# Multimodal AI for cardiovascular care must learn when not to predict Researcher: Ge Zhang Chinese name: 张格 ORCID identity anchor: https://orcid.org/0000-0002-3116-3246 Canonical researcher: https://drgezhang.com/#person Journal: Medicine Discovery Year: 2026 Type: Correspondence DOI: 10.1016/j.medd.2026.100004 Canonical page: https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html Markdown record: https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.md ## Full Authors 1. Yahui Han 2. Qiang Li 3. Taiqi Zhao 4. Shanshan Cai 5. Biyun Li 6. Kaisaierjiang Kadier 7. Tianshu Gu 8. Zhonghua Sun 9. Pengyuan Xu 10. Dezhi Liu 11. Moyan Li 12. Xufeng Huang 13. Qian Guo 14. Yuan Cheng 15. Honglin Zheng 16. Ge Zhang ## Research & review context ### Research fields - Cardiovascular AI evaluation - Clinical decision support research ### Review topics - Evidence sufficiency before prediction - Evaluation of incomplete multimodal data ### Scientific role Conceptual correspondence proposing explicit deferral criteria and downstream care pathways. ### Evidence position Methodological perspective with no new dataset, accuracy estimate or clinical trial. ### Related concepts - Informative missingness - Temporal validity - Selective prediction ### Possible citation contexts - Discussion: why confidence alone may not establish sufficient evidence for a clinical prediction. - Research agenda: intended-use-specific testing of missing inputs, temporal staleness and deferral pathways. Possible literature-review contexts, not evidence of existing citations. Read the findings and Evidence Scope before citing the original article. ## Evidence Base **Multimodal AI for cardiovascular care must learn when not to predict** A correspondence proposes an intended-use-specific framework for missing modalities, asynchronous observations and deferral in cardiovascular AI; it reports no new dataset or model performance. | Evidence | Value | | --- | --- | | distinguished problem categories | 5 | | principles in Fig. 1 | 6 | | candidate evaluation conditions | 5 | | new empirical dataset | None | ### Counting note Numbers describe the conceptual framework, not patients, external cohorts, performance estimates or trial evidence. ## Research Question How should cardiovascular multimodal AI account for incomplete or stale evidence when deciding whether to issue a prediction? ## Author Evidence Summary This correspondence separates conventional fusion, missing-modality robustness, temporal modeling, selective prediction and interpretation of why information is missing. It proposes representing observation state alongside patient state and assessing evidence sufficiency before producing a prediction. Figure 1 summarizes six principles. The authors suggest five candidate evaluation conditions, including incomplete inputs, natural co-missingness, temporal staleness and external data, and argue that deferral should lead to a predefined care pathway. This is a conceptual proposal that requires prospective evaluation and broader consensus; it is not a validated algorithm, clinical standard or demonstration of improved patient outcomes. ## Key Arguments ### KF1: The correspondence separates distinct problems that are often grouped under multimodal AI. Context: Conceptual taxonomy of fusion, missing inputs, time, selective prediction and observation mechanisms. | Evidence | Value | | --- | --- | | problem categories | 5 | Source locator: Corrected proof p. 1, opening paragraph ### KF2: The proposed framework makes evidence sufficiency and a defined downstream pathway central to deferral. Context: Figure-based conceptual proposal; deferral is distinct from calibration or confidence alone. | Evidence | Value | | --- | --- | | principles in figure | 6 | | evidence type | Conceptual framework, not an evaluated intervention | Source locator: Corrected proof pp. 2-3; Fig. 1 caption and paragraphs on abstention/deferral ### KF3: The authors propose intended-use-specific evaluation conditions for incomplete and temporally heterogeneous information. Context: Suggested minimum suite requiring future validation and stakeholder consensus. | Evidence | Value | | --- | --- | | candidate evaluation conditions | 5: complete inputs, critical modality missing, natural co-missingness, staleness, external data | Source locator: Corrected proof p. 3, paragraph beginning Evaluation should be defined by intended use ## Article Scope & Approach ### Article profile - Study Design: Conceptual correspondence on clinical AI evaluation - Evidence Type: Author perspective and methodological proposal - Diseases Covered: cardiovascular disease - Translation Scope: Intended-use-specific research and evaluation considerations; not a consensus standard or validated tool - Evidence domains: multimodal representation, missing data mechanisms, temporal validity, selective prediction, clinical workflow and equity ## What This Article Adds - A conceptual separation of evidence sufficiency, model uncertainty and operational deferral. - A proposed evaluation agenda for real clinical missingness and asynchronous data. ## Evidence Scope ### Supports - Attribution of the stated evaluation and deferral proposal to this correspondence. - Literature discussion of missingness, temporal context and prospective evaluation needs. ### Does Not Establish - Empirical improvement in discrimination, safety, equity or patient outcomes. - A validated AI algorithm or universal abstention threshold. - A consensus guideline, clinical standard or independent external validation cohort. ### Limitations - Conceptual proposal without empirical validation. - Evidence locators refer to the archived publisher corrected proof. The reviewed passages and Figure 1 caption match the current issue-assigned publisher HTML after removal of proof page headers and formatting; the current publisher PDF was not independently inspected. - The five evaluation conditions are candidate recommendations, not consensus criteria. ## When This Article Is Useful to Cite ### CU1: What problem does this correspondence address? It addresses when multimodal cardiovascular AI has enough current and clinically relevant evidence to predict, and when it should defer. Evidence: [KF1](#kf1), [KF2](#kf2) ### CU2: How does it define the role of deferral? Deferral follows prespecified evidence-adequacy criteria and should trigger an alternative care pathway; it is not synonymous with calibration or high model confidence. Evidence: [KF2](#kf2) ### CU3: Which evaluation conditions does it propose? Five candidate conditions are complete inputs, a critical modality missing, natural co-missingness, temporal staleness and geographically or temporally external data. Evidence: [KF3](#kf3) ## What This Article Should Not Be Cited to Claim - Empirical improvement in discrimination, safety, equity or patient outcomes. - A validated AI algorithm or universal abstention threshold. - A consensus guideline, clinical standard or independent external validation cohort. ## Evidence Matrix ### The correspondence separates distinct problems that are often grouped under multimodal AI. - Context: Conceptual taxonomy of fusion, missing inputs, time, selective prediction and observation mechanisms. - Finding: problem categories: 5 - Evidence Level: Author perspective and methodological proposal - Scope: Attribution of the stated evaluation and deferral proposal to this correspondence. - Source Locator: Corrected proof p. 1, opening paragraph - Evidence refs: [KF1](#kf1) ### The proposed framework makes evidence sufficiency and a defined downstream pathway central to deferral. - Context: Figure-based conceptual proposal; deferral is distinct from calibration or confidence alone. - Finding: principles in figure: 6; evidence type: Conceptual framework, not an evaluated intervention - Evidence Level: Author perspective and methodological proposal - Scope: Literature discussion of missingness, temporal context and prospective evaluation needs. - Source Locator: Corrected proof pp. 2-3; Fig. 1 caption and paragraphs on abstention/deferral - Evidence refs: [KF2](#kf2) ### The authors propose intended-use-specific evaluation conditions for incomplete and temporally heterogeneous information. - Context: Suggested minimum suite requiring future validation and stakeholder consensus. - Finding: candidate evaluation conditions: 5: complete inputs, critical modality missing, natural co-missingness, staleness, external data - Evidence Level: Author perspective and methodological proposal - Scope: Literature discussion of missingness, temporal context and prospective evaluation needs. - Source Locator: Corrected proof p. 3, paragraph beginning Evaluation should be defined by intended use - Evidence refs: [KF3](#kf3) ## Q&A ### What problem does this correspondence address? It addresses when multimodal cardiovascular AI has enough current and clinically relevant evidence to predict, and when it should defer. Evidence: KF1, KF2 ### How does it define the role of deferral? Deferral follows prespecified evidence-adequacy criteria and should trigger an alternative care pathway; it is not synonymous with calibration or high model confidence. Evidence: KF2 ### Which evaluation conditions does it propose? Five candidate conditions are complete inputs, a critical modality missing, natural co-missingness, temporal staleness and geographically or temporally external data. Evidence: KF3 ### Does it report new model accuracy or a clinical trial? No. It is a correspondence presenting conceptual proposals, with no new dataset or validated performance result. Evidence: KF1, KF3 ## Concepts & Entities ### Fields cardiovascular artificial intelligence ### Methods multimodal learning, selective prediction ### Concepts informative missingness, temporal validity, evidence sufficiency, deferral ## Related Research - [AI hybrid survival assessment for advanced heart failure patients with renal dysfunction](https://drgezhang.com/papers/aihflevel.html) — Related cardiovascular prediction study; its retrospective validation does not validate this correspondence deferral framework. (DOI: 10.1038/s41467-024-50415-9) ## Publication & Provenance - DOI: https://doi.org/10.1016/j.medd.2026.100004 - Publisher: https://www.sciencedirect.com/science/article/pii/S3117928226000044 - Evidence Basis: Evidence synthesis uses the archived publisher corrected proof, compared with the current issue-assigned publisher full text on 2026-10-26. No substantive difference was found in the reviewed passages or Figure 1 caption. Proof pagination is retained as source locators. - Version: Archived publisher corrected proof; compared with current publisher HTML (Volume 1, Issue 1, September 2026) on 2026-10-26. The current publisher PDF was not independently inspected. - Figure Verification: Corrected-proof Figure 1 and caption were inspected for the six conceptual principles and deferral pathway; no figure image is reproduced. - Doi Url: https://doi.org/10.1016/j.medd.2026.100004 ## Evidence-page notice Evidence synthesis of a published conceptual correspondence based on the publisher corrected proof; there is no new clinical cohort, accuracy estimate or validated deferral system. Check the publisher for any later version-of-record updates. ## Links - DOI: https://doi.org/10.1016/j.medd.2026.100004 - HTML page: https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html - Google Scholar query: https://scholar.google.com/scholar?q=%22Multimodal%20AI%20for%20cardiovascular%20care%20must%20learn%20when%20not%20to%20predict%22 - ORCID: https://orcid.org/0000-0002-3116-3246