TIER-MoE Enhances Multimodal Biomedical Classification with Risk-Guided Fusion

Researchers have introduced TIER-MoE, a novel Trust-Informed Expert Routing via Conditional Modality Risk model designed to enhance multimodal fusion, particularly in biomedical classification. Traditional multimodal models often struggle with situations where combining more evidence doesn't necessarily lead to better predictions, especially when certain modalities are unreliable, redundant, or poorly matched to specific experts. TIER-MoE addresses this by defining and learning a sample-specific modality reliability, quantified as the expected prediction loss from its unimodal predictor. This risk is derived from out-of-fold predictions, ensuring an unbiased assessment. The model then combines this estimated risk with expert-specific subspace compatibility to sparsely route modalities to the most appropriate experts, while a shared pathway maintains overall multimodal complementarity. Evaluations on four public biomedical datasets, including Alzheimer's disease and skin lesion malignancy, demonstrate that TIER-MoE outperforms state-of-the-art methods in predictive performance and probability calibration, showing strong generalization capabilities.