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

Parkinson's disease (PD) affects 10 million globally, with accurate staging essential for personalized treatment planning. Current UPDRS assessments achieve < 93% accuracy due to subjective clinical judgment and unimodal data limitations, failing to capture complex genetic-neuroimaging-clinical interactions driving disease heterogeneity. This study introduces MAFNet, a novel deep learning framework pioneering Iterative Adaptive Vold-Kalman Filter (IAVKF) temporal denoising, Accelerated Binary Particle Swarm Optimization (ABPSO) swarm feature selection, Multilayer Perceptron-Lagrangian Support Vector Machine (MLP-LSVM) classification, and Graph-Attention Based Multimodal Fusion Network (GAMF). Applied to PPMI cohort (200 patients) with genetic SNPs (50), neuroimaging voxels (1,024), and UPDRS-III scores, the end-to-end pipeline delivers 97.6% accuracy, 98.2% precision, 96.8% recall, and 97.3% F1-score-outperforming CNN (92.4%), Autoencoder (90.8%), InceptoFormer (96.6%), and HCT (97.0%). IAVKF boosts SNR + 15.2dB (+ 2.9% accuracy vs. PCA/t-SNE); ABPSO reduces 1,276→340 features (73% reduction); regularization cuts overfitting gap to 0.9% (vs. 4.2% baseline). SHAP interpretability validates clinical plausibility (top predictors: LRRK2 SNPs, UPDRS-III tremor, hippocampal volume). Five-fold CV confirms stability with the Indian cohort external validation. Real-time inference (0.2s/patient, RTX 3090) enables clinical deployment. Future scope includes longitudinal temporal modelling, modality-agnostic fusion, edge deployment, federated learning, and extension to Alzheimer's/ALS. MAFNet transforms PD staging from subjective assessments to objective precision medicine, enabling biomarker discovery, progression forecasting, and personalized therapies across diverse global populations.