Mixture-of-Experts and Gating Mechanisms in Multimodal Biomedical Learning: A Routing-Centric Analysis

Clinical decision-making is fundamentally multimodal, requiring the joint interpretation of radiological images, pathology slides, electronic health records, and physiological signals. However, conventional fusion strategies rely on fixed combination rules that cannot adapt to modality quality, data completeness, or patient-specific diagnostic relevance. Gating mechanisms and Mixture-of-Experts (MoE) architectures address this limitation by enabling sample-adaptive, conditional integration of heterogeneous clinical data. Although recent surveys have examined MoE in large language models and from a broader big-data perspective, no dedicated survey has focused on gating and MoE architectures for biomedical multimodal fusion. Against this background, we present a systematic review of this rapidly evolving field. This survey proposes a novel three-axis taxonomy covering Gating Architecture, Application Level, and Supporting Techniques. It provides a structured comparative analysis of representative methods and traces the evolution of gating strategies. Four persistent open research directions are highlighted: multi-level simultaneous fusion, per-sample inference-time routing, evaluation beyond downstream accuracy, and interpretable routing decisions. To empirically ground the proposed taxonomy, we present a controlled comparison of six gating architectures on three biomedical benchmarks, VQA-RAD, SLAKE, and Derm7pt, demonstrating that the five Axis 1 gating types produce measurably distinct performance profiles and that no single architecture dominates across all settings. These contributions provide a structured reference for researchers designing adaptive multimodal fusion systems in clinical AI.