M2KD-Net: A multimodal multi-domain knowledge-driven framework for Parkinson's disease diagnosis

Abstract

Parkinson's disease (PD) is a challenging neurodegenerative condition often prone to diagnostic errors, where early and accurate diagnosis is critical for effective clinical management. However, existing diagnostic methods often fail to fully exploit multimodal data or systematically incorporate expert domain knowledge. To address these limitations, we propose M2KD-Net, a multimodal and knowledge-driven diagnostic framework that integrates imaging and non-imaging clinical data with structured expert insights to enhance diagnostic performance. The framework consists of three key modules: (1) a contrastive learning-based multimodal feature extractor for improved alignment between imaging and non-imaging data. (2) an expert feature modeling module that encodes domain-specific knowledge through structured annotations, and (3) a cross-modal interaction module that enhances the integration of heterogeneous features across modalities. Experimental results on the Parkinson's Progression Markers Initiative (PPMI) dataset show that M2KD-Net achieves a classification accuracy of 89.6% and an AUC of 0.935 in distinguishing PD patients from healthy controls. This evidence suggests that the developed method provides a dependable, interpretable, and clinically useful solution for PD diagnosis.