AI-Driven Identification of Parkinson’s Disease Subtypes from Functional Brain Networks with Prototype-Guided Graph Learning
Main article
Abstract
Parkinson’s disease (PD) is the second most prevalent neurodegenerative disorder worldwide, characterized by significant clinical heterogeneity that complicates both diagnosis and therapeutic stratification. Resting-state functional MRI (rs-fMRI) provides non-invasive access to large-scale brain network disruptions associated with PD; however, conventional machine learning approaches lack the representational depth and interpretability required for reliable subtype identification. This paper proposes a Prototype-Guided Graph Learning (PGL) framework that integrates three computational advances: (1) a graph contrastive pretraining stage using a Graph Transformer encoder to derive generalizable brain-network embeddings from multi-site fMRI data; (2) a graph variational autoencoder (GVAE) that maps functional connectivity matrices to a structured latent space enabling uncertainty-aware topology learning; and (3) a prototype-guided classification module that partitions the latent space into clinically meaningful subtype prototypes, simultaneously yielding subtype labels and explainable subgraph signatures. Evaluated on a multi-site dataset comprising 312 PD patients and 184 healthy controls, PGL achieves an accuracy of 93.4%, an AUC of 0.941, and an F1 score of 0.927 for PD versus healthy control discrimination—outperforming all compared state-of-the-art baselines by at least 4.0 percentage points. For three-way PD subtype classification (olfactory-deficit, postural-instability, and tremor-dominant subtypes), the macro-average AUC reaches 0.893. Explainability analysis reveals subtype-specific disruption patterns involving the orbitofrontal cortex and amygdala for olfactory-deficit PD, the supplementary motor area and striatum for postural-instability PD, and the putamen and cerebellum for tremor-dominant PD, consistent with established neurobiological findings. The PGL framework offers a clinically actionable pathway for AI-assisted Parkinson’s diagnosis and subtype stratification.
