Classifying metro drivers' cognitive distractions during manual operations using machine learning and random forest-recursive feature elimination.

Metro drivers are more likely to trigger accidents if they suffer from cognitive distractions during manual driving. However, identifying metro drivers' cognitive distractions faces challenges as generally no obvious behavior can be found during the distractions. To address the challenge, this paper identifies metro drivers' cognitive distractions based on Electrocardiogram (ECG) signals collected by wearable devices in simulated driving experiments. The ECG signals are processed to generate ultra-short-term heart rate and heart rate variability (HR-HRV) features. The HR-HRV features are extracted by 30-s and 60-s time-windows in driving phase, and 25-s time-windows in parking phase, respectively. Machine learning approaches are developed to identify distractions (binary) and distinguish the degrees of distractions (multi-class). The optimal input features are determined by a random forest and recursive feature elimination (RF-RFE) algorithm. Results show that the DT with only one HR-HRV feature extracted from 30-s time-windows and XGBoost with 20 h-HRV features extracted from 60-s time-windows are optimal models for binary and multi-class classification for distractions during driving phase, respectively. The features including NN20, pNN20, SD1/SD2, Max-HR, Min-HR, and MEDNN are the most critical HR-HRV features associated with distractions. Cognitive distractions in parking phase are difficult to be detected using HR-HRV features.