Applications, Safety Challenges and Future Directions of AI Agents in Cardiovascular Prediction: A Review of Technological Evolution, System Architecture and Clinical Translation

Abstract

Cardiovascular disease remains a major contributor to mortality and long-term morbidity worldwide. For this reason, risk prediction, early detection, and decision support have become central tasks in digital health research. Cardiovascular prediction has developed from statistical scores to machine learning and deep learning models using clinical data, electrocardiograms, imaging, electronic health records, and multimodal inputs. AI agents add a system layer to this trajectory: they connect large language models, retrieval-augmented generation, tool use, code execution, multi-agent collaboration, and evidence tracing. Taking AgentMD, HeartAgent, and CardioAgent as representative systems, this review examines the architecture, clinical roles, safety risks, and future development of AI agents in cardiovascular prediction. The central argument is that AI agents are valuable not merely because they may improve prediction metrics, but because they can link risk calculators, deep learning models, medical knowledge bases, and clinical workflows into prediction systems that are explainable and auditable. At the same time, their deployment introduces prompt injection, hallucination, tool misuse, privacy leakage, adversarial attacks, multi-agent error propagation, and accountability risks. Future research should give more attention to lightweight reasoning, privacy-preserving implementation, prospective validation, and human-agent evaluation.