Predicting driving performance from resting-state EEG using explainable gradient boosting

Abstract Sleep deprivation substantially impairs driving performance, underscoring the need for Electroencephalography (EEG)-derived markers associated with driving-performance vulnerability under varying levels of sleepiness. In this study, resting-state EEG signals were recorded from 40 healthy participants following fragmented sleep or total sleep deprivation, simulating conditions of elevated sleep pressure. Recordings were collected across 12 sessions under eyes-open and eyes-closed conditions and acquired throughout the day and night to incorporate variability related to vigilance state, circadian phase, and sleep pressure. Participants subsequently completed a 20-min driving simulation, with performance quantified using the standard deviation of lateral position. We developed a subject-independent framework to predict driving performance (Safe vs. Risky) from resting-state EEG using a minimal number of channels under heterogeneous recording conditions. Spectral features were extracted from subject-specific frequency sub-bands aligned to individual Alpha peak frequencies, and Shapley additive explanations (SHAP)-based feature selection was employed to improve interpretability while maintaining predictive performance. SHAP were additionally used to characterize feature contributions to model predictions. Classification was performed using CatBoost, a gradient-boosted decision-tree algorithm. The model achieved a subject-independent classification accuracy of 0.78 using EEG-derived features alone and 0.82 when age was included, despite substantial inter- and intra-individual variability. Delta- and Theta-band power at the Pz electrode during eyes-closed rest emerged among the features contributing most strongly to model predictions. Using only two EEG channels (Cz and Pz), the gradient-boosting models maintained comparable predictive performance. These findings support the feasibility of subject-independent, low-density EEG-based modeling for predicting driving-performance-related outcomes under controlled experimental conditions.