Medical-AI trial: agreement between ChatGPT-5 and anesthesiologists on preoperative risk class (ASA-PS) — preoperative triage (NCT07459491)

Before surgery, the overall condition of a patient is assessed to plan anesthesia and the operation and to prepare transfusion and testing. The global reference for this is the ASA-PS classification, which grades patients from healthy to those with severe systemic disease. Because the class depends on clinician interpretation, judgments can differ between evaluators (inter-observer variability).

This trial applies a large language model (LLM), ChatGPT-5, to that preoperative assessment and, as an observational study, measures how well the AI agrees with anesthesiologists. Per the registry summary, the same anonymized preoperative clinical information is given to both clinicians and the AI, and the agreement between the ASA-PS classes they assign is the primary endpoint.

Secondarily, it examines how the recommended preoperative labs differ between AI and clinicians, and how the assessments relate to perioperative red-blood-cell (transfusion) use. Participants are adults scheduled for elective surgery, with a target of 703; the study is at the completed stage.

What is measured here is not whether AI is "better" than clinicians, but how closely the AI agrees with them. An LLM can process structured and unstructured medical information together and may help standardize assessment, yet whether its output aligns with clinical judgment needs testing.

Preoperative evaluation is performed in large volumes daily, and the risk class feeds resource decisions such as testing and transfusion preparation; measuring AI-clinician agreement and differences in test recommendations on real data is a meaningful validity check before deploying medical AI (this study evaluates agreement and does not establish AI superiority or safety).