A multimodal Bayesian decision fusion model for Parkinson's disease prediction

The early symptoms of Parkinson's disease exhibit cross-modal heterogeneity (e.g., microexpression tremors, speech distortion), making it difficult for traditional unimodal intelligent diagnostic methods to comprehensively capture pathological features. Moreover, existing multimodal fusion strategies face challenges in achieving reliable decision-making in noisy environments due to static weight allocation and feature manifold distortion. To address these challenges, this paper proposes a Multimodal Bayesian Decision Fusion (MBDF) method. The method constructs a dynamic confidence evaluation module using kernel density estimation and Bayesian inference, enabling adaptive weight allocation for different modalities. In addition, a feature stability optimization module based on mean-variance statistical properties is designed to suppress high-dimensional noise interference. Experimental results demonstrate the effective synergy between the two modules, significantly enhancing diagnostic robustness in complex scenarios. The proposed method provides a scalable multimodal fusion paradigm for intelligent diagnosis of neurodegenerative diseases.