Detection of Cognitive Workload Using Optimized Number of EEG Features

Cognitive load analysis during mental tasks can help to diagnose individuals with neurodegenerative diseases like learning disability and develop personalized learning materials. The multichannel electroencephalogram(EEG) is one of the popular non-invasive devices used to assess cognitive load. For more insight analysis, it is important to identify the optimal number of EEG features consisting of electrodes and frequency bands that mostly contribute to identifying the cognitive load. In this paper, our aim is to find the optimal number of EEG features that are significant to distinguish brain states; before mental tasks (BMT) and during mental tasks (DMT). For the purpose of implementation, EEG data with 19 electrodes have been collected from open-access datasets for 34 healthy subjects while performing rest (BMT) or mental arithmetic tasks (DMT). The power spectral density (PSD) analysis is performed on the pre-processed EEG data to compute five frequency bands alpha, beta, theta, delta, and gamma. From the five frequency bands, 49 different power ratios are computed for each of the 19 electrodes, resulting in a total of 931 features. The ReliefF and Chi-Square tests are employed to rank the features according to their significance. The ranked features are provided one by one as input to three classifiers, k-nearest neighbour (KNN), random forest (RF) and XGBoost. Considering ReliefF ranking method with only 46 features, RF achieves the highest accuracy of $\mathbf{8 4. 9 6 \%}$. The study found Fp1, Fp2, F3 and F4 electrodes and $\gamma / \delta$ and $\gamma /(\beta+\delta)$ power ratios are the most significant features that primarily contribute to classifying BMT or DMT.