Assessment and prediction of driving risk at tunnel entrance and exit sections on mountainous roads

OBJECTIVES: To investigate the spatial evolution of driving visual load under different road alignment conditions and its relationship with driving risk at tunnel entrance and exit sections on mountainous two-lane roads, and to develop an integrated framework for risk assessment and prediction. METHODS: Real-vehicle tests were conducted across nine mountainous two-lane road tunnels in Chongqing, China; valid data from 27 of 30 recruited drivers were retained after preprocessing. Eye-movement, vehicle-speed, and vehicle-state data were collected. Driving visual load was quantified using non-negative matrix factorization based on pupil area change rate, blink frequency, fixation duration, and saccade velocity. Driving behavior risk was represented by the safe speed difference, and comprehensive driving risk was calculated using the improved entropy weight-Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method, followed by K-means clustering. Several machine-learning models were compared for risk-level prediction, and the best-performing Light Gradient Boosting Machine (LightGBM) model was further optimized using the whale optimization algorithm (WOA) and interpreted using SHapley Additive exPlanations (SHAP). RESULTS: Driving visual load clustered near tunnel portals and increased markedly at approximately 70 m before the entrance portal and 50 m before the exit portal. Across alignment conditions, driving visual load was highest in horizontal curve sections, followed by grade and straight sections. In contrast, comprehensive driving risk was highest in combined curve-grade sections, followed by horizontal curve sections. The entrance high-risk zone was concentrated from 30 m outside to 40 m inside the portal, whereas the exit high-risk range was narrower. CONCLUSIONS: Abrupt light-environment transitions and complex alignment conditions jointly increase driving visual load and risk clustering. The proposed framework linking visual load, behavioral risk, comprehensive risk, and risk prediction supports lighting optimization, speed management, and safety facility deployment.