JMIR Publications

ABSTRACT

Multimodal clinical decision support combines imaging, clinical text, structured records, and molecular data. It is usually judged by a single comparison: whether the fused model outperforms a model built on one modality. That comparison establishes less than it appears to. A higher score can reflect one dominant modality, an availability pattern rather than clinical signal, or a dataset shortcut, and it is measured under an input condition that deployment does not reproduce. In this viewpoint, we ask which of the reporting requirements a multimodal claim depends on are already met by current clinical-AI guidance, and which are not. We mapped 10 audit dimensions against the complete published item sets of the 3 most recent deployment-facing documents: FUTURE-AI, PROBAST+AI, and STARD-AI. Six dimensions are already covered, and work on them adapts existing guidance. One, robustness to a missing modality, sits beside an existing item that addresses a different construct. Three have no corresponding item in any of the 3 documents, and a single cause runs through them: none of these documents treats the individual modality as a unit of analysis. We therefore highlight 3 reporting gaps. First, modality ablation, because showing which inputs a model used differs from showing that it still works when one is unavailable. Second, claim provenance, because existing traceability items govern the accountability log and the origin of the dataset, leaving open whether a generated statement can be traced to its supporting evidence. Third, contradiction handling, because no item addresses 2 modalities that are individually clear and mutually inconsistent. We illustrate each gap against published systems, name the nearest existing item so that the distinction is visible, and set out the reporting that would close it. The missing items belong inside the guidance that already exists.