An Explainable Transfer Learning Method for EEG-based Seizure Type Classification.

Epilepsy, traditionally conceptualized as a neurological disorder characterized by a persistent inclination toward epileptic seizures, is commonly diagnosed and monitored through EEGs. However, manual analysis of EEG data can be exceedingly time-consuming. The integration of automated seizure classification methods represents a valuable resource for clinicians engaged in epilepsy analysis. In this study, we introduce an explainable transfer learning method designed to classify seizure types within EEG recordings. Our method relies on the utilization of spectrograms derived from EEGs that capture seizure events, enabling the discrimination between normal EEG patterns and focal or generalized seizures. To achieve this, we employed four pre-trained transfer learning models, namely Inception, ResNet, DenseNet, and VGG16, using spectrogram data from 19 EEG channels as inputs. The model demonstrated high accuracy when assessed on an independent test dataset. To enhance clinician trust, we leveraged the LIME technique to elucidate model predictions, creating heatmap visualizations that emphasize explanation weights. The incorporation of a colorbar further facilitates clinician comprehension of predictions and aids in the identification of EEG seizure events. This method holds great promise for the effective classification of epilepsy types, providing support to neurologists in their comprehensive analysis of epilepsy cases.