Brain natural frequencies as physiologically meaningful biomarkers for machine-learning detection of Parkinson's disease

In this study, we investigated whether individual brain maps of natural frequencies derived from EEG can serve as physiologically meaningful biomarkers for Parkinsons disease (PD), a disorder characterized by altered oscillatory activity. Data were gathered from three independent, openly available EEG databases. Natural frequency maps were extracted for 57 PD participants and 57 sex- and age-matched healthy controls (HC). The group-level brain maps showed region-specific slowing of oscillatory activity in PD, particularly in frontomedial and left frontolateral cortices, as well as hemispheric asymmetries consistent with the lateralized onset of motor symptoms. A subset of 100 participants (50 PD, 50 matched controls) was then used to train and internally validate three machine learning classifiers--support vector machine (SVM), random forest, and logistic regression--under different PCA-based dimensionality reduction and feature normalization schemes. Among all models, SVM achieved the best performance (AUC {approx} 77%). This classifier was subsequently retrained on the full set of 100 participants and evaluated on an independent hold-out test set of 14 participants, achieving an AUC of 75.5% with 71.4% accuracy. Feature contribution analysis highlighted the regions that were most informative for classification toward each class, complementing the group-level observations. Overall, these results indicate that natural frequency mapping captures disease-related alterations in cortical dynamics and provides interpretable features for EEG-based diagnostic tools, opening new avenues for biomarker development in oscillopathies.