Adapting artificial neural networks to a specific driver enhances detection and prediction of drowsiness

Monitoring car drivers for drowsiness is crucial but challenging. The high inter-individual variability observed in measurements raises questions about the accuracy of the drowsiness detection process. In this study, we sought to enhance the performance of machine learning models (Artificial Neural Networks: ANNs) by training a model with a group of drivers and then adapting it to a new individual. Twenty-one participants drove a car simulator for 110 min in a monotonous environment. We measured physiological and behavioral indicators and recorded driving behavior. These measurements, in addition to driving time and personal information, served as the ANN inputs. Two ANN-based models were used, one to detect the level of drowsiness every minute, and the other to predict, every minute, how long it would take the driver to reach a specific drowsiness level (moderately drowsy). The ANNs were trained with 20 participants and subsequently adapted using the earliest part of the data recorded from a 21st participant. Then the adapted ANNs were tested with the remaining data from this 21st participant. The same procedure was run for all 21 participants. Varying amounts of data were used to adapt the ANNs, from 1 to 30 min, Model performance was enhanced for each participant. The overall drowsiness monitoring performance of the models was enhanced by roughly 40% for prediction and 80% for detection.