Eeg-based analysis of Driver cognitive load in subsea tunnel vs. Open Road: A Hybrid Machine learning and Statistical Modeling Approach

Driver cognitive load significantly impacts traffic safety, particularly in high-risk environments such as subsea tunnels. However, the complex effects of environmental factors and individual characteristics on cognitive load remain inadequately understood. To address this gap, this study employed a two-stage research framework to estimate cognitive load and examine its contributing factors. First, a standardized auditory N-back task was used to develop an EEG-based cognitive-load calibration model using features spanning the time domain, frequency domain, nonlinear dynamics, and functional connectivity. Subject-independent leave-one-subject-out (LOSO) validation yielded an accuracy of 87.2 %. Out-of-fold SHAP analysis identified cross-channel functional-connectivity and multiscale-complexity features as leading contributors to the model predictions. The calibrated model was then applied to real-world driving data to derive cognitive-load estimates in two driving environments: an open road and a subsea tunnel. The model-derived cognitive-load estimate was highest at the subsea tunnel exit. Mixed-effects ordered logit models showed that, within the tunnel, illuminance variability was associated with higher odds of assignment to a higher model-derived load category, OR = 1.147. Greater upslope grade was also associated with higher odds in both violation groups. This association was stronger in the higher-violation group, with conditional ORs of 1.630 and 1.181, respectively, and an interaction OR of 1.380. The upslope association was curvilinear and became steeper above approximately 2 %. These findings provide essential theoretical and empirical support for the development of context-sensitive intelligent driver-assistance systems and targeted infrastructure safety measures.