Bayesian Relation Propagation for Interpretable Multimodal Medical Diagnosis

Multimodal medical models benefit from flexible interactions among heterogeneous inputs, but predictive objectives alone may lead to unstable or incidental feature dependencies. On the other hand, imposing predefined relational structures can improve structural transparency while suppressing useful predictive interactions. We propose a method of Bayesian Relation Propagation (BRP) that preserves the original predictive pathway while adaptively incorporating relational information through a separate relation-propagation branch. BRP first identifies stable feature relations that persist under repeated subsampling and uses them to determine a relational direction. A disease prior provides feature relevance retrieved using a calibrated predictive posterior, while predictive uncertainty and Bayesian network reliability jointly control the propagation strength. The relational direction and propagation strength are then combined through residual relation propagation to adaptively update the predictive representations. We further characterize sufficient conditions under which relation propagation can simultaneously reduce local relational discrepancy and prediction loss. On MILK10k and CheXchoNET, BRP achieves a Macro-F1 of 0.55 and an exact match ratio of 0.87, improving predictive performance over the strongest competing methods by 1.85% and 2.35%, respectively. Compared with a strong predictive baseline of SOTA, BRP further reduces normalized Structural Hamming distance by 51.9% and 56.5%, respectively, implying stronger agreement with expert-informed relational structures. Additional analyses further support the robustness and model-agnostic applicability of BRP. Our code will be released upon publication.