Bounded Agentic Orchestration for Multimodal Cardiovascular Risk and Follow-up

We present a bounded and auditable agentic framework for multimodal cardiovascular risk assessment and follow-up support using chest radiographs (CXRs) and electronic health records (EHRs). Rather than introducing a new foundation model, our primary contribution is the coordination of existing multimodal models within a constrained workflow that selects a high-quality CXR and temporally relevant EHR evidence, invokes specialist review when needed, and applies deterministic verification and safety controls. The system fails closed when evidence is missing, inconsistent, unsafe, or outside scope, while recording evidence identifiers, model outputs, stop reasons, and guard decisions for reproducibility and review. We evaluated the workflow retrospectively in an emergency-department cohort of 100 patients. The study-defined composite MACE endpoint included stroke, heart failure, myocardial infarction, or death after the index CXR. MACE classification achieved sensitivities of 72.97\%, 85.45\%, and 83.67\% at 1 week, 1 month, and 1 year. The downstream follow-up recommendation achieved a sensitivity of 95.6\% at 1 year. These results suggest that the framework can transform routinely collected ED data into evidence-grounded and reviewable cardiovascular follow-up recommendations, while clinical utility and prospective outcome improvement remain to be established.