Missing-Modality Robustness in Clinical Multimodal Prediction: A Mechanism-Level Review

Multimodal medical artificial intelligence (AI) aggregates complementary evidence from imaging, omics, pathology, and clinical records, yet hospital deployment routinely confronts structurally incomplete inputs. This review synthesizes 77 algorithmic studies on missing-modality robustness in clinical prediction, selected through a Scopus search (16 May 2026) and a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) study-selection pathway, and organizes them in a three-family taxonomy: imputation and generative reconstruction (Family A), representation learning and robust fusion (Family B), and knowledge transfer and distillation (Family C). Cross-tabulating families against segmentation (Seg), diagnosis (Dx), and survival prognosis (Surv) endpoints yields 107 family–task assignments, within which Family B is most frequently represented (51 assignments), while reconstruction and relational transfer provide complementary approaches where spatial fidelity matters. Missingness is characterized along data dimensionality, temporal stage, and selection mechanism, and each study is mapped onto this joint space. At least 23 of the 77 studies (29.9%) test robustness by artificially withholding clinical inputs at test time, and 6 (7.8%) evaluate three or more missingness levels. Clinical transport requires validation under selective acquisition, including possible missing-not-at-random (MNAR) mechanisms. We therefore connect the taxonomy to explicit input-pattern validation and mechanism-aware evaluation. We outline benchmark, foundation-model, and federated priorities for deployment-credible missing-modality research.