Enhancing Parkinson's Disease Staging: An Integrative Deep Learning Framework for Multimodal Feature Selection.

Area of Science:

• Computational neuroscience
• Artificial intelligence in medicine
• Biomedical data science

Background:

• Parkinson's disease (PD) impacts millions globally, necessitating accurate staging for personalized treatment.
• Current staging methods (e.g., UPDRS) have <93% accuracy due to subjective clinical judgment and limited data integration.
• Disease heterogeneity arises from complex interactions between genetic, neuroimaging, and clinical factors.

Purpose of the Study:

• To introduce MAFNet, a novel deep learning framework for accurate and objective Parkinson's disease staging.
• To integrate multimodal data (genetics, neuroimaging, clinical scores) for improved PD classification.
• To enhance clinical decision-making and enable precision medicine for Parkinson's disease.

Main Methods:

• Developed MAFNet, a deep learning pipeline incorporating Iterative Adaptive Vold-Kalman Filter (IAVKF) for temporal denoising, Accelerated Binary Particle Swarm Optimization (ABPSO) for feature selection, and a Graph-Attention Based Multimodal Fusion Network (GAMF).
• Applied the framework to a cohort of 200 patients from the PPMI study, utilizing genetic SNPs, neuroimaging data, and UPDRS-III scores.
• Validated the model's performance and interpretability using SHAP analysis and external validation with an Indian cohort.

Main Results:

• MAFNet achieved 97.6% accuracy, 98.2% precision, 96.8% recall, and 97.3% F1-score, outperforming existing models like CNN and Autoencoder.
• IAVKF improved signal-to-noise ratio by +15.2dB, while ABPSO reduced feature dimensionality by 73% (1,276 to 340).
• SHAP analysis identified LRRK2 SNPs, UPDRS-III tremor, and hippocampal volume as key predictors, confirming clinical relevance. Real-time inference achieved 0.2s/patient.

Conclusions:

• MAFNet provides a highly accurate, objective, and efficient method for Parkinson's disease staging, surpassing traditional approaches.
• The framework's ability to fuse multimodal data and provide interpretable results facilitates biomarker discovery and personalized treatment strategies.
• MAFNet represents a significant advancement towards precision medicine in neurology, with potential for clinical deployment and extension to other neurodegenerative diseases.