Designing and evaluating large language model-enabled clinical decision support for heart failure: a modular and risk-tiered framework

Heart failure (HF) care requires repeated decisions across suspected disease, diagnostic confirmation, phenotyping, guideline-directed medical therapy, device consideration, worsening HF, transition care, and advanced HF planning. Large language models (LLMs) may support this work by synthesizing structured and unstructured electronic health record data, retrieving current evidence, and presenting patient-specific reasoning. However, an HF-specific LLM clinical decision support system should not be framed as a single autonomous agent. We present the Heart Failure Intelligent Agent (HF-IA) as a modular, risk-tiered conceptual framework in which agent functions have different data requirements, reference standards, risk levels, and validation pathways. We argue that evaluation should combine node-level tests, longitudinal case replay, silent prospective validation, and post-deployment monitoring. This framework is conceptual and does not claim clinical effectiveness; its value is to clarify design, evaluation, and governance requirements for future LLM-enabled HF decision support.