image: Cardiovascular care generates inherently multimodal data that are frequently incomplete, asynchronously acquired, and dynamically altered by clinical events and treatment. A clinically meaningful patient representation should therefore capture not only what was observed, but also the reliability and timing of each observation, patterns and reasons for missingness, and changes in clinical state over time. Multimodal information can then be encoded and interpreted within its temporal and clinical context. Before a prediction is issued, evidence sufficiency should be assessed to determine whether the available information supports prediction. When prespecified criteria are met, the model may provide clinical decision support; when evidence is insufficient, it should defer to an alternative pathway such as clinician review, additional testing, or usual care. Performance, calibration, model-derived uncertainty, clinical utility, deferral behaviour, generalisability, and equity should be evaluated separately across relevant care settings and patient populations. The lower panel summarises six principles for trustworthy multimodal cardiovascular AI: representing the patient rather than the dataset; treating missingness as clinically meaningful; respecting temporal and clinical context; assessing evidence sufficiency and supporting deferral; validating generalisability and equity; and aligning model outputs with clinical decision-making and workflow.
Credit: Ge Zhang / Medicine Discovery
A correspondence in Medicine Discovery argues that a model performing well only when every input is present may still be unusable for the patient in front of the clinician — and proposes evidence sufficiency, selective prediction, and deferral pathways as explicit design requirements
Cardiovascular medicine is inherently multimodal. Cardiac imaging provides anatomical and functional information; electrocardiography and physiological waveforms capture electrical and hemodynamic dynamics; electronic records provide longitudinal clinical context; and omics and wearable measurements add molecular and out-of-hospital resolution. Multimodal artificial intelligence (AI) is often presented as the natural solution to this fragmentation, on the expectation that combining complementary data may improve phenotyping, risk prediction, treatment selection, and longitudinal monitoring.
In practice, however, the complete multimodal patient rarely exists. A patient with suspected coronary disease may undergo computed tomography but not cardiovascular magnetic resonance; echocardiography may precede laboratory testing by months; and wearable data may cover only selected periods. Even within one modality, vendors, acquisition protocols, image quality, measurement frequency, and clinical indications vary. Recent reviews identify incomplete datasets, heterogeneous data characteristics, and missing modalities as persistent barriers to multimodal clinical AI.
A correspondence published in Medicine Discovery argues that the central problem is not how to fuse more data, but how to reason when clinically relevant data are incomplete, temporally misaligned, or no longer representative of the patient's current state. " A model that performs well only when every input is present and synchronised may solve a benchmark problem while remaining unusable for the patient in front of the clinician," said Dr. Ge Zhang, a corresponding author of the correspondence and a researcher at Zhengzhou University.
The correspondence distinguishes five linked but non-equivalent problems that are frequently treated as one: conventional multimodal fusion; robustness to missing modalities; temporally aware modeling; uncertainty-aware selective prediction; and clinically informed interpretation of why a modality is unavailable. The last of these connects the clinical data-generating process to the disease process, and is central to the framework the authors propose.
Missingness is a clinical observation
The authors argue that most multimodal AI research treats missingness as an anomaly to be cleaned, whereas a test may in fact be absent because it was not clinically indicated, was ordered but not completed, failed quality control, was unavailable, or was lost during data linkage. These situations produce the same empty field but have different meanings. Missingness can therefore carry information about patient state, clinician decisions, access to care, institutional resources, and the health system.
Modality availability, the authors suggest, should be represented as an observation state alongside, but not conflated with, patient state. They caution that using availability as a predictor is a predictive modeling choice rather than a causal claim about disease biology, and may introduce or amplify selection bias by encouraging models to learn institutional workflow rather than patient biology. Random modality dropout may improve robustness to absent inputs, but it does not reproduce naturally occurring informative co-missingness.
Timing changes what a measurement means
A second challenge is asynchrony. Cardiovascular observations are not collected at a single biological moment, and their meaning depends on timing, intervening events, and changes in disease state. The correspondence recommends that each use case prespecify the clinical index time, the prediction horizon, the look-back window, the at-risk population, and the time at which each input actually became available for decision-making. Ordering, acquisition, report sign-off, and availability in the clinical system should be distinguished, and interventions or state transitions should be recorded to avoid future-information leakage.
Temporal validity, the authors note, is modality-, disease-, and decision-specific. A six-month-old genomic measurement may remain relevant, whereas left ventricular ejection fraction may be invalidated by intervening revascularization or decompensation; a troponin result from several days earlier may be irrelevant to an acute presentation; and wearable rhythm data may require a different validity window.
Fusion strategy should follow clinical use
Fusion strategy, the correspondence argues, should follow the intended clinical use rather than architectural fashion. Early fusion may capture low-level interactions but usually requires strong alignment and is vulnerable to missing inputs; intermediate fusion can learn modality-specific representations before combining them in a shared space; and late fusion can support variable modality subsets only when an aggregation, gating, or routing scheme has been designed and evaluated for those subsets. No strategy is intrinsically superior across cardiovascular applications.
Deferral is not the same as uncertainty
Central to the framework is the operational definition of abstention, or deferral: a prespecified decision not to issue the model's intended prediction because the available evidence does not satisfy use-case-specific adequacy criteria. The authors stress that this is not synonymous with probability calibration, uncertainty estimation, or distribution-shift detection. Good aggregate calibration should not be treated as evidence of reliable case-level uncertainty, and high confidence should not be used to exclude distribution shift.
Deferral triggers need not be weighted equally. Clinically mandatory inputs, or failure of prespecified quality criteria, may function as non-compensatory hard stops, whereas graded signals such as uncertainty, staleness, cross-modal conflict, and distribution shift may be combined using a prespecified composite rule. As an illustration, the authors describe a hypothetical emergency-department model estimating 30-day major adverse cardiovascular events in patients with acute chest pain at the point of disposition. Such a model might defer when a current electrocardiogram or troponin result is unavailable or fails quality checks, when relevant information has become temporally invalid, or when prespecified uncertainty or shift criteria are exceeded. Coverage and selective risk would then be examined overall and by clinically relevant subgroup. Deferral, the authors emphasize, must trigger care rather than suspend it: clinician review, additional testing, or return to the usual clinical pathway should be specified in advance.
A complement to existing guidance, not a replacement
The framework complements rather than replaces existing clinical AI guidance. TRIPOD+AI addresses prediction-model reporting; PROBAST+AI assesses quality, risk of bias, and applicability; FUTURE-AI provides broader lifecycle guidance; and DECIDE-AI, CONSORT-AI, and SPIRIT-AI address clinical evaluation and trial reporting or protocols.
"Compliance with a reporting checklist is not, by itself, evidence that a model is safe," the authors write. The distinctive contribution, they argue, is an intended-use-specific framework for cardiovascular multimodality that separates observation state from patient state, preserves asynchronous clinical context, and makes evidence sufficiency, selective prediction, and downstream deferral pathways explicit.
Evaluation should be defined by intended use
The correspondence proposes five candidate conditions for a minimum evaluation suite, to be adapted to the population, setting, decision point, prediction horizon, available inputs, action threshold, and consequences of error for each application: complete inputs; each clinically critical modality unavailable in turn; naturally occurring patterns of co-missingness; clinically relevant degrees of temporal staleness; and geographically and temporally external data.
Reporting, the authors suggest, should cover discrimination and calibration among non-deferred predictions, coverage, selective risk or error across coverage levels, the frequency and causes of deferral, subgroup-specific error and deferral rates, and deterioration as information becomes older or less complete. "Critical modality," "realistic co-missingness," and "reliable prediction" are use-case-specific rather than universal, and current care or a simpler model can serve as comparators where appropriate.
Deferral and equity
Subgroup analysis should also determine whether elevated deferral rates reflect clinically appropriate uncertainty or unequal access to the evidence required for prediction. Persistent excess deferral among socioeconomically disadvantaged or other underserved groups should be treated as a potential equity concern rather than accepted as a neutral safety mechanism, prompting examination of input availability, deferral thresholds, and access to fallback care. Mitigation may require improving data availability or fallback pathways rather than simply relaxing deferral thresholds.
From development to deployment
Clinical translation, the authors suggest, should follow a staged pathway: retrospective model development, followed by geographically and temporally independent validation; silent prospective evaluation to determine whether predictions remain stable when generated from live clinical data without influencing care; early clinical studies assessing usability, human factors, workflow integration, failure modes, and communication of missing information, uncertainty, and deferral; and only then pragmatic trials with patient-centered and health-economic outcomes.
Implementing this requires datasets that retain provenance metadata on why tests were not performed; when observations were ordered, acquired, reported, and made clinically available; intervening clinical events; quality-control failures; reasons for data loss; and how inputs were transformed. Many publicly available datasets do not capture these fields consistently, and the authors argue that dataset curators should treat such information as core metadata if reproducible evaluation of deferral-aware multimodal AI is to become feasible.
Governance must also continue after deployment. A system that is safe at launch may deteriorate as populations, diagnostic pathways, imaging hardware, treatment practices, modality availability, or relationships among data sources change. Deploying institutions and model owners should define responsibility for surveillance and prespecify triggers for revalidation; when a trigger is crossed, the model should be revalidated and, where safety cannot be established, paused or rolled back.
Article information:
Multimodal AI for Cardiovascular Care Must Learn When Not to Predict.
Medicine Discovery, 2026.
Original article link: https://www.sciencedirect.com/science/article/pii/S3117928226000044
DOI: 10.1016/j.medd.2026.100004
About the journal
Medicine Discovery (ISSN: 3117-9282) is a peer-reviewed, international, open-access journal published by Zhujiang Hospital of Southern Medical University in collaboration with KeAi Publishing and distributed on the ScienceDirect platform. The journal is dedicated to publishing high-quality translational medical research that bridges mechanistic discovery and clinical implementation, and covers five sections spanning mechanisms and systems medicine, translational and clinical research, digital medicine and data-driven healthcare, therapeutic innovation and biomedical engineering, and methodology, standards, and implementation science.
Website: https://www.sciencedirect.com/journal/medicine-discovery
Submit: https://www.editorialmanager.com/medd/default2.aspx
Method of Research
Commentary/editorial
Article Title
Multimodal AI for cardiovascular care must learn when not to predict