Intelligent Driver Distraction Detection Using Functional Near-Infrared Spectroscopy and Ensemble Learning with Feature Expansion

Automated detection of driver distraction is vital for improving road safety and reducing accidents caused by human error. This study investigates the use of functional near-infrared spectroscopy (fNIRS) signals to classify multiple types of distraction, cognitive, visual, and manual, in simulated driving scenarios. Unlike previous studies that focused on limited features or distraction types, we apply a comprehensive feature engineering approach combined with a permutation-based importance analysis in oxygenated hemoglobin, deoxygenated hemoglobin, and their combination. To enhance classification robustness, we introduce a feature-importance-driven selection strategy integrated into an Extreme Gradient Boosting (XGBoost)-based stacking ensemble framework. The primary AI contribution of this work is the integration of feature importance driven selection within an ensemble learning model, while the engineering application lies in applying this approach to fNIRS data for real-time driver distraction detection in intelligent transportation systems. The experimental results show that the proposed method outperforms benchmark machine learning models, achieving an overall accuracy of 67.04% in distinguishing three distraction levels, with per-class accuracies of 89.07 ± 3.75% (baseline), 73.15 ± 1.98% (driving without distraction), and 71.85 ± 1.15% (driving with distraction). These findings advance smart road safety research and recognition of human distractions.