A NEWLY published retrospective study has shown that AI, particularly deep-learning algorithms, can significantly reduce the rate of misdiagnosis in paediatric elbow fractures. The study analysed 755 children (median age: 8 years), of which 352 (47%) had an elbow fracture, joint effusion, and/or dislocation. The AI assistance improved physicians’ sensitivity (correctly identifying fractures) by 21.6% (from 77.3–98.9%; p<0.001), although with the AI, physicians’ specificity (correctly identifying normal/unfractured elbows) fell by 24.8% (from 88.3% to 63.5%; p<0.001). Importantly for the technological approach, however; the stand-alone AI algorithm (without physician review) achieved a sensitivity of 98% and specificity of 70%.
Challenges in Assessing Paediatric Elbow Fractures
Elbow fractures are common paediatric injuries, accounting for 15–20% of all fractures in children. However, interpreting paediatric elbow radiographs is notoriously complex due to the evolving anatomy of growing children, where multiple centres of ossification can both obscure and mimic fractures. Missed diagnoses can lead to future complications, especially as children’s bones have limited capacity for remodelling.
What Deep-Learning AI Can and Can’t Deliver
Emergency clinicians play a key role in interpreting radiographs, but they frequently face difficulties due to a shortage of emergency radiologists, high patient volumes, all of which elevate the risk of diagnostic errors. Indeed, the number of trauma consultations and radiographs performed can make real-time specialist review unfeasible.
As such, the authors of the study propose that a deep-learning AI algorithm, integrated into clinical practice, would significantly enhance emergency clinicians correct diagnosis of paediatric elbow fractures, especially in identifying and prioritising those patients that require further specialist radiologist review.
But, the authors also note the current limitations to this technology, which has to date been trained towards selectivity and identifying true positive fractures. As such, while the deep-learning AI algorithm identified 95% of the fractures missed by the unassisted clinician (from n=80 down to n=4), there was a significant increase in false positive diagnoses (from n=47 up to n=147).
Going forward the authors propose that the AI performance would be improved through further training of the algorithms. However, this should be accompanied with studies to assess the impact of AI on patient outcomes, to ensure AI systems can perform as well as trained practitioners, and are effectively integrated into real-world settings.
Reference
Costa JD et al. Assessing deep learning artificial intelligence support for detecting elbow fractures in the pediatric emergency department. Eur J Radiol. 2025;DOI: 10.1016/j.ejrad.2025.112498.
Author: Adam Michael, Director and Founder, 80th Atom, Greater Cambridge Area, UK
