Abstract
BACKGROUND: Hand surgeons are increasingly asked to assign American Society of Anesthesiologists (ASA) physical status classifications for local anesthesia-only procedures despite minimal training. Artificial intelligence (AI) is a powerful tool that could assist in ASA class assignment. We evaluated the concordance/agreement between surgeon-rated and AI-rated ASA class. METHODS: Local anesthesia-only hand surgical episodes (n = 236) performed by 4 fellowship-trained hand surgeons at ambulatory surgery centers were included. We used ChatGPT to estimate ASA class-based patient medical history. Interrater agreement between surgeon and AI ratings were compared with the weighted Cohen κ. The most recent anesthesia-rated ASA class from another anesthesia event was also recorded (if present) and compared. We used multivariable logistic regression models to evaluate whether patient factors were associated with concordance. We also recorded 30-day postoperative emergency department visits and hospitalizations. RESULTS: Overall concordance between surgeon and AI ratings was 56%. In 72% of discordant cases, surgeon-rated ASA class was lower than the AI rating. Surgeon-rated ASA class exhibited fair agreement with the AI rating (κ = 0.38). Artificial intelligence-rated ASA class exhibited moderate agreement with prior anesthesia-rated ASA class (κ = 0.44). In the multivariable model, age, sex, and presence of a prior anesthesia-rated ASA class were not associated with concordance between surgeon and AI ratings. Postoperative emergency department visits or hospitalization occurred in 7 (3%) patients, for which surgeon-rated ASA class was lower than AI rating in 4 cases. CONCLUSIONS: Generative AI could be used as a tool to support ASA class assignment for local anesthesia-only hand surgeries.