Comparative Machine Learning Classification of Age Groups Using Resting-State EEG Spectral Features from OpenNeuro ds005385

Resting-state electroencephalography (EEG) provides spectral information on age-related changes in brain activity. This study developed a leakage-controlled machine learning framework for classifying age groups using channel-wise spectral band-power features from the OpenNeuro ds005385 dataset. The dataset consisted of 64-channel resting-state EEG recordings from 608 participants aged 20–70 years, grouped into Adult, Middle-aged, and Elderly classes. Non-overlapping 2-second epochs were generated, while all data partitioning was performed at the subject level to prevent data leakage. Power spectral density was estimated using Welch’s method and summarized into delta, theta, alpha, beta, and gamma bands. Five classifiers were evaluated: Random Forest, K-Nearest Neighbors, XGBoost, Multilayer Perceptron, and Support Vector Machine. Statistical reporting included Wilson 95% confidence intervals and 5-fold group cross-validation statistics. Multilayer Perceptron achieved the best held-out performance, with 78.87% accuracy, 78.05% macro F1-score, and 0.921 macro AUC, followed by Support Vector Machine with 77.80% accuracy, 77.40% macro F1-score, and 0.918 macro AUC. Because paired prediction vectors were unavailable, model superiority was interpreted descriptively rather than as statistically significant. This study provides a reproducible comparative benchmark for EEG-based age-group classification using interpretable spectral features and subject-level validation.