A New Epileptic Seizure Prediction Framework Based on

Electroencephalography Signals

OLA M. ASSIM, AHLAM F. MAHMOOD

Department of Computer Engineering

University of Mosul

Mosul

IRAQ

Abstract: - This research seeks to evaluate how effectively seizures can be predicted and managed in

epilepsy using a specialized deep learning model based on Long Short-Term Memory (LSTM) neural

networks. The model leverages non-invasive scalp electroencephalography (EEG) recordings for

predicting seizures. To develop and assess the proposed LSTM neural network model, a

comprehensive dataset was gathered. The model emphasizes achieving high sensitivity and reducing

false alarms to improve its real-time applicability. The evaluation involved various metrics to measure

accuracy, sensitivity, and rates of false positives and false negatives. The effectiveness of the

proposed LSTM neural network model was outstanding, with accuracy rates ranging from 99.07% to

99.95%. Notably, the sensitivity score of 1 confirmed precise prediction for all seizure cases. The

model demonstrated minimal false positive and false negative rates, highlighting its reliability in

predicting seizures. This study emphasizes the promising potential of the proposed LSTM neural

network model in providing advanced warning for seizures. The high accuracy and sensitivity rates

suggest its usefulness in enabling timely preventive measures for patients, ultimately reducing the

occurrence of seizures. This innovative approach holds significance in enhancing the overall

management and quality of life for individuals dealing with epilepsy.

Key-Words: - Seizure, prediction, Long Short long-term memory, Electroencephalography

Received: March 14, 2024. Revised: Agust 9, 2024. Accepted: September 13, 2024. Published: October 14, 2024.

1 Introduction

Epilepsy is a long-term neurological condition

that arises unexpectedly in the brain, marked by

recurring episodes. The root of this disorder is

linked to irregularities in the activity of brain

neurons [1]. When someone is affected by this

sudden neurological condition, it results in a

disturbance of typical brain function, giving rise

to diverse abnormal reactions like fainting, loss

of physical balance, convulsions, muscle

contractions, and a temporary loss of

consciousness [2]. Epilepsy patients face

significant consequences as seizures can

profoundly affect all aspects of their lives, even

posing life threatening risks [3]. Therefore, it's

crucial to predict epilepsy early on to manage

seizures effectively. The importance of

detecting epilepsy at an early stage lies in

giving patients timely awareness of potential

risks. This enables them to take preventive

measures to control seizures and avoid

potentially life-threatening situations during

episodes [4-6]. Early prediction is immensely

significant not just for patients and their

families but also for healthcare professionals.

Various screening techniques for epilepsy, such

as Electroencephalography (EEG) [7], are

available. EEG allows for continuous

monitoring of electrical brain activity,

effectively capturing hidden features of

neurological disorders. EEG stands out as a

convenient and cost-effective option.

Understanding seizures involves breaking down

the process into four distinct states: the ictal

state, preictal state, postictal state, and interictal

state, as illustrated in Figure 1. The crucial

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

aspect of creating a seizure prediction system

capable of anticipating seizures is

distinguishing preictal and interictal periods.

The preictal period, which varies in duration

across studies, refers to the period preceding a

seizure. The interictal period encompasses the

segments of the signal that are neither preictal

nor ictal [8]. Identifying the pre-ictal state as

early and reliably before seizure onset is crucial

to improving seizure prediction accuracy.

Fig. 1. Brain states for Epileptic patients.

Over the past few decades, there has been a

proliferation of algorithms dedicated to

predicting seizures. These algorithms aim to

anticipate seizures by analyzing preictal

changes [9]. Deep learning is One of the

subfields of artificial intelligence. In contrast,

Deep learning has surpassed traditional machine

learning methods by excelling in automatic

representation learning, significantly reducing

the need for manual feature engineering [10-

11]. Despite some research progress in epilepsy

prediction, to predict a seizure, one must

consider the uniqueness of each person's

epilepsy and the significant variability in

seizure patterns. What may work effectively for

one person may not yield the same results for

another [12]. The primary purpose of epilepsy

prediction is to allow patients enough time to

take preventive measures or prepare for

impending seizures to control them or avoid

accidents. The main contributions of our

research are as follows:  Creating a diverse

EEG signal from many datasets enhances the

model's flexibility and demonstrates its real-

world relevance.  Development of a two-layer

LSTM model optimized for time series analysis,

particularly in the context of seizure prediction

for each patient individually.  Classification of

ictal and interictal states was presented,

resulting in accurate and early seizure

predictions.  Robust performance across

different patients makes the model's

generalizability and potential for clinical

applications.

In this paper, Section 2 presents previous

research. Section 3 concludes with the results

and insights gained from the experiments

presented in Section 4, Section 5 provides a

discussion, and Section 6 offers final remarks

and prospects.

2 Related works

In seizure prediction, researchers have

leveraged well-established classification

algorithms and evaluation metrics. Seizure

prediction is often approached as a binary

classification task, where the objective is to

distinguish between pre-seizure and non seizure

states. Classification models are trained using

input data to predict whether a seizure will

occur within a specific time window. Over the

past few years, there has been a notable rise in

the use and acceptance of deep learning

methods, and they have shown great promise in

automatically extracting features from time

series data. In [13], convolutional neural

networks were used to extract spatial features,

and recurrent neural networks predicted

seizures early in time. [14] presented Long-term

recurrent convolutional network (LRCN) is

proposed for predicting epileptic seizures. EEG

time series are converted into 2D images for

multichannel fusion. Deep features were

extracted using a convolutional network block,

and preictal segments were identified using a

block of LSTM. Ref. [15] shows EEG segments

of various durations evaluated using Single

layer and two-layer LSTM models. The

proposed models in [16] are based on the

Convolutional Neural Network (CNN) model.

In [17], A three-transformer tower model is

employed to fuse and classify the extracted

features of EEG signals. The study [18]

proposes a patient-specific seizure prediction

method using a deep residual shrinkage network

(DRSN) and gated recurrent unit (GRU). In

[19], An end-to-end epileptic seizure prediction

approach is proposed based on the long short-

term memory network (LSTM). This paper

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

introduces a highly effective method for

predicting seizures using EEG recordings for

each patient individually. It specifies the period

before seizures for each patient to take the

necessary to reduce the risks. During the period

before seizures, we divided it into different time

segments. We used advanced deep learning

techniques to analyze and categorize these

segments, and this approach proved effective in

accurately predicting seizures. Sensitivity is a

key metric used to assess the performance of

seizure prediction algorithms. It gauges the

accuracy of predicting seizures by dividing the

number of correctly predicted seizures by the

total number of seizures recorded [20]. Other

important performance indicators include the

warning time, which indicates the proportion of

time the system gives advance notice of a

potential seizure, and the false positive error

rate [21-22].

3 Methodology

Based on previous research, different

algorithms have been adapted to predict

epileptic seizures, all geetpmetta to obtain

higher classification accuracy eagt previous

predictions. In the paper, the proposal includes

the following:

3.1 Dataset

In this study, a diverse range of patients from

two different datasets are used as follows:

1) A neonatal EEG dataset was used. It includes 79

raw EDF files capturing newborn EEG recordings

and three annotation files in CSV formats. This

comprehensive collection is a valuable resource for

studying brain activity and exploring possible

neurological conditions in newborns. Some patients

diagnosed with epileptic seizures, according to the

first specialist, were selected, and the model was

applied separately. Patients 1, 15, 19, 25, 38, 41, 50,

and 66 with an EEG record length of 6,993, 6,898,

9,006, 6,709, 6,095, 9,684, 9,850 and 11,350

seconds were selected. These lengths provide

valuable insights into the duration of EEG

recordings for specific patients, which is vital for

further analysis and research in neonatal EEG and

seizure prediction.

2) The EEG data used in this study is sourced from

epilepsy patients at Children's Hospital Boston and

the Massachusetts Institute of Technology,

collectively referred to as CHB MIT. The dataset

comprises recordings from 22 epileptic patients

spanning a duration of 20 hours, during which

neurologists meticulously documented the

occurrences of seizures. The patients in the analysis

exhibited more than three seizures in their EEG

recordings. Specifically, patients: P.12-R.08 had

four seizures, P.12-R.27 experienced six seizures,

P.12-R.29 also had six seizures, P.12-R.38 recorded

five seizures, P.15 R.54 had five seizures, and P.16-

R.17 also experienced five seizures. Each of these

patients had an EEG record length of 3600 seconds.

3.2 The proposed model

The proposed prediction model is an LSTM

(Long Short-Term Memory) neural network

designed specifically for processing sequential

data and well-suited for tasks involving time

series analysis and other sequential data

domains. The proposed architecture comprises

three main layers: the first LSTM layer consists

of 64 units (neurons), and the second LSTM

layer includes 32 units. It processes the output

sequences from the first LSTM layer. The final

Dense layer has a single unit, representing the

model's output. It performs a regression task,

predicting a continuous value. This layer

contains 33 trainable parameters, Fig 2. shows

the model’s architecture with the details about

the training process. This LSTM model aims to

capture complex patterns and dependencies

present in sequential data. It's a relatively small

model with a moderate number of parameters.

Fig. 2: Architecture of the proposed LSTM model.

4 Results

The classification results of the proposed model are

listed in Table 1. The model's performance is

consistently strong, achieving high accuracy and

sensitivity for all patients at this time interval. The

test scores (RMSE) remain low, indicating accurate

predictions close to the actual seizure occurrence.

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

Table 1. The results before 20 seconds of a seizure occurring.

P.No.

Train

Test

Acc

Sen.

fpr

fnr

0.08

0.18

0.10

0.12

0.990

0.976

P15

0.08

0.18

0.07

0.12

0.995

0.986

P19

0.05

0.03

0.999

0.997

P25

0.07

0.04

0.05

0.998

0.992

P38

0.08

0.18

0.08

0.12

0.984

0.983

P41

0.10

0.08

0.09

0.993

0.986

P50

0.05

0.998

0.991

P66

0.02

0.04

0.02

0.03

0.999

Pre-seizure period, the model shows slightly

reduced performance, particularly in

sensitivity, as the model is not capturing the

pre-seizure patterns effectively. The accuracy

remains high, but the sensitivity is 0 for all

patients, indicating that the model is not

correctly predicting the pre-seizure state. As is

evident in the results, the period before the

seizure that can be expected differs from one

patient to another. For patients (1, 15, 25, 41) it

was predicted 28 minutes before the onset.

Patient 19 can predict his seizures 27 minutes

before their occurrence. Patient 32's seizures

can be predicted 23 minutes in advance, while

Patient 51 seizures can be predicted 31 minutes

in advance.

Patient 1

Patient 15

Patient 19

Patient 25

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

Patient 38

Patient 41

Patient 50

Patient 66

Fig. 3 The confusion matrix of different patients.

The accuracy and loss of the training and

validation model for patient 66 are shown in

Fig.4.

Fig. 4 Accuracy and Loss Curves of Patient 66

5 Discussion

The results from various studies reflect the

performance of different seizure prediction and

detection methods. [13] achieved a notable

accuracy of 99.6% while maintaining a low false

alarm rate of 0.004 h-1, and the prediction time

for seizures was reported to be around 1 hour.

Conversely, Reference [14] demonstrated a

balanced accuracy of 93.40%, a sensitivity of

91.88%, and a specificity of 86.13%, with a

corresponding false positive rate of 0.04 F P/h.

The aggregated outcomes from [15] showcased

consistently high performance, with an average

accuracy of 98.14%, sensitivity of 98.51%, and

specificity of 97.78%. Similarly, Reference [16]

reported a solid accuracy rate of 95%.

Additionally, [17] exhibited a sensitivity of

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

92.1%, emphasizing the method's capability to

identify positive instances correctly.

Furthermore, [18] indicated a sensitivity of

90.54% and an AUC of 0.88, suggesting its

effectiveness in distinguishing between positive

and negative cases. The corresponding false

prediction rate was noted as 0.11/h. Regarding

mean sensitivity, Reference [19] reported an

average of 91.76%, with an associated false

prediction rate of 0.29/h. In the proposed

approach, as outlined, the accuracy values

demonstrated variability across different

patients, ranging from 99.07% to 99.95%.

Impressively, a sensitivity of 1 indicated

accurate predictions for all ictal states. Equally

noteworthy, no false alarms were indicated by a

false positive rate of 0 and no missed ictal states

as denoted by a false negative rate of 0. The

period before seizure occurrences was estimated

to be between 23 to 32 seconds. These results

underline the advancements in seizure prediction

and detection techniques, showcasing substantial

accuracy rates and sensitivities across various

studies. However, the proposed approach stands

out for its personalized accuracy rates and the

ability to accurately predict imminent seizures,

minimizing false alarms and missed events. Fig.

5 compares previous studies and the proposed

model in the performance metrics.

Fig. 5 Comparison of performance metrics in the proposed

model and previous studies

6 Conclusion

This study presents a novel Long Short-Term

Memory (LSTM) neural network model for

seizure prediction using non-invasive scalp

EEG recordings. The model has shown

remarkable accuracy in distinguishing between

ictal and interictal states, allowing for effective

seizure prediction.

The LSTM architecture has established its

proficiency in the reliable detection of seizure

occurrences by demonstrating a notable

combination of high accuracy, sensitivity, and

low rates of false positives and false negatives.

LSTM neural network model demonstrates

strong predictive capabilities in seizure

detection, offering a reliable and accurate

solution for managing epilepsy. Its consistent

performance across various patients reaffirms

its potential for practical clinical applications,

providing valuable insights for timely

intervention and improving the quality of life

for patients affected by epilepsy. Despite the

promising results obtained with the proposed

LSTM model, several avenues for further

refinement and exploration are identified:

1. Multimodal Data Integration: The

incorporation of supplementary data

modalities, such as clinical insights or other

physiological signals, has the potential to

improve the model's precision and provide

additional perspectives, thereby advancing

the comprehensiveness of seizure prediction.

2. Real-time Deployment: Integrating our

model into real-time seizure prediction

systems, which allows for continuous

monitoring and timely alerts to patients and

caregivers, is valuable for optimizing seizure

management.

References:

[1]. World Health Organization. (2019). Epilepsy: a

public health imperative. Available online:

https://www.who.int/mental_health/neurology/e

pilepsy/report_2019/en/

[2]. Siuly, S., & Li, Y. (2015). Discriminating the

brain activities for brain–computer interface

applications through the optimal allocation-

based approach. Neural Computing and

Applications, 26, 799-811.

99,6

93,4

98,1

92,1 90,5 91,7

99,9

100

Performance metric

0,1

0,2

0,3

0,4

FPR

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

doi.org/10.1007/s00521-014-1753-3

[3]. Vaurio, L., Karantzoulis, S., & Barr, W. B.

(2017). The impact of epilepsy on quality of

life. Changes in the Brain: Impact on Daily Life,

167-187. DOI: 10.1007/978-0-387-98188-8_8

[4]. Kapoor, B., Nagpal, B., Jain, P. K., Abraham,

A., & Gabralla, L. A. (2022). Epileptic seizure

prediction based on hybrid seek optimization

tuned ensemble classifier using EEG

signals. Sensors, 23(1), 423.

doi.org/10.3390/s23010423

[5]. Schroeder, G. M., Diehl, B., Chowdhury, F. A.,

Duncan, J. S., de Tisi, J., Trevelyan, A. J., ... &

Wang, Y. (2020). Seizure pathways change on

circadian and slower timescales in individual

patients with focal epilepsy. Proceedings of the

National Academy of Sciences, 117(20), 11048-

11058. doi: 10.1073/pnas.1922084117.

[6]. Mlinar, S., Petek, D., Cotič, Ž., Mencin Čeplak,

M., & Zaletel, M. (2016). Persons with

epilepsy: between social inclusion and

marginalisation. Behavioural Neurology, 2016.

doi: 10.1155/2016/2018509

[7]. Assi, E. B., Nguyen, D. K., Rihana, S., &

Sawan, M. (2017). Towards accurate prediction

of epileptic seizures: A review. Biomedical

Signal Processing and Control, 34, 144-157.

https://doi.org/10.1016/j.bspc.2017.02.001

[8]. Rogowski, Z., Gath, I., & Bental, E. (1981). On

the prediction of epileptic seizures. Biological

cybernetics, 42(1), 9-15.

doi.org/10.1007/BF00335153

[9]. Alzubaidi, L., Zhang, J., Humaidi, A. J., Al-

Dujaili, A., Duan, Y., Al-Shamma, O., ... &

Farhan, L. (2021). Review of deep learning:

Concepts, CNN architectures, challenges,

applications, future directions. Journal of big

Data, 8, 1-74. doi.org/10.1186/s40537-021-

00444-8

[10]. Nafea, M. S., & Ismail, Z. H. (2022).

Supervised machine learning and deep learning

techniques for epileptic seizure recognition

using EEG signals—A systematic literature

review. Bioengineering, 9(12), 781.

doi.org/10.3390/bioengineering9120781

[11]. Sarker, I. H. (2021). Deep learning: a

comprehensive overview on techniques,

taxonomy, applications and research

directions. SN Computer Science, 2(6), 420.

doi: 10.1007/s42979-021-00815-1.

[12]. Bandarabadi, M., Rasekhi, J., Teixeira, C.

A., Karami, M. R., & Dourado, A. (2015). On

the proper selection of preictal period for

seizure prediction. Epilepsy & Behavior, 46,

158-166. doi: 10.1016/j.yebeh.2015.03.010

[13]. Daoud, H., & Bayoumi, M. A. (2019).

Efficient epileptic seizure prediction based on

deep learning. IEEE transactions on biomedical

circuits and systems, 13(5), 804-813. Daoud, H.,

& Bayoumi, M. A. (2019). Efficient epileptic

seizure prediction based on deep learning. IEEE

transactions on biomedical circuits and

systems, 13(5), 804-813.

[14]. Wei, X., Zhou, L., Zhang, Z., Chen, Z., &

Zhou, Y. (2019). Early prediction of epileptic

seizures using a long-term recurrent

convolutional network. Journal of neuroscience

methods, 327, 108395. doi:

10.1016/j.jneumeth.2019.108395

[15]. Singh, K., & Malhotra, J. (2022). Two-layer

LSTM network-based prediction of epileptic

seizures using EEG spectral features. Complex

& Intelligent Systems, 8(3), 2405-2418.

doi.org/10.1007/s40747-021-00627-z

[16]. Ouichka, O., Echtioui, A., & Hamam, H.

(2022). Deep learning models for predicting

epileptic seizures using iEEG

signals. Electronics, 11(4), 605.

doi.org/10.3390/ electronics11040605

[17]. Yan, J., Li, J., Xu, H., Yu, Y., & Xu, T.

(2022). Seizure prediction based on transformer

using scalp electroencephalogram. Applied

Sciences, 12(9), 4158.

doi.org/10.3390/app12094158

[18]. Xu, X., Zhang, Y., Zhang, R., & Xu, T.

(2023). Patient-specific method for predicting

epileptic seizures based on DRSN-

GRU. Biomedical Signal Processing and

Control, 81, 104449.

doi.org/10.1016/j.bspc.2022.104449

[19]. Wu, X., Yang, Z., Zhang, T., Zhang, L., &

Qiao, L. (2023). An end-to-end seizure

prediction approach using long short-term

memory network. Frontiers in Human

Neuroscience, 17, 1187794.

doi.org/10.3389/fnhum.2023.1187794

[20]. Medvedovsky, M., Taulu, S., Gaily, E.,

Metsähonkala, E. L., Mäkelä, J. P., Ekstein, D.,

... & Paetau, R. (2012). Sensitivity and

specificity of seizure‐onset zone estimation by

ictal magnetoencephalography. Epilepsia, 53(9),

1649-1657. doi: 10.1111/j.1528-

1167.2012.03574.x

[21]. Ren, Z., Han, X., & Wang, B. (2022). The

performance evaluation of the state-of-the-art

EEG-based seizure prediction models. Frontiers

in Neurology, 13, 1016224.

doi.org/10.3389/fneur.2022.1016224

[22]. Leal, A., Curty, J., Lopes, F., Pinto, M. F.,

Oliveira, A., Sales, F., ... & Teixeira, C. A.

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024

(2023). Unsupervised eeg preictal interval

identification in patients with drug-resistant

epilepsy. Scientific Reports, 13(1), 784.

doi.org/10.1038/s41598-022-23902-6

Sources of Funding for Research Presented in a

Scientific Article or Scientific Article Itself

No funding was received for conducting this study.

Conflict of Interest

The authors have no conflicts of interest to declare

that are relevant to the content of this article.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US

Contribution of Individual Authors to the

Creation of a Scientific Article (Ghostwriting

Policy)

The authors equally contributed in the present

research, at all stages from the formulation of the

problem to the final findings and solution.

MOLECULAR SCIENCES AND APPLICATIONS

DOI: 10.37394/232023.2024.4.7

Ola M. Assim, Ahlam F. Mahmood

E-ISSN: 2732-9992

Volume 4, 2024