Cross-Subject Drowsiness Recognition Based on EEG Signals of Frontal Area

For EEG-based drowsiness recognition, cross-subject recognition is desirable since calibrating for each subject is time-consuming. Although EEG provides objective measures of fatigue with very high temporal resolution, its practical application is often limited by the complexity of multichannel systems. Although frontal EEG signals are more practical and portable, they may suffer from higher levels of noise and less robust signal quality compared to full-cap setup. In this paper, we introduce a Squeeze and Excitation Convolutional Neural Network (SECNN) aimed at improving cross-subject recognition of drowsiness utilizing solely forehead EEG signals. The network features a compact structure and employs separable convolutions specifically to reduce the number of parameters and computational complexity, thereby improving the network's convergence and efficiency. Furthermore, the use of separable convolutions aids in further noise and artifact removal, enhancing the richness of the extracted features from the forehead EEG signals. Results show that the model achieves an average accuracy of 80.68% for cross-subject drowsiness recognition on a public dataset using only three forehead EEG channels. This performance significantly exceeds that of traditional baseline methods (53.40% -72.68%) and surpasses state-of-the-art deep learning methods (68.66%- 78.03%) that utilized 30 EEG channels.