Subject-Aware User State Classification with Deep Learning Models: An Exploratory Study

Haram Kwon, Jaehoon Choi, Jin Woo Choi, Sungho Jo

Abstract: Inter-subject variability is one of the critical issues that hinder electroencephalogram-based brain-computer interfaces from wide usage. Recent studies that aimed to tackle such problems utilize deep learning methods such as domain adaptation to have their feature extractor learn domain-invariant features. As such approaches employ user state classification as well as subject classifiers to contribute toward domain-invariant feature extractions, considering the performance on both the state classification and subject identification for designing the feature extraction model may further benefit such approach. Thus in this work, we aim to improve widely used convolutional neural network-based feature extractors by enhancing subject identification accuracy while preserving user state classification. Along with our approach of using multi layer perceptron, we trained and evaluated our method using the visual imagery dataset and the speech imagery dataset collected from five participants. By training with EEG dataset of one paradigm and evaluating with the other dataset, our proposed feature extraction method achieved higher subject identification accuracy than the baseline models, while preserving their user state classification performance.