Aperiodic EEG spectral features as consistent indicators of Parkinson's disease: a machine learning approach...

A major challenge in Parkinson's disease (PD) research is the identification of non-invasive neurophysiological markers that can distinguish PD patients from healthy controls while providing interpretable information about cortical dysfunction. This study investigates whether aperiodic EEG spectral parameters extracted using Fitting Oscillations and One-Over-F decomposition can serve as candidate indicators for exploratory PD classification and medication-state analysis. Resting-state EEG data from 15 PD patients and 16 healthy controls were analyzed. For PD participants, both medicated and unmedicated recordings were considered. Periodic and aperiodic powers spectral densities were separated out, and channel-wise exponents and offsets were derived. A leakage-free machine-learning pipeline was implemented using stratified subject-grouped tenfold cross-validation, in which normalization and hybrid t-test, MRMR, and ReliefF feature selection were performed strictly within each training fold. The Random Forest classifier achieved the best performance, with an accuracy of 0.843 ± 0.188, a pooled out-of-fold accuracy of 0.848, and a pooled AUC of 0.811. Permutation testing indicated that the observed performance exceeded chance-level classification. FDR-corrected statistical analysis showed significant group-level differences in global exponent and offset, with regional effects mainly involving frontal, temporal, and central areas. Medication-state analysis suggested that the global exponent was sensitive to dopaminergic state, whereas offset effects were weaker. Overall, the findings suggest that aperiodic EEG parameters may provide interpretable candidate markers for PD-related cortical alterations. However, due to the limited public dataset size and lack of external validation, the results should be considered exploratory and require confirmation in larger independent cohorts.