DEEP LEARNING-ASSISTED EPILEPSY DETECTION AND PREDICTION

Abstract

Epilepsy is a multifaceted neurological disorder characterized by superfluous and recurrent seizure activity. Electroencephalogram (EEG) signals are indispensable tools for epilepsy diagnosis that reflect real-time insights of brain activity. Recently, epilepsy researchers have increasingly utilized Deep Learning (DL) architectures for early and timely diagnosis. This research focuses on resolving the challenges, such as data diversity, scarcity, limited labels, and privacy, by proposing potential contributions for epilepsy detection, prediction, and forecasting tasks without impacting the accuracy of the outcome. The proposed design of diversity-enhanced data augmentation initially averts data scarcity and inter-patient variability constraints for multiclass epilepsy detection. The potential features are extracted using a graph theory-based approach by analyzing the inherently dynamic characteristics of augmented EEG data. It utilizes a novel temporal weight fluctuation method to recognize the drastic temporal fluctuations and data patterns realized in EEG signals. Designing the Siamese neural network-based few-shot learning strategy offers a robust framework for multiclass epilepsy detection. Subsequently, Federated Learning (FL) architecture enables epileptic seizure prediction and enhances the generalization capability by utilizing numerous seizure patterns across diversified and globally distributed epileptic patients. By capturing the potential patterns, the hybrid model design potentially offers superior prediction accuracy by integrating a spiking encoder with graph convolutional neural networks. The preictal probability of each local model then aggregates the weights of the local medical centers with the global FL. Furthermore, applying the adaptive neuro-fuzzy inference system ensures a patient-specific preictal probability by combining the local model with patientspecific clinical features. Finally, epileptic seizure forecasting utilizes Self-Supervised Learning (SSL) capabilities to overcome the limitations of annotated EEG data. This selfsupervised transfer learning improves the training efficiency in massively arriving EEG data streams. The dual-feature embedding enhances the learning ability while a lightweight prediction utilizes the embeddings from the pretext task for epilepsy forecasting in the downstream task. The performance testing on the benchmark datasets reveals the accuracy of epilepsy detection, prediction, and forecasting by addressing the limitations of the existing approaches for effective patient management. The research outcomes ultimately enable real-time, transparent, and personalized solutions to ensure commitment towards the quality of life.