Bridging Imaging and Clinical Scores in Parkinson's Progression via Multimodal Self-Supervised Deep Learning

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Neurodegenerative diseases present complex research challenges requiring advanced modeling.
• Latent generative models offer data-driven approaches to understand neurodegeneration within the manifold hypothesis framework.

Purpose of the Study:

• To develop a joint multi-modal, common latent generative model for a comprehensive understanding of Parkinson's disease (PD).
• To leverage coupled variational autoencoders (VAEs) for modeling neuroimaging and clinical data from the Parkinson's Progression Markers Initiative (PPMI).

Main Methods:

• Utilized coupled variational autoencoders (VAEs) to jointly model a common latent space for neuroimaging and clinical data.
• Investigated alternative loss functions, normalization procedures, and model interpretability for latent generative models.
• Employed data from the Parkinson's Progression Markers Initiative (PPMI) dataset.

Main Results:

• Achieved high prediction accuracy for clinical symptomatology (UPDRS), with R2 up to 0.86 (same-modality) and 0.441 (cross-modality using neuroimaging).
• Demonstrated the model's capability in predicting clinical outcomes based on integrated multi-modal data.
• Highlighted the interpretability and explainability of the latent generative model.

Conclusions:

• The proposed framework provides a foundation for advancing clinical research and decision-making in Parkinson's disease.
• The model offers direct interpretability for understanding neuroimaging patterns associated with PD.
• This approach enhances the comprehensive understanding of the neurodegenerative landscape in PD.