A Comparative Analysis on Different Feature Extraction Techniques for EEG-based Mental Workload Classification

Mental workload (MWL), the cognitive effort required to perform tasks, is a critical factor in fields like clinical medicine, system design, and industrial productivity. This study enhances MWL classification accuracy by exploring feature extraction methods from EEG data, including effective brain connectivity, statistical features, and power spectral density (PSD). Using the Simultaneous Task EEG Workload (STEW) dataset from 48 subjects, we optimized preprocessing with bandpass filtering, artifact removal, and re-referencing. Brain connectivity features were extracted using the direct directed transfer function (dDTF), with the top 30 connections selected per frequency band. Statistical features and PSD were also computed, as PSD is a key indicator of MWL. Various feature combinations were evaluated using machine learning models with sevenfold cross-validation. Results showed that statistical features combined with PSD achieved 93% accuracy using neural networks (NN) in the gamma band. Support vector machines (SVM) reached 86% accuracy with connectivity features alone, while decision trees (DT) and ensemble learning (EL) achieved 89% accuracy in the theta band using connectivity, statistical features, and PSD. Naive Bayes achieved 89.3% accuracy with connectivity and PSD features. The highest accuracy (93%) was observed using connectivity and statistical features in the theta band. Explainable AI (XAI) identified the most impactful features, highlighting the importance of feature selection. This study demonstrates that combining statistical features and PSD in the gamma band, along with connectivity-based features in the theta band, significantly improves MWL classification accuracy.