Feature Fusion-Based Capsule Network for Cross-Subject Mental Workload Classification

In a complex human-computer interaction system, estimating mental workload based on electroencephalogram (EEG) plays a vital role in the system adaption in accordance with users' mental state. Compared to within-subject classification, cross-subject classification is more challenging due to larger variation across subjects. In this paper, we targeted the cross-subject mental work-load classification and attempted to improve the performance. A capsule network capturing structural relationships between features of power spectral density and brain connectivity was proposed. The comparison results showed that it achieved a cross-subject classification accuracy of 45.11%, which was superior to the compared methods (e.g., convolutional neural network and support vector machine). The results also demonstrated feature fusion positively contributed to the cross-subject workload classification. Our study could benefit the future development of a real-time workload detection system unspecific to subjects.