Reproduce “TIER-MoE: Trust-Informed Expert Routing via Conditional Modality Risk for Multimodal Fusion in Biomedical Classification” on OpenResearch

About this paper

Yu Chang, Anzhe Cheng, Chenwei Wu, Zhuoran Wang, Jiahao Chen, Tamoghna Chattopadhyay, Sophia I. Thomopoulos, Paul M. Thompson and 2 more

A framework for multimodal fusion, TIER-MoE, learns subject-specific modality reliability as conditional expected prediction loss, informing expert routing while preserving cross-modal complementarity. This approach consistently outperforms state-of-the-art methods in biomedical classification, achieving a Macro-F1 of 0.648 and a Brier score of 0.189 on the ADNI dataset, and shows robustness to unreliable data and zero-shot generalization.

Key insights

• Defining modality reliability directly as conditional expected prediction loss, derived from label-honest out-of-fold predictions, provides a more accurate and robust measure of data trustworthiness.
• Explicitly decoupling subject-specific modality reliability, cross-modal complementarity, and expert specialization within the architecture addresses the reliability–complementarity trade-off in multimodal fusion.
• A combination of a reliability-weighted unimodal prediction branch and a sparsely routed expert interaction branch leads to superior overall predictive performance and better calibrated probabilities.

Summary by alphaXiv.