EEG Features Classify Mild Cognitive Impairment but Face Cross-Dataset Limits

In an exploratory evaluation of electroencephalogram (EEG)-derived indices across two independent cohorts, a machine learning classifier using lower alpha band features separated patients with mild cognitive impairment from healthy controls with within-dataset accuracies above 86 percent, though performance declined sharply when models were evaluated across datasets. Researchers evaluated EEG recordings from a Japanese dataset consisting of 7 patients with mild cognitive impairment and 10 healthy controls alongside an Iranian dataset comprising 29 patients with mild cognitive impairment and 32 healthy controls to examine within-dataset and cross-dataset classification performance.

For each cohort, common 16-channel EEG data were separated into five frequency bands, from which power spectrum values and Synchronization Likelihood (SL) values were calculated. The investigators also derived graph metrics based on Synchronization Likelihood values and computed both standardized and exponentially transformed versions of these indices to select candidate features for classification. A 34-dimensional feature subset selected from standardized and exponentially transformed EEG-derived indices in the lower alpha band achieved classification accuracies of 88.24 percent in the Japanese dataset and 86.89 percent in the Iranian dataset.

To minimize information leakage and optimistic bias, the researchers conducted nested leave-one-out cross-validation using exponentially transformed EEG-derived indices for the lower alpha band, yielding classification accuracies of 88.24 percent in the Japanese dataset and 83.61 percent in the Iranian dataset. However, in exploratory bidirectional cross-dataset evaluations designed to test whether classifiers generalize across different recording conditions, performance dropped to 40.98 percent when trained on the Japanese dataset and tested on the Iranian dataset, and 58.82 percent when trained on the Iranian dataset and tested on the Japanese dataset.

These findings suggest that resting-state EEG-derived indices for the lower alpha band contain informative signals useful for classifying mild cognitive impairment and healthy controls within a given cohort. At the same time, the authors noted that classifiers based on these indices may partly reflect dataset-specific recording conditions and characteristics, underscoring the challenge of establishing cross-dataset generalizability in dementia-related classification.

Subject of Research: Electroencephalogram feature evaluation for mild cognitive impairment classification

Article Title: Exploratory evaluation of electroencephalogram-derived features for classification of mild cognitive impairment in two independent datasets

Article References: Takeoka, C., Yada, T., Yamazaki, T., Kuroiwa, Y., Hirai, T., Fujino, K., & Nakane, M. (2026). Exploratory evaluation of electroencephalogram-derived features for classification of mild cognitive impairment in two independent datasets. Frontiers in Computational Neuroscience, 20, Article 1940507. https://doi.org/10.3389/fncom.2026.1940507

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1940507

Keywords: mild cognitive impairment, electroencephalogram, lower alpha band, machine learning, synchronization likelihood, graph metrics, cross-dataset evaluation, Exploratory, evaluation, electroencephalogram-derived, features, classification