Large language models in sports cardiology: Challenges, governance, and human-in-the-loop integration.
TL;DR
It is argued that current evaluation paradigms inadequately capture the uncertainty-intensive nature of athlete cardiovascular assessment and proposed the text-proxy paradigm, a governance-oriented framework in which LLMs operate as semantic integration layers that process clinician-verified and structured outputs generated by validated upstream systems rather than directly interpreting physiological signals.
Abstract
Background
Large language models (LLMs) are increasingly integrated into clinical artificial intelligence (AI) workflows, with emerging applications in medical documentation, information synthesis, guideline retrieval, and patient communication. Recent advances in foundation models, including multimodal systems, have further expanded these capabilities. However, current evidence derives predominantly from examination-style benchmarks, retrospective vignettes, and hospital-based populations, limiting its direct applicability to athlete-centered cardiovascular care. Sports cardiology represents a uniquely challenging environment for AI because of low disease prevalence, physiological-pathological overlap, longitudinal uncertainty, and eligibility-driven decision-making. MAIN BODY This narrative review critically examines the conceptual, technical, regulatory, and governance challenges associated with LLM deployment in sports cardiology. We argue that current evaluation paradigms inadequately capture the uncertainty-intensive nature of athlete cardiovascular assessment and propose the text-proxy paradigm, a governance-oriented framework in which LLMs operate as semantic integration layers that process clinician-verified and structured outputs generated by validated upstream systems rather than directly interpreting physiological signals. Within this architecture, LLMs may support documentation, longitudinal synthesis, guideline retrieval, contextual reasoning, and athlete communication under continuous human oversight. We further discuss uncertainty propagation across layered AI systems, workflow-oriented validation, regulatory implications, and a staged framework for responsible clinical implementation.
Conclusion
Future progress will depend less on increasingly capable models than on trustworthy clinical ecosystems that prioritize governance, calibration, transparency, auditability, and clinician oversight. Rather than serving as an early adopter of autonomous AI, sports cardiology may provide a benchmark for the responsible integration of foundation models into high-stakes clinical medicine.