined in

(Yang et al., 2021a) when the hyper-parameters are

set to their most extreme values (0 or ∞). The Leave-

One-Out (LOO) approach’s local density and preci-

sion serve as the cornerstone of the parameter op-

timization strategy, which is suggested to increase

computing efﬁciency. The kernel width of each sam-

ple is dependent on the local density necessary to en-

sure a higher separability in feature space, and the

LOO approach reﬁnes the grid search to identify the

ideal penalty value. The proposed method is evalu-

ated for validity by comparison to the grid method

(Yang et al., 2021a).

The impact of various kernel functions on SVM

characteristics is highly variable (Yang et al., 2021a).

Here, we used the Gaussian kernel presented in equa-

tion 8. An optimized hyperplane in kernel space is

generated by a Gaussian function of kernel scale 4

and kernel offset 0.1. Training instances are separable

for these parameters and yield optimal results. The

feature space that training data is mapped to is deter-

mined by the kernel width. As a result, the success of

an SVM’s training procedure is profoundly affected

by its accuracy.

k(x

, x

) = exp(−

||x

− x

2σ

) (8)

Some of the key contributions of the proposed

method are:

• We have used sparse representation generated by

the AE for seizure prediction

• Our framework provides individual EEG channel

analysis for seizure prediction

• Using autoencoder with only one hidden layer

will reduce the computational complexity of the

algorithm

• Multiple Hidden sizes of autoencoder are ana-

lyzed to obtain optimum sparse representation ca-

pable of effectively classifying the data as inter-

ictal or pre-ictal

• Our method outperforms state of the art seizure

prediction techniques

Using Sparse Representation of EEG Signal from a Shallow Sparse Autoencoder for Epileptic Seizure Prediction

127

2 PROPOSED METHODOLOGY

AE based feature extraction method achieved great

success in generating implicit features of high dimen-

sional data (Meng et al., 2017). AE use artiﬁcial neu-

ral networks to reduce dimensionality by minimizing

the reconstruction loss. Due to this, AE and its exten-

sions demonstrate a promising ability to extract mean-

ingful features, particularly in signal processing do-

main (Meng et al., 2017).

The fundamental processing architecture em-

ployed in this paper to predict epileptic seizures is

depicted in Figure 1. Firstly, a sparse AE with only

one hidden layer is trained to obtain a sparse signal

directly from the raw EEG data. SVM classiﬁer pro-

cesses these sparse signal samples to categorize the

data as inter-ictal or pre-ictal. Ability of various hid-

den sizes of the AE are analyzed for seizure predic-

tion.

Using the raw EEG data directly without any

transformation or feature extraction and AE with

only one hidden layer will reduce the computational

complexity of the algorithm. Avoiding the compli-

cated process for extracting features, a lot of memory

for storing high-precision parameters, and complex

arithmetic computations will reduce the hardware re-

sources requirement. It is desirable in wearable de-

vices that demand low-power consumption and real-

time operation. Therefore, the proposed methodol-

ogy is computationally efﬁcient, and its classiﬁcation

performance is comparable to that of existing seizure

prediction techniques.

EEG data is divided into 1-second segments for

each channel. These 1-second segments are fed into

the AE. Different seizure prediction methods in the

literature commonly use EEG data segments ranging

from 1 to 30 seconds in length (Zhang and Li, 2022),

(Khan et al., 2021). The proposed algorithm pro-

cesses one-second EEG segments using a single EEG

channel. The EEG dataset used in this work has pre-

viously been preprocessed for noise and artifact re-

moval. As a result, to minimize the computational

overhead of preprocessing, we use raw EEG data di-

rectly.

AE with single hidden layer processes each 1-sec

EEG trial to produce its sparse representation. In

the training phase, an encoder is used to determine

a sparse representation of the signal, and a decoder is

designated to restore the signal to its original form.

The sparse signal samples are used as input to classi-

ﬁer model to categorize the data as inter-ictal or pre-

ictal. However, the decoder part of the AE is dis-

carded in the testing phase and only the encoder is uti-

lized for seizure prediction. Numerous hidden sizes

of AE are analyzed to obtain an optimum sparse sig-

nal samples capable of classifying the data precisely.

Figure 1: Work ﬂow architecture of the proposed seizure

prediction method.

2.1 Evaluation Data

In order to evaluate the effectiveness of the proposed

seizure prediction system, we experimented with the

Children’s Hospital Boston-collected PhysioNet scalp

EEG database CHB-MIT. There are 163 seizure oc-

currences recorded from 23 pediatric patients with

916 hours of continuous scalp EEG (sEEG) monitor-

ing. Recordings from 22 participants are organized

into 23 cases and sampled at 256 Hz.

We evaluate the common 18 channels for each pa-

tient to ensure model consistency and patient wise

uniformity as in (Ryu and Joe, 2021), (Sun et al.,

2021), (Gao et al., 2022). Because there are numerous

patients in the experiment using different channels,

we used the same 18 channels that all patients had.

Therefore, we utilize 18 channels (“FP1-F7”, “F7-

T7”, “T7-P7”, “P7-O1”, “FP1-F3”, “F3-C3”, “C3-

P3”, “P3-O1”, “FP2-F4”, “F4-C4”, “C4-P4”, “P4-

O2”, “FP2-F8”, “F8-T8”, “T8-P8”, “P8-O2”, “FZ-

CZ”, “CZ-PZ”).

2.2 System Parameters

We experimented the proposed seizure prediction al-

gorithm with various hidden sizes of the AE. For a an

input size of 256, the encoder layer length varies from

64 to 8, corresponding to the sparsity levels 4 to 32 as

demonstrated in equation 7. Reducing the AE’s hid-

den size will decrease the computational cost of the

algorithm. In addition, analyzing the EEG signals to-

wards seizure prediction with different hidden sizes

of the AE will help to select an optimum hidden size

as we do not have a speciﬁc formula to build an AE

with the most appropriate hidden size.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

128

Table 1: Subjects information used in this study from the

CHB-MIT database.

Patient

No. of seizure

events

Inter-ictal Duration

(hrs)

1 7 14

2 3 23

3 6 22

5 5 14

9 4 46

10 6 26

13 5 14

14 5 5

18 6 24

19 3 25

20 5 20

21 4 22

23 5 13

Total 64 268.6

The time interval between at least 4 hours (hrs) be-

fore the start of a seizure and 4 hrs after it has ended

is known as the inter-ictal phase. CHB-MIT dataset

demonstrates that many seizures can take place near

together. In the seizure prediction task, we are inter-

ested in forecasting a seizure episode that will occur

within roughly 30 minutes of the previous one. As a

result, we treat seizures that happen within 30 min-

utes of one another as a single seizure, with the ﬁrst

seizure’s onset serving as the start of the combined

seizure. Additionally, we only take into account pa-

tients with less than 10 seizures per day for the predic-

tion task because it is not necessary to predict seizures

occurrence for patients experiencing seizures on av-

erage every 2 hours. These parameters indicate that

13 participants have sufﬁcient data (at least three pri-

mary seizures and three hrs of interictal recording).

The subject ID, total number of seizure occurrences,

and length of the inter-ictal period for each subject are

included in the Table 1 thorough description of each

subjects’ information.

The Seizure Prediction Horizon (SPH) is the time

between the alert triggered in anticipation of the

seizure occurrence and the actual ictal state onset.

For an accurate forecast, a seizure must occur after

the SPH and before the Seizure Occurrence Period

(SOP), which is the estimated time span for seizure

occurrence. A false alarm will be generated if the pre-

diction algorithm produces a positive signal (a seizure

is imminent) but there is no seizure during the SOP.

After the alert has been triggered, the ideal therapeutic

application of SPH is giving the patient sufﬁcient time

to take preventative measures. A patient’s SPH must

be lengthy enough to allow for adequate safety mea-

sures once the alarm is activated. To comfort the pa-

Figure 2: Variations in Seizure prediction performance of

the proposed method using different hidden sizes (Sparsity

Levels) of the AE for channel ‘FzCz’ (best case).

tient, however, the SOP should not be overly lengthy.

In this research, we employ the SPH period of 10 min-

utes and the SOP period of 30 minutes.

Inter-ictal EEG recordings are more prevalent than

pre-ictal EEG recordings because the ictal state is un-

common across lengthy hours of EEG recordings. In

machine learning methods, it is commonly assumed

that data from different classes should be distributed

uniformly (Rasheed et al., 2021). A classiﬁer trained

on a greater number of examples for one class in

comparison to other classes will be biased and pre-

fer an uneven decision. To overcome the class imbal-

ance problem, numerous strategies are discussed in

the literature, including overlapping window (Truong

et al., 2018),(Dissanayake et al., 2021), generative ad-

versarial network (GAN) (Rasheed et al., 2021), and

synthetic minority over-sampling technique (SMOTe)

(Usman et al., 2021). We utilize SMOTe to produce

additional pre-ictal data during the training phase,

hence addressing the issue of data imbalance as de-

scribed in (Usman et al., 2021). By supplement-

ing pre-ictal state data with an overlapping window,

SMOTe can alleviate the class imbalance issue.

3 SYSTEM EVALUATION

3.1 Experimental Setup

Using the CHB-MIT database, the AE model for

sparse representation of the input signal is trained us-

ing 10 minutes of pre-ictal EEG data (1-second du-

ration segments) for each seizure. For example, as

indicated in Table 1, Patient 1 has experienced seven

seizures. Therefore, we extract 70 (7 × 10) minutes

of pre-ictal data and the same amount of data for

the inter-ictal phase for this patient. Various model

assessment strategies have been presented in the lit-

Using Sparse Representation of EEG Signal from a Shallow Sparse Autoencoder for Epileptic Seizure Prediction

129

Table 2: Patient wise classiﬁcation results of the proposed seizure prediction method with AE’s hidden size 64.

Patient ID

Channel ‘FzCz’ (Best Case) Channel ‘FP1F7’ (Worst Case)

Ac Sen Spe AUC Ac Sen Spe AUC

1 99.0 100 97.8 100 92.6 89.2 96.1 97.0

2 93.8 98.7 91.8 98.0 89.7 85.8 93.4 94.0

3 96.3 98.1 93.0 99.0 88.0 97.4 78.4 89.0

5 96.1 98.1 95.0 99.0 83.4 74.7 92.2 92.0

9 97.5 98.3 94.6 99.0 89.2 82.4 95.9 90.0

10 95.0 99.3 91.0 100 91.9 86.0 97.8 94.0

13 95.7 99.4 91.8 98.0 88.0 87.2 88.9 94.0

14 94.1 91.2 98.0 99.0 87.2 78.3 96.9 95.0

18 95.4 99.1 91.3 100 86.8 86.5 87.1 88.0

19 94.1 97.0 91.5 98.0 94.0 97.8 90.3 95.0

20 97.0 99.0 96.1 100 97.1 96.0 98.2 95.0

21 95.8 99.3 92.5 99.0 87.1 85.3 88.8 91.0

23 92.0 91.0 92.0 95.0 86.1 84.6 87.7 90.0

Average 96.0 98.0 93.5 98.8 89.3 87.1 92.0 92.6

erature, including patient-wise data partitioning into

train and test sets (Zhang et al., 2021a), k-fold CV

(Rasheed et al., 2021), (Ryu and Joe, 2021), and LOO

CV (Gao et al., 2022). As a result, for a fair compari-

son and to avoid over-ﬁtting, we evaluated the perfor-

mance of the proposed seizure prediction model using

the same 10-fold CV as the majority of existing ap-

proaches. Raw trials/segments of duration 1 second

were utilized without any preprocessing. The com-

pressed form of these segments was utilized to train

the 10-fold CV-based classiﬁer.

3.2 System Performance

The sampling rate of the dataset is 256, while the hid-

den size of AE is 64 (at sparsity level 4). This im-

plies that the sparse signal length is 64, which the

SVM classiﬁer uses to identify the data as pre-ictal

or inter-ictal. Among the 18 available EEG channels

of the dataset, we present the results of best and worst

performing channels. Channel ‘FzCz’ corresponds to

the the best case providing the highest average clas-

siﬁcation results for all the patients, whereas, chan-

nel ‘FP1F7’ is the worst case. Table 2 provides a

summary of the patient-speciﬁc seizure prediction re-

sults achieved with these two channels. Performance

metrics include prediction accuracy (Ac), sensitiv-

ity (Sen), speciﬁcity (Spe), and area under the curve

(AUC). Even when using a single EEG channel, the

proposed approach can produce results comparable to

those of previous studies on the prediction of seizures.

The prerequisite is the Neurologist’s selection of the

channel.

The proposed classiﬁcation approach is also eval-

uated using an AE with variable hidden sizes that gen-

erates sparse signals of varying lengths. Figure 2 il-

lustrates the average Ac, Sen, Spe, and AUC of the

proposed method for seizure prediction at each spar-

sity level. Statistical analysis of performance eval-

uation demonstrates that the proposed classiﬁcation

model can be utilized in a variety of system-required

scenarios. The highest sparsity level reported is 32,

which corresponds to compressing an input signal of

length 256 to merely 8 samples.

3.3 Performance Comparison

Table 3 illustrates the performance comparison be-

tween the proposed algorithm and other existing

epileptic seizure prediction approaches for the CHB-

MIT dataset. However, even if the same datasets were

used, the model performance would be affected by the

selection of samples by different methods, the wide

variety of channel conﬁgurations, the division of data,

the duration of the pre-ictal period, and training ap-

proaches (such as processing for unbalanced data) and

evaluation techniques (LOO, k-fold CV etc). We used

a framework that is quite relevant to the existing high

performance approaches in order to achieve a com-

parative analysis that is pretty fair.

Methods for classifying EEG at the segment level

(1-30 sec intervals) and providing classiﬁcation re-

sults in terms of accuracy, sensitivity, speciﬁcity, and

AUC are compared. In this table, we present the best

single-channel results. The AE utilized sparsity level

4. Classiﬁcation performance statistics demonstrate

that our method yields comparable results to the state-

of-the-art, particularly in terms of prediction sensitiv-

ity, which is the most crucial factor to ensure that no

seizure event prediction is missed.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

130

Table 3: Average classiﬁcation results comparison with some recent seizure prediction methods for CHB-MIT database.

Ref. - Year No.

of Pa-

tients

Channels Segment

Length

SPH

(min)

Evaluation Data

balanc-

ing

(%)

Sen

(%)

Spe

(%)

AUC

(%)

(Yang et al.,

2021b)

13 22 5 sec 30 10-fold CV Overlap

window

92.0 87.8 92.8 91.3

(Rasheed

et al., 2021)-

2021

13 22 60 sec 10 10-fold CV GAN 92.0 90.9 89 -

(Truong

et al., 2018)

13 22 30 sec 30 LOO-CV

Seizure wise

Overlap

window

81.2 - -

(Ryu and

Joe, 2021)

23 18 10 sec 10 k-fold CV - 92.6 91.2 94.1 -

(Zhang et al.,

2021b)

19 23 8 sec 15 Train/Test

split

Overlap

window

89.9 92.9 87.0 -

(Usman

et al., 2021)

22 23 29 sec 32 k-Fold CV SMOTe - 93 92.5 -

(Liang et al.,

2022)

13 All 30 sec 30 LOO-CV Pa-

tient Wise

Overlap

window

- 88.3 - -

(Dissanayake

et al., 2021)

24 23 10 sec 60 10-fold CV Overlap

window

91.5 92.4 89.9 96.9

(Zhang et al.,

2021a)

13 23 5 sec 5 Test 1 patient;

Train Rest

- 80.0 - 74.1 -

(Sun et al.,

2021)

17 18 1.5 sec 30 4-fold CV Overlap

window

- 97.1 95.6 91.7

(Li et al.,

2021)

19 All 5 sec 15 LOO-CV

Seizure Wise

- - 95.5

- 93.8

(Gao et al.,

2022)

16 18 4 sec 30 LOO-CV Overlap

window

- 93.3 - -

(Halawa

et al., 2022)

16 18 10 sec 8 70% Train

30% Test

Overlap

window

93.4 - - 86.5

This work

(Channel

‘FP1F7’,

Worst Case)

13 1 1 sec 10 10-fold CV SMOTe 89.3 87.1 92.0 92.6

This work

(Channel

‘FzCz’ ,

Best case)

13 1 1 sec 10 10-fold CV SMOTe 96.0 98.0 93.5 98.8

4 CONCLUSIONS

Using a single-layer autoencoder and SVM classi-

ﬁer, we introduced a hybrid method for developing

a baseline model for the early prediction of epileptic

episodes in this study. Two phases of EEG data pro-

cessing are proposed for the prediction of seizures.

First, a sparse signal is generated using a dimension-

ality reduction methodology based on a deep learning

method for unsupervised learning. The SVM classi-

ﬁer is then trained to classify the data as either inter-

ictal or pre-ictal. Analysis of a single EEG chan-

nel are provided to predict the beginning of epileptic

episodes. Evaluation at the level of individual sig-

nals reveals the seizure-predicting capability of each

EEG channel. This method provides the possibility of

combining analysis of fewer EEG channels to get re-

liable seizure prediction on their basis. The proposed

method outperformed state-of-the-art methodologies,

with an average prediction sensitivity of 98% percent

and an area under the curve (AUC) of 98%. Develop-

ing such a high-performance system is beneﬁcial for

the construction and usage of wearable devices and

sensor networks, making them more comfortable to

Using Sparse Representation of EEG Signal from a Shallow Sparse Autoencoder for Epileptic Seizure Prediction

131

use and smaller in size as a result of the decreased

processing requirements. In the future, the current

effort will be broadened to produce more exhaustive

ﬁndings.

REFERENCES

Alotaiby, T. N., Alshebeili, S. A., Alotaibi, F. M., and Alr-

shoud, S. R. (2017). Epileptic seizure prediction using

csp and lda for scalp eeg signals. Computational in-

telligence and neuroscience, 2017.

Cui, S., Duan, L., Qiao, Y., and Xiao, Y. (2018). Learning

eeg synchronization patterns for epileptic seizure pre-

diction using bag-of-wave features. Journal of Am-

bient Intelligence and Humanized Computing, pages

1–16.

Dissanayake, T., Fernando, T., Denman, S., Sridharan, S.,

and Fookes, C. (2021). Deep learning for patient-

independent epileptic seizure prediction using scalp

eeg signals. IEEE Sensors Journal, 21(7):9377–9388.

Gao, Y., Chen, X., Liu, A., Liang, D., Wu, L., Qian, R., Xie,

H., and Zhang, Y. (2022). Pediatric seizure prediction

in scalp eeg using a multi-scale neural network with

dilated convolutions. IEEE Journal of Translational

Engineering in Health and Medicine, 10:1–9.

Halawa, R. I., Youssef, S. M., and Elagamy, M. N. (2022).

An efﬁcient hybrid model for patient-independent

seizure prediction using deep learning. Applied Sci-

ences, 12(11):5516.

Khan, G. H., Khan, N. A., Altaf, M. A. B., and Abid, M.

U. R. (2021). Classifying single channel epileptic eeg

data based on sparse representation using shallow au-

toencoder. In 2021 43rd Annual International Confer-

ence of the IEEE Engineering in Medicine and Biol-

ogy Society (EMBC), pages 643–646. IEEE.

Li, Y., Liu, Y., Guo, Y.-Z., Liao, X.-F., Hu, B., and

Yu, T. (2021). Spatio-temporal-spectral hierarchical

graph convolutional network with semisupervised ac-

tive learning for patient-speciﬁc seizure prediction.

IEEE transactions on cybernetics.

Liang, D., Liu, A., Li, C., Liu, J., and Chen, X.

(2022). A novel consistency-based training strategy

for seizure prediction. Journal of Neuroscience Meth-

ods, 372:109557.

Meng, Q., Catchpoole, D., Skillicom, D., and Kennedy, P. J.

(2017). Relational autoencoder for feature extraction.

In 2017 International Joint Conference on Neural Net-

works (IJCNN), pages 364–371. IEEE.

Rasheed, K., Qadir, J., O’Brien, T. J., Kuhlmann, L., and

Razi, A. (2021). A generative model to synthesize eeg

data for epileptic seizure prediction. IEEE Transac-

tions on Neural Systems and Rehabilitation Engineer-

ing, 29:2322–2332.

Ryu, S. and Joe, I. (2021). A hybrid densenet-lstm model

for epileptic seizure prediction. Applied Sciences,

11(16):7661.

Sun, B., Lv, J.-J., Rui, L.-G., Yang, Y.-X., Chen, Y.-G., Ma,

C., and Gao, Z.-K. (2021). Seizure prediction in scalp

eeg based channel attention dual-input convolutional

neural network. Physica A: Statistical Mechanics and

its Applications, 584:126376.

Tautan, A.-M., Dogariu, M., and Ionescu, B. (2019). Detec-

tion of epileptic seizures using unsupervised learning

techniques for feature extraction. In 2019 41st Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 2377–

2381. IEEE.

Truong, N. D., Nguyen, A. D., Kuhlmann, L., Bonyadi,

M. R., Yang, J., Ippolito, S., and Kavehei, O. (2018).

Convolutional neural networks for seizure prediction

using intracranial and scalp electroencephalogram.

Neural Networks, 105:104–111.

Usman, S. M., Khalid, S., and Bashir, Z. (2021). Epileptic

seizure prediction using scalp electroencephalogram

signals. Biocybernetics and Biomedical Engineering,

41(1):211–220.

Usman, S. M., Usman, M., and Fong, S. (2017). Epilep-

tic seizures prediction using machine learning meth-

ods. Computational and mathematical methods in

medicine, 2017.

Yang, J., Wu, Z., Peng, K., Okolo, P. N., Zhang, W., Zhao,

H., and Sun, J. (2021a). Parameter selection of gaus-

sian kernel svm based on local density of training

set. Inverse Problems in Science and Engineering,

29(4):536–548.

Yang, X., Zhao, J., Sun, Q., Lu, J., and Ma, X. (2021b).

An effective dual self-attention residual network for

seizure prediction. IEEE Transactions on Neural Sys-

tems and Rehabilitation Engineering, 29:1604–1613.

Zhang, Q., Ding, J., Kong, W., Liu, Y., Wang, Q.,

and Jiang, T. (2021a). Epilepsy prediction through

optimized multidimensional sample entropy and bi-

lstm. Biomedical Signal Processing and Control,

64:102293.

Zhang, S., Chen, D., Ranjan, R., Ke, H., Tang, Y., and

Zomaya, A. Y. (2021b). A lightweight solution to

epileptic seizure prediction based on eeg synchroniza-

tion measurement. The Journal of Supercomputing,

77(4):3914–3932.

Zhang, X. and Li, H. (2022). Patient-speciﬁc seizure pre-

diction from scalp eeg using vision transformer. In

2022 IEEE 6th Information Technology and Mecha-

tronics Engineering Conference (ITOEC), volume 6,

pages 1663–1667. IEEE.

Zhao, S., Yang, J., Xu, Y., and Sawan, M. (2020). Binary

single-dimensional convolutional neural network for

seizure prediction. In 2020 IEEE International Sym-

posium on Circuits and Systems (ISCAS), pages 1–5.

BIOSIGNALS 2023 - 16th International Conference on Bio-inspired Systems and Signal Processing

132