GPT-5 Missed 27.5% of Confirmed Pediatric Growth Plate Fractures

Researchers at Ankara Sincan Training and Research Hospital found that GPT-5 missed 221 of 805 confirmed pediatric growth plate fractures, according to Bioengineer.org. The model recorded a sensitivity of 72.5%, leaving a false negative rate of 27.5%.

The retrospective study examined anonymized X rays from children aged zero to 12 who sustained confirmed fractures during 2025. Researchers evaluated each case through the GPT-5 API in March and April 2026, providing the model with the child's age, anatomical region, and a brief description of the complaint or injury. Three orthopedic surgeons established the reference diagnosis by consensus.

The model missed 29.9% of Salter-Harris type I and II injuries, compared with 17.9% of type III and IV injuries. False negative rates were highest for elbow fractures at 36.9% and distal tibia fractures at 31.2%, while the rate for distal radius fractures was 20.1%.

Because every case involved a confirmed fracture, the study could not measure specificity, predictive values, or overall diagnostic accuracy. It also did not compare GPT-5 with human readers or assess its use in clinical care.