A new method for mental workload classification with EEG connectivity maps

Abstract

The classification of mental workload from electroencephalogram (EEG) signals is important for cognitive monitoring and adaptive human-machine interaction. In this work, we propose a new mental workload classification method that combines dDTF-derived EEG connectivity maps with a compact feature-learning pipeline. Multichannel EEG recordings are first transformed into multiband brain connectivity maps using the direct Directed Transfer Function (dDTF), which captures directed interactions among EEG channels. The resulting connectivity features are then refined through a two-stage feature selection process using discriminant feature test (DFT) ranking and auxiliary XGBoost (AXB). The selected features are further transformed by a least-squares normal transform (LNT) and classified using XGBoost. Experiments on the Simultaneous Task EEG Workload (STEW) dataset, conducted in subject-dependent and subject-independent settings, show that dDTF-based connectivity maps provide a more effective input representation than raw EEG signals. Under the same dDTF input setting, our method achieves strong classification performance relative to several deep learning baselines and Safari’s method as demonstrated by experimental results.