In the context of electroencephalogram (EEG)-based driver drowsiness recognition, it is still a challenging task to design a calibration-free system, since there exists a significant variability of EEG signals among different subjects and recording sessions. As deep learning has received much research attention in recent years, many efforts have been made to use deep learning methods for EEG signal recognition. However, existing works mostly treat deep learning models as blackbox classifiers, while what have been learned by the models and to which extent they are affected by the noise from EEG data are still underexplored. In this paper, we develop a novel convolutional neural network that can explain its decision by highlighting the local areas of the input sample that contain important information for the classification. The network has a compact structure for ease of interpretation and takes advantage of separable convolutions to process the EEG signals in a spatial-temporal sequence. Results show that the model achieves an average accuracy of 78.35% on 11 subjects for leave-one-out cross-subject drowsiness recognition, which is higher than the conventional baseline methods of 53.4%-72.68% and state-of-art deep learning methods of 63.90%-65.61%. Visualization results show that the model has learned to recognize biologically explainable features from EEG signals, e.g., Alpha spindles, as strong indicators of drowsiness across different subjects. In addition, we also explore reasons behind some wrongly classified samples and how the model is affected by artifacts and noise in the data. Our work illustrates a promising direction on using interpretable deep learning models to discover meaning patterns related to different mental states from complex EEG signals.
Download