A self-attention domain adaptation network based on Riemannian representation and contrastive learning for EEG driver drowsiness detection

EEG-based deep learning approaches have demonstrated effectiveness in detecting driver drowsiness, thereby enhancing classification accuracy. Despite this success, inter-subject EEG variability poses a significant challenge in building robust deep learning models. To mitigate this challenge, we introduce a deep neural network and a multi-source domain adaptation framework. The network first extracts multiscale features using a multiscale module, followed by a multi-head self-attention module to learn local and long-range dependencies. Then it employs a Riemannian manifold module to enrich the features by computing second-order inter-feature correlations in the complex EEG structure. The multi-source domain adaptation framework trains the network to overcome inter-subject EEG variability by combining coarse-grained marginal alignment and fine-grained conditional alignment. The coarse-grained marginal alignment is achieved using the maximum mean discrepancy (MMD), whereas fine-grained conditional alignment is achieved using label-aware maximum mean discrepancy (LMMD) and a supervised contrastive loss (Supcon). The proposed approach was evaluated on two publicly available EEG datasets for driver drowsiness detection. Under the challenging leave-one-subject-out (LOSO) protocol, it achieves an average accuracy of 88.95% on the SAD dataset, outperforming recent state-of-the-art methods, and demonstrates robust generalization on the SEED-VIG dataset with an accuracy of 93.05%.