Physiological signal-based drowsiness detection using machine learning: Singular and hybrid signal approaches

Abstract

Introduction: Drowsiness is one of the main contributors to road-related crashes and fatalities worldwide. To address this pressing global issue, researchers are continuing to develop driver drowsiness detection systems that use a variety of measures. However, most research on drowsiness detection uses approaches based on a singular metric and, as a result, fail to attain satisfactory reliability and validity to be implemented in vehicles.

Method: This study examines the utility of drowsiness detection based on singular and a hybrid approach. This approach considered a range of metrics from three physiological signals - electroencephalography (EEG), electrooculography (EOG), and electrocardiography (ECG) - and used subjective sleepiness indices (assessed via the Karolinska Sleepiness Scale) as ground truth. The methodology consisted of signal recording with a psychomotor vigilance test (PVT), pre-processing, extracting, and determining the important features from the physiological signals for drowsiness detection. Finally, four supervised machine learning models were developed based on the subjective sleepiness responses using the extracted physiological features to detect drowsiness levels.

Results: The results illustrate that the singular physiological measures show a specific performance metric pattern, with higher sensitivity and lower specificity or vice versa. In contrast, the hybrid biosignal-based models provide a better performance profile, reducing the disparity between the two metrics.

Conclusions: The outcome of the study indicates that the selected features provided higher performance in the hybrid approaches than the singular approaches, which could be useful for future research implications. Practical Applications: Use of a hybrid approach seems warranted to improve in-vehicle driver drowsiness detection system. Practical applications will need to consider factors such as intrusiveness, ergonomics, cost-effectiveness, and user-friendliness of any driver drowsiness detection system.

Keywords: Accuracy; Drowsiness; Features; Ground truth; Machine learning; Physiological signals; Sensitivity; Specificity.