ChatGPT delivers mostly satisfactory but occasionally inaccurate and potentially unsafe answers to hip arthroscopy questions during the surgeon's learning curve.
other · Level V
Where this comes from
- Record sourced from PubMed, PMID 42255385.
- Also identified by DOI 10.1002/jeo2.70770 and PMC identifier 13239738.
- Licence recorded as CC BY.
- The licence permits redistribution, so the abstract is shown in full and the full text is available from the publisher.
Abstract
To assess the accuracy, potential safety concerns and readability of single-shot answers generated by the free GPT-4o ChatGPT interface to 15 predefined surgeon-level hip arthroscopy (HAS) learning-curve questions, using expert ratings (Mika scale) and interrater reliability analysis. Fifteen questions were selected based on frequency in HAS teaching courses. Each question was submitted once to ChatGPT in a new chat without additional prompting. Eight high-volume hip arthroscopists, serving as faculty and trainers, independently rated every answer using the 4-point Mika scale (1 = excellent, 4 = unsatisfactory). Consensus ratings were defined by the modal score or, in case of ties, by panel discussion with safety-oriented adjudication. Interrater reliability was evaluated using intraclass correlation coefficients (ICCs). Readability metrics were assessed using the Flesch Reading Ease Score (FRES) and the Flesch-Kincaid Grade Level (FKGL). After consensus, 2 of 15 responses (13.3%) were rated excellent, 9 (60%) satisfactory with minimal clarification required, 3 (20%) satisfactory with moderate clarification required and 1 (6.7%) unsatisfactory, yielding a mean accuracy score of 2.2 ± 0.8 (median, 2.0; range, 1-4). The single unsatisfactory answer addressed patient positioning, and pharmacologic venous thromboembolism prophylaxis was rated satisfactory but raised safety concerns. Interrater reliability was moderate for single ratings (ICC(2, 1) = 0.58) and excellent for the mean of all raters (ICC(2, 8) = 0.92). Readability indicated a college-level demand (mean FRES 34, mean FKGL 13). GPT-4o provided mostly satisfactory and useful answers to common HAS questions posed by surgeons in their learning curve, but a minority of responses required substantial clarification or were judged unsafe if applied uncritically. These findings support the use of large language models as an adjunct educational tool, while highlighting the need for expert verification in safety-critical topics. Level IV, cross-sectional, comparative simulation study.