Assessment models for driving comfort and fatigue status based on heart rate variability

This study, from an interdisciplinary perspective of human factors engineering and biomedical engineering, aims to develop a real-time assessment model for driver status to mitigate the risks associated with fatigue and discomfort during prolonged driving. The primary objective is to construct a machine learning model based on Heart Rate Variability (HRV) that enables continuous and objective monitoring of driver fatigue and comfort levels.

A total of 64 healthy drivers holding international Class B licenses with at least 2.5 years of driving experience were recruited. The experiment was conducted in a six-degree-of-freedom driving simulator employing a 2 (speed: low vs. high) × 2 (road type: smooth vs. rough) factorial design, simulating realistic driving conditions (low speed: 50 ± 5 km/h; high speed: 100 ± 10 km/h). Electrocardiogram (ECG) signals were continuously recorded using a BIOPAC MP150 system. Subjective ratings of comfort (on a 10-point scale) and fatigue scale (Multidimensional Fatigue Inventory, MFI) were collected synchronously. A comprehensive set of 85 HRV features was extracted from the ECG signals. Six machine learning regression algorithms were evaluated and optimized via five-fold cross-validation. Model performance was assessed using Root Mean Square Error (RMSE) and the coefficient of determination (R²).

The Random Forest model outperformed others in predicting both fatigue (RMSE = 14.55) and comfort (RMSE = 1.56). Feature importance analysis identified six HRV features as most contributory: HRV Triangular Index (HTI), Shannon Entropy (ShanEn), Minimum NN interval (MinNN), Geometric Index (GI), Lorenz Plot Index (PI), and Power of Asymmetry Segment (PAS). These features are physiologically linked to autonomic nervous system activity, providing a mechanistic basis for the model's predictions.

The developed HRV-based machine learning model demonstrates high reliability and potential for real-time application. It advances beyond traditional discrete classification by enabling continuous assessment of driver status. This research provides a quantitative tool for designing in-vehicle human-machine interfaces and early warning systems, contributing to the development of safer and more intelligent transportation systems. Future work should focus on validation in real-world driving conditions.