Integrating Imaging and Clinical Information With Artificial Intelligence When Data Are Missing or Uncertain

Multimodal artificial intelligence can integrate medical imaging with clinical information to support diagnosis and prognosis. However, inputs are frequently missing, acquired at different times, or inconsistent, reflecting clinical judgement, measurement limitations, or barriers to care. This Review critically examines approaches to using available inputs, reconstructing missing information, integrating sources according to reliability, and incorporating clinical knowledge. Particular attention is given to foundation and vision-language models, informative missingness, and the distinction between predictive confidence and demonstrated reliability. A proposed evidence-to-action framework connects information limitations and uncertainty to supporting a decision, obtaining further information, or deferring to expert assessment, illustrated through cardiovascular, ophthalmic, dental, and oncology applications. Evaluation must extend beyond predictive accuracy to calibration, equitable performance, deferral outcomes, workload, diagnostic delays, and patient benefit. The framework requires prospective validation across the complete care pathway. This includes assessing whether missing information disproportionately affects particular patient groups and their access to appropriate further assessment.