Investigating the role of data preprocessing, hyperparameters tuning, and type of machine learning algorithm in the improvement of drowsy EEG signal modeling

Abstract

• Electroencephalography measured driver drowsiness in simulator tests. • Driver drowsiness was classified with four modeling approaches. • Six algorithms of DT, ET, KNN, MLP, RF, and SVC were applied to the dataset. • The type of algorithm is more effective than the modeling approach. • Tree-based ensemble algorithms are more appropriate for drowsy EEG signal modeling. Driver drowsiness leads to fatal road traffic accidents. The effects of drowsy driving on electroencephalogram (EEG) signal are well visible. Accordingly, classifying EEG signal with machine learning (ML) is known as a reliable and accurate drowsiness detection technique. The type of ML algorithm, data preprocessing, and tuning the hyperparameters influence classification results greatly. In this paper, the drowsy EEG signal of 20 participants was measured in simulator tests. Six ML algorithms of decision tree, extra trees, K-nearest neighbors, multi-layer perceptron, random forest, and support vector classification were trained and cross-validated with different approaches to understand how data preprocessing and tuning the ML hyperparameters can improve drowsy EEG signal modeling. Results indicated that the type of ML algorithm has a more notable effect on modeling accuracy and modeling error than the modeling approach. Data preprocessing generally improved modeling results. But tuning the hyperparameters with the random search method wasn't helpful. Comparison of algorithms showed that tree-based ensemble algorithms (extra trees and random forest) were the most accurate models. They are more practical for real-time applications of drowsy EEG signal modeling.