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

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Parkinson's disease (PD) diagnosis is challenging due to potential errors and underutilization of multimodal data.
• Current diagnostic methods often lack systematic integration of expert domain knowledge.
• Early and accurate PD diagnosis is crucial for effective patient management.

Purpose of the Study:

• To develop M²KD-Net, a multimodal and knowledge-driven diagnostic framework for Parkinson's disease.
• To enhance PD diagnostic performance by integrating imaging, non-imaging clinical data, and expert insights.
• To address limitations in current diagnostic approaches by leveraging multimodal data and domain knowledge.

Main Methods:

• Developed M²KD-Net, a framework with three modules: multimodal feature extraction, expert feature modeling, and cross-modal interaction.
• Utilized contrastive learning for improved data alignment between imaging and non-imaging modalities.
• Incorporated structured expert annotations to encode domain-specific knowledge.

Main Results:

• M²KD-Net achieved 89.6% classification accuracy and an AUC of 0.935 on the Parkinson's Progression Markers Initiative (PPMI) dataset.
• The framework effectively distinguished Parkinson's disease patients from healthy controls.
• Demonstrated improved integration of heterogeneous features across different data modalities.

Conclusions:

• M²KD-Net offers a dependable, interpretable, and clinically useful solution for Parkinson's disease diagnosis.
• The knowledge-driven multimodal approach enhances diagnostic accuracy.
• This framework represents a significant advancement in leveraging complex clinical data for neurodegenerative disease diagnosis.