FNIRS-Driven Explainable AI for Robust Mental Fatigue Classification in Human-Robot Collaboration

This study presents an interpretable, subjectindependent framework for detecting mental fatigue in simulated flight environments using functional near-infrared spectroscopy (fNIRS). Leveraging data from thirty-one participants performing prolonged flight monitoring tasks, the framework extracts comprehensive statistical, spectral, and crosschromatic features from oxygenated (HbO) and deoxygenated (HbR) hemoglobin signals. Rigorous leakage prevention and Leave-One-Subject-Out cross-validation (LOSO-CV) were implemented to ensure generalizability across individuals. Among the tested classifiers, the Gradient Boosting model achieved the highest performance for Karolinska Sleepiness Scale (KSS)- based fatigue states, which captures the gradual transition to drowsiness, yielding an F1-score of 0.72, AUROC of 0.782 0.092, and AUPRC of 0.783 0.098. To enhance interpretability, we applied Shapley additive exPlanations (SHAP) analysis. The results highlight that normalized OxyDiff features (reflecting the relative difference between HbO and HbR, standardized within-subject) are the most influential predictors, followed by HbO variability and low-frequency band-power descriptors. This dominance of subject-normalized features and the strong performance on the KSS-based label underscore that the reliable detection of fatigue using frontal fNIRS is driven by slow, relative hemodynamic shifts and the balance between HbO and HbR, rather than absolute signal amplitudes or transient changes..