Enhanced Drowsiness Detection Using Deep Learning: An fNIRS Study

In this paper, a deep-learning-based driver-drowsiness detection for brain-computer interface (BCI) using functional near-infrared spectroscopy (fNIRS) is investigated. The passive brain signals from drowsiness were acquired from 13 healthy subjects while driving a car simulator. The brain activities were measured with a continuous-wave fNIRS system, in which the prefrontal and dorsolateral prefrontal cortices were focused. Deep neural networks (DNN) were pursued to classify the drowsy and alert states. For training and testing the models, the convolutional neural networks (CNN) were used on color map images to determine the best suitable channels for brain activity detection in 0 similar to 1, 0 similar to 3, 0 similar to 5, and 0 similar to 10 second time windows. The average accuracies (i.e., 82.7, 89.4, 93.7, and 97.2% in the 0 similar to 1, 0 similar to 3, 0 similar to 5, and 0 similar to 10 sec time windows, respectively) using DNNs from the right dorsolateral prefrontal cortex were obtained. The CNN architecture resulted in an average accuracy of 99.3%, showing the model to be capable of differentiating the images of drowsy/non-drowsy states. The proposed approach is promising for detecting drowsiness and in accessing the brain location for a passive BCI.