The automotive industry and scientific community are making efforts to develop innovative solutions that would increase successful driver performance in preventing crashes caused by drivers' health and concentration. High stress is one of the causes of impaired driver performance. This study investigates the ability to classify different stress levels based on capacitive electrocardiogram (cECG) recorded during driving by unobtrusive acquisition systems with different hardware implementations. The proposed machine-learning model extracted only four features, based on the detection of the R peak, which is the most reliably detected characteristic point even in inferior quality cECG. Another criterion for selecting the features is their low computational complexity, which enables real-time application. The proposed method was validated on three open data sets recorded during driving: electrocardiogram (ECG) recorded by electrodes with direct skin contact (high quality); cECG recorded without direct skin contact through clothes by electrodes built into a portable multi-modal cushion (middle quality); and cECG recorded through the clothes without direct skin contact by electrodes built into a car seat (lowest quality). The proposed model achieved a high accuracy of 100% for high-quality ECG, 96.67% for middle-quality cECG, and 98.08% for the lower-quality cECG.
Keywords: cECG filter; machine-learning model for stress level detection; unobtrusive health monitoring system.