Open Access

ARTICLE

A Neural ODE-Enhanced Deep Learning Framework for Accurate and Real-Time Epilepsy Detection

Tawfeeq Shawly^1,2, Ahmed A. Alsheikhy^3,*

1 Department of Electrical Engineering, Faculty of Engineering at Rabigh, King Abdulaziz University, Jeddah, 21589, Saudi Arabia
2 King Salman Center for Disability Research, Riyadh, 11614, Saudi Arabia
3 Department of Electrical Engineering, College of Engineering, Northern Border University, Arar, 91431, Saudi Arabia

* Corresponding Author: Ahmed A. Alsheikhy. Email:

Computer Modeling in Engineering & Sciences 2025, 143(3), 3033-3064. https://doi.org/10.32604/cmes.2025.065264

Received 08 March 2025; Accepted 12 May 2025; Issue published 30 June 2025

Abstract

Epilepsy is a long-term neurological condition marked by recurrent seizures, which result from abnormal electrical activity in the brain that disrupts its normal functioning. Traditional methods for detecting epilepsy through machine learning typically utilize discrete-time models, which inadequately represent the continuous dynamics of electroencephalogram (EEG) signals. To overcome this limitation, we introduce an innovative approach that employs Neural Ordinary Differential Equations (NODEs) to model EEG signals as continuous-time systems. This allows for effective management of irregular sampling and intricate temporal patterns. In contrast to conventional techniques, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), which necessitate fixed-length inputs and often struggle with long-term dependencies, our framework incorporates: (1) a NODE block to capture continuous-time EEG dynamics, (2) a feature extraction module tailored for seizure-specific patterns, and (3) an attention-based fusion mechanism to enhance interpretability in classification. When evaluated on three publicly accessible EEG datasets, including those from Boston Children’s Hospital and the Massachusetts Institute of Technology (CHB-MIT) and the Temple University Hospital (TUH) EEG Corpus, the model demonstrated an average accuracy of 98.2%, a sensitivity of 97.8%, a specificity of 98.3%, and an F1-score of 97.9%. Additionally, the inference latency was reduced by approximately 30% compared to standard CNN and Long Short-Term Memory (LSTM) architectures, making it well-suited for real-time applications. The method’s resilience to noise and its adaptability to irregular sampling enhance its potential for clinical use in real-time settings.

---

Keywords

---