Epileptic seizure onset early warning using a lightweight multimodal self-explainable neural network

BackgroundCutting-edge deep learning algorithms are applied to epileptic seizures to uncover hidden patterns and achieve high accuracy. Post hoc algorithms explain black-box models but do not reveal the detailed workings of the algorithms. Therefore, intrinsic methods overcome these issues; however, they suffer from improper use of indicative metrics and performance-to-accuracy trade-offs.ObjectiveThis study implements a lightweight multimodal explainable Neural Network for real-time early warning of epileptic seizure onset.MethodsPeople with drug-resistant epilepsy across all age groups who had regular follow-up at Emanuel Mental Specialized Hospital in Ethiopia during the data collection period were the study population. EEG recordings were performed using the 10/20 system. A pretest was conducted two weeks before actual data collection. EDFbrowser was used for preparing the dataset in CSV format. Learned and identity self-explainable neural network algorithms were used to develop a model. In addition to classical evaluation metrics, explicitness, faithfulness, stability, and clear patterns behind predictions were used. Ethical approval was also obtained from the institutional review board (IRB) with reference number No: AM/146/2/36.ResultsLearned and identity self-explainable neural networks were developed and compared. The models' accuracies are 96.524% and 95.328%, and their stability mean values are 99.88% and 99.88%, respectively. Additionally, explicitness values of 0.0055 and 0.01197 were achieved, respectively.ConclusionThe results demonstrate that using a self-explainable neural network is a promising step toward responsible and transparent artificial intelligence. Comparing the two models shows that the learned self-explainable neural network is slightly more optimistic than the identity self-explainable neural network.