A multimodal MRI framework employing machine learning for detecting beginning cognitive impairment in Parkinson's disease

Abstract

Cognitive deficits affect up to half of the patients with Parkinson's disease (PD) within a decade of diagnosis, placing an increasing burden on patients, families and caregivers. Therefore, the development of strategies for their early detection is critical to enable timely intervention and management. This study aimed to classify cognitive performance in patients with PD using a binary support vector machine (SVM) model that integrates structural (high-resolution anatomical) and functional connectivity (FC; resting state) MRI data with clinical characteristics. We hypothesized that PD patients with beginning cognitive deficits can be detected through MRI in combination with machine learning. Data from 38 PD patients underwent extensive preprocessing, including large-scale FC and voxel-based analysis. Relevant features were selected using a bootstrapping approach and subsequently trained in an SVM model, with robustness ensured by 10-fold cross-validation. Although clinical parameters were considered during feature selection, the final best-performing model exclusively comprised imaging features-including gray matter volume (e.g., anterior cingulate gyrus, precuneus) and inter-network functional connectivity within the frontoparietal, default mode, and visual networks. This combined model achieved an accuracy of 94.7% and a ROC-AUC of 0.98. However, a model integrating clinical and only functional MRI data reached similar results with an accuracy of 94.7% and a ROC-AUC of 0.90. In conclusion, our findings demonstrate that applying machine learning to multimodal MRI data-integrating structural, functional, and clinical metrics-could advance the early detection of cognitive impairment in PD and could therefore be used to support timely diagnosis.

Keywords: Parkinson’s disease; cognitive impairment; functional neuroimaging; machine learning; magnetic resonance imaging; support vector machine; translational neuroscience.

FIGURE 1

Overview of the machine learning pipeline. The pipeline integrates input features (clinical, GMV, and FC) and the target (MoCA group classification score) through sequential stages of feature selection, classification, and model evaluation. Time series data were extracted from all voxels of preprocessed resting-state fMRI volumes, then parcellated into 17 networks using the Yeo2011 atlas, and Pearson correlations were computed between network pairs. T1-weighted (MPRAGE) volumes were used to calculate gray matter volumes (GMV) for brain regions defined by the Neuromorphometrics atlas. Along with clinical data, all MRI features were z-score normalized, and three distinct feature sets were created (1. step: Feature reduction): one based on clinical data and solely GMV features, one on clinical data with and solely FC features, and one combining clinical data with both imaging modalities. For each scenario, bootstrapping with 1,000 iterations was applied to identify the 50 most consistently selected features (for details, see Methods—Feature Reduction). Feature selection (2. step) then identified the optimal combination of 1 to 6 features (i.e., the combination that yields the highest mean accuracy over 10 cross-validation loops), resulting in a total evaluation (3. step) of 18 models (6 feature combinations per scenario across 3 scenarios). FC, functional connectivity; H&Y, Hoehn and Yahr scale; LEDD, Levodopa equivalent daily dose; UPDRS III, Unified Parkinson’s Disease Rating Scale. Figure created with draw.io.

FIGURE 2

Circular diagrams illustrating the selected FC features for n = 6 features. On the left side, the features from the second scenario are shown (clinical and FC features). The features from the third scenario are displayed on the right side (clinical, GMV and FC features). The clinical and GMV features are shown in boxes to the right of the respective diagram. ACgG, anterior cingulate gyrus; PrG, precentral gyrus; Calc, calcarine cortex; DMN, default mode network; FP, frontoparietal network; VIS, visual network; MOT, motor network; DAN, dorsal attention network; LIM, limbic network; VAN, ventral attention network.