Objective:

Investigate whether machine learning analysis of resting-state functional MRI (rs-fMRI) can predict REM sleep behavior disorder (RBD) status in early Parkinson's disease (PD) patients and provide physiological insights into brain network differences underlying distinct phenotypes.

Background:

RBD predicts a malignant non-motor phenotype in PD that significantly influences disease progression and clinical outcomes. While previous studies have characterized non-motor phenotypes based on RBD presence or absence, the underlying physiological mechanisms distinguishing these phenotypes remain incompletely understood.

Design/Methods:

We analyzed rs-fMRI data from 271 early PD patients in the Parkinson's Progression Markers Initiative (PPMI) cohort. We used an 80/20 train-test split. Data preprocessing included MinMax scaling and principal component analysis (PCA), with five principal components selected via elbow method to explain maximum variance. rs-fMRI processing was performed using the CONN toolbox in MATLAB. We compared two feature sets: ROI-ROI connectivity features alone versus combined ROI-ROI and graph theory metrics, using the top 80 features for model training. Logistic regression was employed with 5-fold cross-validation (CV) and prediction performance on a test set for model validation.

Results:

The combined ROI-ROI and graph theory feature model achieved superior performance with balanced accuracy of 0.87, mean CV balanced accuracy of 0.85, and test set prediction 0.92. In contrast, using only ROI-ROI features yielded lower performance (balanced accuracy: 0.80, mean CV balanced accuracy: 0.78, test set balanced accuracy 0.79). Feature importance analysis identified key brain connectivity features contributing to RBD prediction.

Conclusions:

Our findings demonstrate that brain networks in early PD patients with and without RBD exhibit distinct physiological signatures detectable through rs-fMRI analysis. The enhanced predictive performance achieved by incorporating graph theory metrics suggests that network topology provides crucial information beyond simple connectivity patterns. This neuroimaging-based approach for RBD prediction offers potential clinical utility for stratifying patients and predicting long-term disease progression trajectories in early PD.