Interpretable Electrophysiological Features of Resting-State EEG Capture Cortical Network Dynamics in Parkinsons Disease

Abstract

Parkinsons disease (PD) alters cortical neural dynamics, yet reliable non-invasive electrophysiological biomarkers remain elusive. This study examined whether interpretable EEG features capturing complementary aspects of neural dynamics can discriminate Parkinsonian neural states. A comprehensive set of interpretable features was extracted and grouped into Standard descriptors (spectral power, phase synchronization, time-domain statistics) and Dynamical descriptors (aperiodic activity, cross-frequency coupling, scale-free dynamics, neuronal avalanche statistics, and instantaneous frequency measures). A multi-head attention transformer classifier was trained using strict LOSO validation. Group-level comparisons were performed to identify electrophysiological differences associated with disease and medication state. Standard feature sets achieved strongest performance in discriminating medication states (PDoff vs PDon), whereas Dynamical performed competitively in contrasts between PD patients and healthy controls. Random feature ablation analyses indicated that Dynamical descriptors provide complementary information distributed across features while correlation analysis revealed low redundancy within both feature sets. Group-level comparisons revealed medication-sensitive reductions in delta power and voltage variance, modulation of neuronal avalanche statistics, persistent increases in theta phase synchronization in PD patients, and disease-related alterations in cross-frequency interactions. Traditional spectral and synchronization features primarily reflect medication-related neural modulation, whereas dynamical descriptors reveal broader alterations in cortical network organization associated with disease but also with medication. These findings support multivariate EEG representations as a promising framework for developing non-invasive biomarkers of PD.

Figures (9)

• Figure 1: Classification accuracy across diagnostic conditions using EEG feature configurations in a multi-head attention transformer model. Performance is shown for four tasks: three-class classification (CN, PDoff, PDon) and pairwise contrasts (CN--PDoff, CN--PDon, PDoff--PDon). Models were trained separately on the Standard and Dynamical feature sets using strict leave-one-subject-out (LOSO) cross-validation. Fusion corresponds to the decision-level model output obtained by averaging the softmax probabilities of the Standard and Dynamical feature sets. Bars indicate mean accuracy with 95% bootstrap confidence intervals (5000 samples). Statistical comparisons were performed using McNemar’s test on contingency tables of classification correctness with FDR correction (*p $<$ 0.05; NS, not significant).
• Figure 2: Receiver operating characteristic (ROC) curves for transformer models across diagnostic conditions. Curves are shown for models trained on the Standard and Dynamical feature sets and for Fusion, which corresponds to the decision-level model output obtained by averaging the softmax probabilities of the Standard and Dynamical feature sets. For the three-class problem, ROC curves were computed using a one-vs-rest strategy for each group (CN, PDoff, PDon). Performance is summarized by the weighted ROC--AUC (TPR, True Positive Rate; FPR, False Positive Rate).
• Figure 3: Accuracy of EEG feature configurations compared with random feature reductions across diagnostic contrasts (3-Class, CN--PDoff, CN--PDon, PDoff--PDon) using the Standard and Dynamical feature sets and Fusion. Fusion corresponds to the decision-level output obtained by averaging the softmax probabilities of the Standard and Dynamical sets. For each feature set, the solid horizontal line indicates the accuracy of the full model. Scatter points represent accuracies from one hundred models trained on randomly reduced feature subsets (50% features removed). The dashed horizontal line indicates the mean accuracy of the random ablations while the vertical lines represent its standard deviation.
• Figure 4: Pairwise correlations among features of the Standard EEG feature set. (A) Spearman correlation matrix computed across all subjects. The lower triangle displays correlation coefficients ($\rho$), while the upper triangle shows the corresponding 95% bootstrap confidence intervals (5000 resamples). (B) Matrix of FDR-corrected p-values (Benjamini--Hochberg) for the same pairwise comparisons, shown for the lower triangle only. The Standard feature set includes relative spectral power ($\delta$, $\theta$, $\alpha$, $\beta$, $\gamma$ bands), phase synchronization metrics—phase locking value (PLV) and phase lag index (PLI) for the same frequency bands—and time-domain signal statistics (mean voltage and variance). Relative power represents the percentage of each frequency band power to the total spectral power.
• Figure 5: Pairwise correlations among features of the Dynamical EEG feature set. (A) Spearman correlation matrix computed across all subjects. The lower triangle displays correlation coefficients ($\rho$), while the upper triangle shows the corresponding 95% bootstrap confidence intervals (5000 resamples). (B) Matrix of FDR-corrected p-values (Benjamini--Hochberg) for the same pairwise comparisons, shown for the lower triangle only. The Dynamical feature set includes measures of spectral aperiodic activity (aperiodic exponent and offset), aperiodic-corrected spectral peak characteristics (central frequency, bandwidth, and power), long-range temporal correlations via detrended fluctuation analysis (DFA), functional excitation--inhibition balance (fEI), bicoherence phase--amplitude coupling (bicPAC), instantaneous frequency dynamics (range[rng_freqSld], rate of change[dif_freqSld], inter-electrode correlation[corr_freqSld], and modulation frequency[mod_freqSld]), harmonic locking between frequency bands (HarmLock), and neuronal avalanche statistics, including avalanche size and deviation from a power-law scaling ($\kappa$ size).