$\mathcal{S}^{3}$1DCNN: A Compact Stacked Spectral-Spatial Attention 1DCNN for Seizure Prediction with Wearables

Yang Zhang, Yvon Savaria, Mohamad Sawan, François Leduc-Primeau

Abstract: Seizure prediction has become a crucial field of research that aims to improve the lives of patients with drug-resistant epilepsy by reducing their anxiety and allowing the implementation of precautionary measures. Recently, deep learning has shown remarkable advancements in epilepsy prediction. However, this progress comes with increased computational demands and memory usage, which makes it unsuitable for low-power wearable devices. This work proposes a compact stacked spectral-spatial attention 1DCNN ($\mathcal{S}^{3}$1DCNN) leveraging the short-time Fourier transform (STFT). This model aims to enhance the interpretable ability to analyze non-stationary electroencephalography (EEG) signals, making it suitable for implementation in wearable biomedical devices. The results demonstrate that the proposed method outperforms recent state-of-the-art methods, achieving an average sensitivity of 92.1 %, an average false prediction rate (FPR) of 0.0081h, an average area under the ROC curve (AUC) of 0.980, and an estimated energy consumption of 0.21 µJ per inference on the American Epilepsy Society Seizure Prediction Challenge (AES) dataset. It demonstrates our method's promising application potential in low-power and energy-efficient wearable devices.