Context-Aware Graph Neural Networks (CAGNN) for Multimodal Prediction of Parkinson’s Disease Dementia

Abstract

Parkinson’s disease dementia is a severe cognitive decline in up to 95% of Parkinson’s disease sufferers within 10–20 years after diagnosis. Diagnosis relies on comprehensive clinical evaluation; however, research indicates that underlying physiological changes to white matter tracts precede symptomatic presentation. Existing work has focused on symptomatic data or single imaging modalities, without exploiting 3D structural information available through fiber tractography. To the best of our knowledge, no existing framework effectively combines 3D fiber tractography and clinical population graphs for early PDD classification. A modular, context-aware geometric deep learning framework is developed to classify subjects into cognitive stages from baseline data, enabling future prediction. The method combines two graph neural network modules, processing white matter tractography as point-based graphs and inter-subject similarity using 16 clinical and cognitive features. Each module generates prediction probabilities, which are combined through a lightweight ensemble classifier to produce a final diagnostic label. Results demonstrate individual component elements outperform baseline models by approximately 10%, achieving accuracies between 75% and 82%. Similarly, a combined framework of both modules with an ensemble strategy further improves accuracy to 88%, exceeding individual module performance by over 10%, underscoring the importance of combining structural and contextual information for classification of cognitive states in Parkinson’s disease.