Effectiveness of Class Weight Balancing in CNN Models for EEG-Based Detection of Driver Drowsiness

This study presents a deep learning approach to detect driver drowsiness using EEG signals, leveraging Power Spectral Density (PSD) features extracted from EEG data collected during a simulated driving session with 27 participants. PSD analysis revealed distinct frequency patterns: higher frequencies like beta waves were more prevalent in alert states, while lower frequencies such as theta and delta increased with drowsiness. Two configurations of a Convolutional Neural Network (CNN) were tested: Conf 1, without weight balancing, and Conf 2, with class weight balancing to address the imbalance between alert and drowsy states. Conf 1 achieved a higher macro accuracy of 0.824 but had low sensitivity for detecting drowsiness. In contrast, Conf 2 showed a more balanced performance, with a slightly lower accuracy of 0.711 but a significantly improved sensitivity for the drowsy class. These results indicate that incorporating weight balancing enhances the model's ability to detect drowsiness, making Conf 2 more suitable for real-world applications. Future improvements could include oversampling techniques and integrating additional physiological data, such as heart rate or eye movements, to further enhance detection accuracy.