Construction and Validation of a Coronary Artery Disease-Specific Large Language Model and Its Evaluation Framework (CorAI)

Coronary artery disease (CAD) affects an estimated 11.39 million people in China, and mortality has continued to rise since 2012. Quality-of-care data indicate persistent gaps in diagnostic and therapeutic guideline adherence, incomplete implementation of secondary prevention, and suboptimal risk-factor control. Large language models (LLMs) can rapidly retrieve and integrate literature, medical records, and multimodal clinical data, and may support clinical decision-making. However, most existing medical LLMs are general-purpose, are trained predominantly on published literature and web text rather than authentic clinical records, and have not been validated with real patient data. No disease-specific LLM for CAD currently exists, and no evaluation framework tailored to CAD has been established; existing benchmarks are general-purpose, lack specialty-specific risk grading, are not rigorously anchored to current guidelines, and emphasize literal accuracy over clinically critical reasoning steps and prescription concordance.

This prospective, multicenter, observational study will enroll approximately 640 patients with confirmed or suspected CAD at Fuwai Hospital and collaborating centers. No study-specific intervention, examination, biospecimen collection, or additional follow-up visit is performed; all clinical decisions are made independently by the treating physician according to routine standards of care. For consenting participants, routinely generated clinical documentation from the index outpatient visit or hospitalization - including medical history, laboratory results, imaging and coronary angiography/coronary CT angiography reports, physician notes, online consultation records, and follow-up records - will be de-identified and used to construct a high-quality multimodal CAD dataset. This dataset will support the development of a CAD-specific large language model (CorAI), the construction of an evaluation framework spanning eight predefined clinical scenarios and multiple assessment dimensions, and external validation of the model at participating centers.

The primary objectives are to characterize the completeness and quality of the constructed prospective cohort dataset, and to quantify the guideline concordance of CorAI-generated clinical recommendations as adjudicated by a blinded panel of cardiologists.