Abstract
Subject-independent cognitive workload decoding using functional near-infrared spectroscopy (fNIRS) remains challenging because workload-related hemodynamic responses are delayed and vary across participants. This study evaluated whether delayed temporal characteristics of fNIRS signals may support cross-subject workload classification. Two public n-back fNIRS datasets were analyzed under a strict subject-wise validation protocol, with 0-back and 2-back defined as low- and high-workload conditions, respectively. In each fold, test participants were completely excluded from training and validation to evaluate generalization to unseen individuals. Considering that task-evoked hemodynamic responses evolve gradually after task onset, an adaptive temporal modeling framework was developed to retain complete task blocks while learning soft temporal weights for informative response periods. In the primary Tufts 0-back versus 2-back setting, the proposed model achieved an accuracy of 73.60%, an F1-score of 75.67%, and an AUC of 77.44%. Further analyses showed that HbO/HbR differences between workload conditions became more evident during delayed response periods, and model-derived temporal importance exhibited a broadly similar temporal distribution. Temporal occlusion analysis further showed that removing delayed middle-to-late windows reduced decoding performance. These results indicate temporal correspondence between delayed HbO/HbR differences and model behavior, supporting the consideration of delayed temporal information in subject-independent fNIRS workload decoding.Supplementary information
The online version contains supplementary material available at https://doi.org/10.1007/s11571-026-10552-x.