EPINET: AN OPTIMIZED, RESOURCE EFFICIENT DEEP GRU-LSTM NETWORK FOR EPILEPTIC SEIZURE PREDICTION

Abstract

The utilization of Electroencephalogram (EEG) as a non-invasive tool to investigate neurological disorders, particularly epilepsy, by capturing pathological biosignal markers indicative of seizures, sets the backdrop for this research endeavor. While previous studies have harnessed deep learning techniques for seizure detection, a pressing need persists for a resource-efficient model that demands minimal training data and time yet upholds commendable specificity and sensitivity. In response to this gap, we introduce an innovative deep Gated Recurrent Unit (GRU)– Long Short-Term Memory (LSTM) network, coined as EpiNET, purposefully crafted for the prediction of epileptic seizures using EEG data. A distinctive feature of EpiNET is its integration of statistical, spectral, and temporal features, chosen for their computational simplicity, thereby enhancing the model’s efficiency. The model is meticulously trained and validated on diverse patient datasets sourced from the CHB-MIT Scalp EEG database, outshining existing deep learning networks regarding seizure prediction accuracy. EpiNET boasts remarkable metrics, with reported sensitivity, accuracy, and specificity values standing at 92.54 ±?0.41%, 96.15 ±?0.45%, and 97.73 ±?0.58%, respectively. This underscores the efficacy of EpiNET while upholding a lean model structure, addressing concerns regarding computational efficiency. A ground-breaking aspect of this study is the introduction of a GRU-LSTM-based deep learning model capable of predicting epileptic seizures at least 2 h (120 min) in advance, marking a significant stride towards timely intervention and heightened patient care. In summary, this research not only advances the field of neurological disorder prediction but also underscores the paramount importance of resource efficiency in model development.