Frontal Lobe EEG Signals Accurately Detect Cognitive Load Across Multiple Datasets

A machine learning framework combining empirical Fourier decomposition and dragonfly-based feature selection achieved classification accuracies exceeding ninety-six percent across multiple benchmark datasets while identifying the frontal lobe as the most informative brain region for cognitive load detection. The authors reported that the approach reached peak classification accuracies of 97.13 percent on the MAT dataset, 96.48 percent on the STEW dataset, and 98.2 percent on the MLR dataset, while individual analysis of the frontal lobe attained 96.7 percent and 96.6 percent accuracy on the MAT and STEW datasets, respectively.

Cognitive load detection plays an essential role in understanding how the human brain responds to mental effort, cognitive demands, and external stimuli by tracking how these processes stimulate neural activity.

The experimental procedure analyzed electroencephalogram data recorded from individual brain lobes over discrete intervals lasting four seconds each. To extract meaningful patterns from these temporal segments, empirical Fourier decomposition was applied to systematically decompose the raw neural signals into ten distinct intrinsic mode functions, or IMFs.

This algorithm searched the extracted entropy feature space to identify and select the most informative subsets, refining the data representations before passing them to downstream machine learning classifiers.

The classification performance of the selected feature subsets was subsequently evaluated across three established benchmark datasets: the MAT, STEW, and MLR datasets. Different machine learning classifiers were deployed to evaluate the efficacy of the proposed EPFD and HLBDA pipeline across these distinct corpora. When classifying overall cognitive load, the framework achieved an accuracy of 97.13 percent on the MAT dataset and 96.48 percent on the STEW dataset using the K-nearest neighbors, or KNN, classifier. Furthermore, when evaluated on the MLR dataset, the framework reached an overall classification accuracy of 98.2 percent using an ensemble subspace K-nearest neighbors classifier, denoted as ESSK, confirming strong detection capabilities across varied experimental collections.

In addition to assessing aggregate whole-brain signals, the investigation examined lobe-wise cognitive load detection to provide detailed, lobe-by-lobe data regarding neural activity during cognitive tasks. The authors conducted individual lobe-wise analyses across five distinct brain lobes using data from the MAT and STEW datasets to evaluate their respective discriminative capabilities. Across these evaluations, the frontal lobe demonstrated superior discriminative capability for distinguishing cognitive load states across a variety of cognitive tasks. Specifically, utilizing the ensemble subspace K-nearest neighbors classifier, frontal lobe signals yielded classification accuracies of 96.7 percent on the MAT dataset and 96.6 percent on the STEW dataset, indicating that the frontal region serves as a particularly informative anatomical area for evaluating cognitive strain.

To maintain methodological rigor and ensure the validity of the reported classification rates, the researchers implemented a subject-wise leave-one-out cross-validation technique, known as LOOCV. This validation design was specifically employed to prevent data leakage between training and testing sets across individual subjects, thereby reducing the risk of overly optimistic performance estimates. The authors concluded that the integrated framework—combining empirical Fourier decomposition, hyper-learning binary dragonfly optimization, and K-nearest neighbor or ensemble subspace classifiers—demonstrated competitive classification performance when compared with recently reported cognitive load detection methods, reinforcing the value of lobe-specific electroencephalogram analysis for reliable mental workload assessment.

Subject of Research: Electroencephalogram-based cognitive load detection using feature selection and machine learning

Article Title: Cognitive load detection using hyper-learning dragonfly optimization for feature selection

Article References: Chervitha, K., & SHARMA, L. D. (2026). Cognitive load detection using hyper-learning dragonfly optimization for feature selection. Frontiers in Physiology, 17, 1952032. https://doi.org/10.3389/fphys.2026.1952032

Image Credits: AI Generated

DOI: 10.3389/fphys.2026.1952032

Keywords: cognitive load detection, electroencephalogram, empirical Fourier decomposition, intrinsic mode functions, hyper-learning binary dragonfly algorithm, feature selection, entropy-based features, K-nearest neighbors, ensemble subspace KNN, frontal lobe, lobe-wise analysis, leave-one-out cross-validation, MAT dataset, STEW dataset, MLR dataset, mental effort, neural activity