Participant-Specific Adaptive Label Smoothing Using Subjective Workload Responses for Cognitive Workload Classification from EEG

Background: Cognitive workload classification from electroencephalography is affected by interindividual variability in EEG signals and subjective workload responses. We propose participant-specific adaptive label smoothing, in which the change in perceived workload between resting and task conditions controls the target confidence, while the experimental class remains unchanged. Methods: The method was evaluated on the STEW dataset and compared with five alternative supervision strategies. Two classification models were used: a feature-based convolutional neural network and a multi-scale Swin Transformer network (MST-Net) operating on time-frequency representations. Global and participant-specific stratified normalization were evaluated using participant-disjoint cross-validation. The participant-specific approach used normalization statistics estimated from each unseen participant’s own unlabeled EEG data. Results: Under stratified normalization, adaptive label smoothing produced the highest macro F1 for both models, with values of 0.768 and 0.796, and ROC AUC values of 0.844 and 0.878, respectively. It also maintained high recall for both classes. Conclusions: The findings suggest that participant-specific subjective workload responses can be used to modulate target confidence, with the most consistent benefit observed under participant-specific normalization. As the strategies differ in target definitions and no hard rest/task or fixed label-smoothing control was evaluated, the differences are descriptive.