Diagnostic performance of GPT-5.2 ınstant for pediatric elbow fracture detection on radiographs: a prospective single-center diagnostic accuracy study

Abstract

Multimodal large language models can interpret medical images, but their performance for pediatric elbow radiographs remains uncertain. We evaluated the diagnostic performance of GPT-5.2 Instant as accessed through the ChatGPT web interface during the defined study period. In this prospective, single-center diagnostic accuracy study, consecutive patients aged 0–17 years presenting with acute elbow trauma between December 2025 and May 2026 were included. De-identified anteroposterior and lateral radiographs exported from the institutional PACS as PDF files were evaluated by GPT-5.2 Instant through the standard ChatGPT web interface, once per case, using a fixed prompt and no task-specific fine-tuning. The final adjudicated composite reference standard was based on concordant assessments by two emergency physician readers; discordant cases underwent independent radiologist review and adjudication. Diagnostic performance was summarized with 95% confidence intervals. Among 412 patients, the reference-standard fracture prevalence was 10.0% (41/412). GPT-5.2 Instant classified 74 cases (18.0%) as fracture-positive. It correctly identified 23 of 41 fractures and missed 18. Sensitivity was 56.10% (95% CI, 39.75–71.53), specificity 86.25% (82.32–89.59), positive likelihood ratio 4.08 (2.81–5.92), negative likelihood ratio 0.51 (0.36–0.72), positive predictive value 31.08% (21.61–42.46), negative predictive value 94.67% (91.71–96.62), and overall accuracy 83.25% (79.29–86.73). Agreement with the reference standard was limited (κ = 0.312). Under the reported access and input conditions, GPT-5.2 Instant demonstrated limited sensitivity and agreement for pediatric elbow fracture detection, insufficient to support independent diagnostic or rule-out use. Clinician–AI collaboration was not evaluated and requires prospective investigation.