Decoding driving states based on normalized mutual information features and hyperparameter self-optimized Gaussian kernel-based radial basis function extreme learning machine

Abstract

This study presents an analysis of driver's unfavorable driving states (UDS) using normalized mutual information (NMI) features and a hyperparameter self-optimized radial basis function extreme learning machine (RBF-ELM). By computing the mutual information across different frequency bands (including delta, theta, alpha, beta, and gamma frequency bands) in EEG signals, brain functional connectivity matrices are constructed to reveal the nonlinear coupling relationships between brain regions. The introduction of NMI reduces the effects of signal dimensionality differences, which ensures the comparability of features across subjects. After preprocessing and band-pass filtering of EEG signals, NMI features from five frequency bands are extracted, and RBF-ELM is then employed for distinguishing UDS. In the RBF-ELM model, an automatic hyperparameter optimization approach is implemented, combining grid search and five-fold cross-validation to select the optimal number of hidden layer neurons and regularization parameters. The experimental results show that the NMI features from the beta band provide excellent classification performance, achieving an accuracy of 94.06 % in detecting UDS. Moreover, the hyperparameter self-optimized RBF-ELM model exhibits outstanding performance on the test set, with an area under the receiver operating characteristic (ROC) curve (AUC) value of 0.9915. Compared to classic machine learning algorithms, the proposed method outperforms support vector machine, ensemble learning, linear discriminant analysis, logistic regression, neural networks, and k-nearest neighbors in terms of accuracy, sensitivity, precision, and specificity. The method presented in this paper provides a promising solution for real-time monitoring of drivers' psychological states and fatigue warning.