Predicting major adverse cardiovascular events from incomplete clinical data through interpretable multimodal AI system

Accurate prediction of major adverse cardiovascular events (MACEs) is crucial for devising personalized treatments for patients with reperfused myocardial infarction. This requires integrating cardiovascular magnetic resonance (CMR) and electronic health record (EHR) data to enable a comprehensive risk assessment. However, both often suffer from missing modality issues—CMR may lack certain imaging sequences like T2 mapping due to equipment limitations or protocol variations, and EHR data may omit variables that are not measured or recorded. Existing methods address neither these gaps nor the transparent, interpretable reasoning required for trustworthy clinical AI. Here, we propose MACE–MAIS, an end-to-end multimodal AI system with integrated reasoning to predict MACE from incomplete CMR and EHR data. It uses a missing-modality-aware contrastive image pretraining to robustly extract features from incomplete CMR data, and a large language model to embed unstructured EHR texts despite missing entries. Furthermore, MACE–MAIS couples each risk prediction with an interpretable rationale, offering clinicians actionable insights and serving as a reliable reference for routine practice. Evaluated across four real-world clinical datasets, MACE-MAIS outperforms baseline methods in MACE risk prediction while providing transparent, clinically relevant reasoning. This system provides a practical and reliable solution for interpretable clinical AI decision support.