Short-term future prediction of driving risk using XGBoost with Personal, Physiological, Biological, and Driving Behavior Data

This study investigated the use of a combination of driving, physiological, and personal features to predict unsafe driving behaviors with large advances. The 54 subjects involved were monitored for one week before the beginning of the experiment. Their personal information was collected, as well as their sleeping quality, stress state, and daily activity. Afterwards, they underwent 12 driving simulation sessions, each divided into 4 laps. During sessions on the driving simulator, saliva samples, physiological data, and driving-based measures were collected. The salivary cortisol level was measured before and after each lap, serving as a stress biomarker. An XGBoost model was trained and used to predict the driving safety level of a lap based on data from the previous lap (10 minutes earlier). Driving safety is categorized based on a 3-level scale. The model identified safe driving, moderately unsafe driving, and severely unsafe driving with precision of 98%, 81%, and 96%.