Learning Subject-independent Representation for EEG-based Drowsy Driving Detection

Abstract

Training a subject-independent Electroencephalography (EEG) classification model is challenging since there are large variations in EEG signals between different subjects. To this end, existing works adopt the subject-dependent training scheme to reduce the individual variations, but training one model per each subject raises expensive costs, especially when the number of subjects is large. In this work, we aim to learn a subject-independent EEG classification model that predicts target labels independent of subjects, which avoids the cost issue. Specifically, we prevent the model from learning subject dependency via minimizing the mutual information between target and subject labels. Our model consists of a feature embedding module, followed by two branches for target and subject label prediction. The subject prediction module is trained adversarially against the feature embedding module, which encourages the feature representation to be encoded invariant to the subjects. To evaluate our method, we conduct experiments on the EEG-based drowsy driving detection task, requiring consistent performances among different subjects to be adapted in real-world applications. Through the analysis on SEED-VIG dataset, we demonstrate that our method achieves meaningful performance in terms of both accuracy and individual differences.