SAFER-Net: a spatial attention and feature enhanced representation network for EEG-based driver drowsiness detection

Abstract

Drowsiness during driving is a risk causing road accidents worldwide, necessitating early detection of drowsiness. Electroencephalography (EEG) is used for developing reliable drowsiness detection system. EEG-based works utilize temporal-spatial representation, channel dependency modeling,attention and feature fusion, these capabilities are addressed through different architectural designs, and there remains a research gap in developing a unified representation learning-based framework that can jointly exploit complementary representations. This research proposes an end-to-end framework, Spatial Attention and Feature Enhanced Representation Network (SAFER-Net) for EEG-based driver drowsiness detection. SAFER-Net pipeline integrates entropy-based signal enhancement, Continuous Wavelet Transform (CWT) based temporal and frequency characteristics of EEG signal, multiscale Convolutional Neural Network (CNN) block with parallel convolutional branches and residual connections for feature extraction, a cascade of attention modules to recalibrate channel, spatial, and long-range temporal dependencies. To further enhance the transparency and model prediction interpretation, SAFER-Net is extended with Explainable AI (XAI) techniques like Gradient-weighted Class Activation Mapping (Grad-CAM) and Layer-wise Relevance Propagation (LRP). The proposed SAFER-Net is evaluated on benchmark SEED-VIG extracted dataset and its performance is compared against relevant architectures to demonstrate the effectiveness for drowsiness detection using EEG signals.