Regional fracture-detection false-negative patterns of a multimodal large language model in confirmed pediatric physeal fractures: a case-only analysis

Abstract

Pediatric physeal fractures can be difficult to assess radiographically. We characterized regional fracture-detection failure patterns of a multimodal large language model (MLLM) in confirmed pediatric physeal fractures. This retrospective, single-center, case-only study included 805 confirmed physeal fractures in children aged 0–12 years during 2025. The recorded GPT-5 API workflow was assessed in March-April 2026 using anonymized radiographs plus age, target anatomical region, and a complaint or mechanism. The reference standard was independent review by three orthopedic surgeons - one pediatric orthopedics specialist, one orthopedic trauma specialist, and one senior orthopedic surgeon, each with at least 8 years of clinical experience in orthopedics and traumatology - followed by consensus with model outputs concealed; no radiologist participated. The primary endpoint was a definite or probable fracture call in the correct anatomical region; explicit physeal recognition or correct Salter-Harris classification was not required. Salter-Harris category, anatomical region, and age were prespecified secondary, hypothesis-generating failure-pattern stratifiers. Each case was evaluated once. Because all cases were fracture positive, specificity, predictive values, overall accuracy, and receiver operating characteristic measures were not estimable. The model met the regional fracture-detection threshold in 584 of 805 cases (sensitivity, 72.5%; 95% confidence interval [CI], 69.4–75.5) and did not meet it in 221 (27.5%; 95% CI, 24.5–30.6). In an unadjusted secondary comparison, false-negative rates were 29.9% for Salter-Harris type I-II and 17.9% for type III-IV injuries (risk ratio, 1.67; 95% CI, 1.15–2.41; p  = 0.004). Excluding type I injuries yielded a similar sensitivity of 72.8%. Descriptive false-negative rates were highest in the elbow (36.9%) and distal tibia (31.2%); the distal radius rate was 20.1%. The age-group difference was not statistically significant (30.0% at 0–6 years vs. 24.9% at 7–12 years; p  = 0.108). Under the recorded context-assisted GPT-5 API workflow, 27.5% of confirmed cases did not receive a qualifying regional fracture call. The study did not assess explicit physeal recognition or Salter-Harris classification and cannot establish full diagnostic accuracy, performance relative to clinicians, or clinical utility. Negative model output should not be used to rule out physeal injury.