TOPLINE:

A retrospective cohort study found that artificial intelligence (AI) assistance significantly improved the sensitivity of emergency clinicians in detecting paediatric elbow fractures, correctly identified 95% of false negatives as abnormal, and showed a standalone sensitivity of 98%.

METHODOLOGY:

• This single-centre study evaluated the impact of an AI-based deep learning algorithm in assisting emergency clinicians to detect elbow fractures in 755 children (median age, 8.3 years; 380 girls) who underwent radiography of the elbow in both frontal and lateral views at the University Hospital of Nantes, France, between January 2019 and April 2020.
• The reference standard included an independent review of radiographs by two experts: one radiologist and one paediatric emergency physician. Diagnostic performance assessment involved the comparison of emergency clinician interpretations without and with AI assistance (BoneView, version 2.0.3.1), including the performance of the standalone AI algorithm.
• Cohen's kappa coefficient was calculated to assess the agreement between the experts constituting the reference standard. Diagnostic performance was based on the calculation of sensitivity, specificity, negative predictive value (NPV) of the clinicians, and AI.
• Researchers conducted a subgroup analysis on the basis of age groups (0-4 years vs 5-15 years) and performed a sensitivity analysis excluding children with isolated joint effusions.

TAKEAWAY:

• The reference standard identified 46.6% of cases as abnormal, of which 27% had isolated joint effusions and 73% had fractures and/or dislocations. Cohen's kappa coefficient for the agreement between the experts was 0.77.
• The sensitivity of emergency clinicians improved by 21.6% with the AI algorithm (from 77.3% to 98.9%; P < .001), whereas specificity decreased by 24.8% (from 88.3% to 63.5%; P < .001); the AI algorithm correctly identified 95% of false-negative cases as abnormal.
• The standalone AI algorithm achieved a sensitivity of 98.0% and an NPV of 97.6%, with a notably higher sensitivity seen in children aged 5-15 years vs 0-4 years (100.0% vs 91.9%; P < .001).
• After excluding children with isolated joint effusions, the sensitivity improved by 9.7% with the AI algorithm (from 89.9% to 99.6%; P < .001). The sensitivity and NPV of the standalone AI algorithm changed to 99.2% and 99.3%, respectively.

IN PRACTICE:

"The AI algorithm evaluated in this study demonstrated robust performance in detecting elbow fractures in children, particularly in terms of Se [sensitivity] and NPV," the authors wrote, further adding that "its integration into clinical practice has the potential to significantly help emergency clinicians, especially by aiding in the prioritization of radiological reviews by specialists."

SOURCE:

This study was led by Julie Da Costa, MD, Pediatric Emergency Department, CHU Nantes, Nantes, France. It was published online on November 01, 2025, in the European Journal of Radiology.

LIMITATIONS:

Experts' fracture diagnoses depending on both radiologic and clinical information were lacking, which may have introduced classification bias. The evaluation was limited to elbow fractures, and the findings may not be generalisable to other regions of the paediatric skeleton. Furthermore, as emergency clinicians did not have access to the results of the AI algorithm during actual patient management, the theoretical framework may have influenced performance in the AI-assisted clinician group.

DISCLOSURES:

This study did not receive any external funding. The authors declared having no relevant conflicts of interest.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.