Prediction of driver alertness levels on mountain roads using machine learning models: A naturalistic driving study...

A method for evaluating driver alertness on mountain roads was developed to enhance dynamic safety monitoring in high-risk sections. An indicator system integrating human and environmental factors was established, with 13 variables used for alertness classification and 17 initial variables applied for quantification.

Field tests were conducted in Guizhou, China, where data on drivers' heart rates, eye movements, and demographics were collected. Kernel principal component analysis (KPCA) was employed to extract four representative factors from 13 driving-related indicators. K-means clustering was employed to categorize drivers into high- and low-alertness groups. Logistic regression scoring, XGBoost, and Tabular Prior-Data Fitted Network (TabPFN) models were developed to assess driver alertness probabilities.

Superior performance was demonstrated by the XGBoost model, achieving an area under the receiver operating characteristic curve (AUC) value of 0.97 compared with 0.80 for logistic regression scoring model, indicating improved classification accuracy and robustness. TabPFN further enhanced performance, yielding the highest results (AUC: 0.98; KS: 0.85; F1: 0.90), confirming its robustness and adaptability in small-sample, high-dimensional data settings.

Driver alertness levels were effectively quantified by integrating human and environmental factors. TabPFN model was found to be more effective than logistic regression scoring and XGBoost, making it the preferred approach for monitoring driver alertness on mountain roads.