A coarse-to-fine machine-learning framework for identifying functional connectivity markers of cognitive impairment in Parkinson's disease.

Abstract

Background: Resting-state functional MRI (rs-fMRI) has been applied to investigate cognitive impairment (CI) in Parkinson's disease (PD). Nevertheless, reported functional connectivity (FC) alterations remain heterogeneous, partly due to reliance on linear analytical approaches and limited validation across datasets.

Objective: To develop a machine-learning framework for identifying generalizable FC markers of CI in PD.

Methods: Rs-fMRI data were obtained from an online cohort (for model training) and an independent local cohort (for external validation) of individuals with PD. Subjects were stratified according to the presence of CI. All images were preprocessed using an identical pipeline to derive whole-brain FC. A coarse-to-fine feature selection strategy was implemented, combining a genetic algorithm for global feature reduction with sequential feature selection using leave-one-out cross-validation.

Results: In the training dataset (n = 181), genetic algorithm-based selection reduced 13,366 ROI-pair features to 229, achieving an accuracy of 0.83. Subsequent sequential selection further reduced the feature set to 10 ROI pairs, improving accuracy to 0.92. In the validation dataset (n = 32), the classification accuracy was 0.88, with FC patterns showing lateralized cortical and cerebellar involvement.

Conclusion: The proposed framework identifies interpretable signatures of rs-fMRI-based markers associated with CI in PD and demonstrates the generalizability.