Classifying Features of Electroencephalography Signal to Detect Driver Drowsiness in the Early Drowsy Stage

Background and Objective: Driver drowsiness is one of the major reasons of severe accidents worldwide. In this study, an electroencephalography (EEG) measurement-based approach has been proposed to detect driver drowsiness. Materials and Methods: The driving tests were conducted in a driving simulator to collect brain data in the alert and drowsy states. Nineteen healthy men participated in these tests. The EEG signals were recorded from the central, parietal, and occipital regions of the brain. 12 features of EEG signal were extracted; then through neighborhood component analysis (NCA), a feature selection method, 6 features including mean, standard deviation (SD), kurtosis, energy, entropy, and power of alpha band in 11-15 Hz, where alpha spindles occur, were selected. For the drowsiness stages assessment, the Observer Rating of Drowsiness (ORD) was applied. Four classifiers including k-nearest neighbor (KNN), support vector machine (SVM), classification tree, and Naive Bayes were employed to classify data. Results: The classification trees detected drowsiness in the early stage with 88.55%. The classification results showed that if only single-channel P4 was used for detecting drowsiness, the better performance could be achieved in comparison to using data of all channels (C3, C4, P3, P4, O1, O2) together. The best performances were 93.13% which were obtained by the classification tree based on data of single-channel P4. Conclusion: This study suggested that the driver drowsiness was detectable based on single-channel P4 in the early stage.