WearNeuroNet: An Interpretable Light-Weight Deep Learning approach for Ictal-Interictal Classification for Limited Channel EEG Wearables

For epileptic patients, distinguishing between seizure (ictal) and non-seizure (interictal) states with limited-channel electroencephalogram (EEG) wearables is challenging. Features, such as magnitude and phase from Fourier and wavelet transforms, showed efficacy on specific datasets but exhibited variability when applied with limited-channel EEG data. This letter proposes a deep learning-based WearNeuroNet architecture that uniquely varies the depth length and kernel size for depthwise convolution layers and the kernel size for 1D convolution layers. It uses skipped GRU layer connections to capture variability in time series EEG data. Various performance metrics, such as accuracy, sensitivity, specificity, and false alarm are analyzed on the seizure challenge public dataset provided by the ICASSP 2023. WearNeuroNet (variable depth) captured the interclass and intraclass variability of Epileptic patients improving the accuracy by 9.08% over baseline EEGNET.