Classification of Epileptic Seizures Based on CNN and Guided Back-Propagation for Interpretation Analysis

Epilepsy is a brain disorder that affects nearly 50 millions people around the world. However, despite its high prevalence, the diagnosis of epilepsy is complex and usually guided by the medical team’s expertise using several neuroimaging techniques such as the Electroencefiphalogram (EEG). In this sense, several studies aim to automatically detect epilepsy events to improve the diagnosis or predict epilepsy crises from EEG recordings to overtake a preventive action. Most recent studies using deep learning approaches to detect Epilepsy Seizures (ES) achieve outstanding classification results but lack interpretability. This work proposes a methodology for classifying ES from EEG signals using intra-subject models in 10 patients from Siena Scalp EEG Database using Time Frequency Representations. In this approach a standard Convolutional Neural Network (CNN) structure is used and several parameters of CNN are optimized using bayesian optimization. Additionally, we used a combination of GradCam and Guided Backpropagation as a strategy to interpret the performance of trained models. The classification results suggest that the intra-subject approach is effective when few data are available. Most of the analyzed subjects got an accuracy that outperform the standard architecture by using optuna tuning. Additionally, the computation of gradients using GradCam-BP show an increased mean gradient for some patients finding the relevant channel in the ES condition.