Agreement Between ChatGPT-5 and Anesthesiologists in Preoperative Risk Assessment: ASA Classification

Overview

Accurate preoperative risk stratification is essential for perioperative planning, resource allocation, and patient safety. The American Society of Anesthesiologists Physical Status (ASA-PS) classification remains the most widely used global system for assessing preoperative health status. However, ASA classification relies on clinician judgment and may demonstrate inter-observer variability. Recent advances in artificial intelligence (AI), particularly large language models (LLMs), have shown potential for assisting clinical decision-making by synthesizing structured and unstructured medical information. In perioperative medicine, AI systems may support more standardized risk assessment and laboratory testing strategies. The objective of this observational study is to evaluate the agreement between ASA classifications assigned by anesthesiologists and those generated by a large language model (ChatGPT-5) using anonymized preoperative clinical information. The study will also examine differences in laboratory test recommendations and explore the relationship between clinician- and AI-generated risk assessments and perioperative erythrocyte suspension utilization. Adult patients scheduled for elective surgery who undergo routine preoperative anesthesia assessment will be included. For each patient, the ASA classification assigned by the anesthesiologist will be recorded and compared with the classification generated by the AI system using the same anonymized clinical information. This study aims to assess whether AI-assisted preoperative evaluation may support more consistent risk stratification and potentially contribute to more standardized perioperative resource utilization.

Background and Rationale Preoperative risk assessment is a fundamental component of perioperative medicine and plays a central role in anesthetic planning, patient safety, and perioperative resource allocation. The American Society of Anesthesiologists Physical Status (ASA-PS) classification system remains the most widely used global method for describing preoperative health status. Despite its widespread adoption, ASA classification depends on clinician interpretation and may vary between evaluators. Advances in artificial intelligence (AI), particularly large language models (LLMs), have introduced new opportunities for supporting clinical decision-making. These systems can process both structured and unstructured clinical information and may assist in standardizing certain medical classification tasks. In perioperative medicine, AI-assisted evaluation may help interpret patient comorbidities and clinical information in a consistent manner. Another important component of preoperative assessment is laboratory test utilization. Preoperative laboratory testing is commonly used to identify potential perioperative risks; however, the number and type of tests ordered may vary among clinicians and institutions. AI-based systems may provide standardized recommendations for laboratory investigations and potentially contribute to more efficient resource utilization. In addition, perioperative erythrocyte suspension (packed red blood cell, PRBC) transfusion represents an objective indicator of surgical physiological stress and perioperative resource use. Evaluating the relationship between risk classification and actual blood product utilization may help determine whether AI-assisted risk assessment has potential clinical relevance. Study Design and Procedures This study is designed as a single-center observational study conducted at the preoperative anesthesia outpatient clinic of Antalya City Hospital. Adult patients undergoing routine preoperative anesthesia assessment before elective surgery during the study period will be included in the analysis. For each patient, the ASA Physical Status classification assigned by the evaluating anesthesiologist during routine clinical care will be recorded. An anonymized summary of the same preoperative clinical information will then be analyzed by the artificial intelligence system (ChatGPT-5), which will generate an independent ASA classification. The study will also compare laboratory test recommendations generated by the AI system with those ordered by anesthesiologists during routine preoperative evaluation. The number and types of laboratory tests recommended by each source will be recorded for comparison. Information regarding perioperative erythrocyte suspension transfusion will be obtained from hospital electronic medical records. These data will be used to explore the relationship between risk classification and actual blood product utilization. All clinical information used in the analysis will be anonymized before being processed by the AI system. AI outputs will not influence patient care or clinical decision-making. Study Significance By comparing clinician-based and AI-generated preoperative assessments, this study aims to explore the potential role of large language models in supporting standardized risk stratification and resource utilization in anesthesia practice. The results may contribute to understanding whether AI-assisted evaluation can provide reliable support for preoperative clinical assessment.