Clinical concordance of large language model–generated and clinician decisions in perioperative planning for dental procedures under general anesthesia

Large language models are increasingly explored for perioperative decision support; however, their concordance with clinician-generated anesthetic decisions remains uncertain. This study evaluated the agreement of ChatGPT-5.2 with an independent anesthesiologist and a clinical comparator across multiple anesthetic decision domains in dental procedures performed under general anesthesia. In this retrospective concordance study, 228 consecutive real-world cases were analyzed. Standardized, anonymized preoperative data were independently assessed by ChatGPT-5.2 and an anesthesiologist blinded to the clinical comparator’s assessments. The responsible anesthesiologist served as the clinical comparator. Decision domains included ASA classification, anesthetic technique (GA vs. sedation), postoperative disposition (same-day discharge vs. >1-day hospitalization), and induction doses. Descriptively, the anesthesiologist demonstrated higher concordance with the clinical comparator than ChatGPT-5.2 for ASA classification and showed greater concordance with the actual clinical disposition. Although agreement for anesthetic technique was high overall, only four cases (1.8%) were classified as sedation by the clinical comparator, and ChatGPT-5.2 identified none of these as sedation. Similarly, only five patients (2.2%) were actually admitted postoperatively, whereas the independent anesthesiologist and ChatGPT-5.2 predicted admission in 4.8% and 72.4% of cases, respectively. Given the extreme class imbalance and the very small numbers of sedation ( n = 4) and admission ( n = 5) cases, performance measures for these categories should be interpreted cautiously and considered exploratory. In induction dosing, the anesthesiologist showed minimal bias relative to the clinical comparator. ChatGPT-5.2 demonstrated systematic bias for propofol (+ 0.286 mg/kg) and fentanyl (+ 0.503 µg/kg), whereas rocuronium dosing showed no significant bias. ChatGPT-5.2 consistently reproduced structured medical knowledge but showed variable performance in context-dependent anesthetic decisions. These findings suggest that AI may support standardization and risk identification, while clinician judgment remains essential for individualized perioperative decision-making.