Physiologically interpretable DNN using blink-related brain activity solely from an EEG device in mental fatigue recognition.

There is a great demand for mental fatigue recognition for enhancing work efficiency and the immersion of contents. Brain activity and eye movements are valuable biomarkers for mental fatigue recognition, as they reflect attentional processes associated with this state. However, a key challenge lies in the lack of mental fatigue recognition models grounded in physiological mechanisms. Blink-related brain activity has been investigated as a potential physiological mechanism, and further exploration is needed from the perspective of its applications to recognition models and minimizing form factors. In this paper, we propose a physiologically interpretable deep neural network (DNN) that extracts and averages blink-related electroencephalography (EEG) features extracted solely from an EEG device for accurate mental fatigue recognition. In this study, the contribution of our proposed architecture was validated using data collected by an EEG device with electrodes across the entire scalp. The results showed that our proposed framework improved the performance of mental fatigue recognition (an average accuracy of 0.73 and a macro-F1 score of 0.72). Furthermore, the results demonstrated that our proposed framework makes use of psychophysiological mechanisms. These findings indicate the potential of our framework for effectively capturing blink-related brain activity reflecting mental fatigue by extracting EEG signals associated with blinks and averaging the extracted features.