Predicting Driver Alertness by Learning Takeover Patterns from Vehicle Telemetry

Driver alertness is a critical factor in road safety, especially in conditionally automated vehicles where human drivers must respond to takeover requests. This study proposes a non-intrusive method to detect driver alertness using only vehicle telemetry data, avoiding the need for physiological sensors or camera-based monitoring. A publicly available dataset with takeover scenarios in urban and rural environments was used to train various machine learning and deep learning models. Features such as steering angle, braking intensity, and reaction time were used to classify driver alertness. Among the models tested, Tab Transformer, Neural Network, and Causal CNN achieved the highest performance, with accuracy up to 83% after hyperparameter tuning. To enhance model interpretability, SHAP analysis was applied to highlight the most influential features in the prediction process. The findings suggest that vehicle telemetry alone can effectively detect driver alertness, supporting the development of scalable and interpretable driver monitoring systems for real-world applications.