Comprehensive Summary
This study evaluated whether BoneView, a deep learning AI tool, could help emergency clinicians better detect elbow fractures and related injuries in children. Investigators included 755 consecutive children (of 0-15 years) seen for elbow trauma at a French pediatric ED between January 2019 and April 2020, all having both frontal and lateral elbow radiographs. A blinded reference standard (radiologist and pediatric trauma emergency physician, with an orthopedic surgeon resolving disagreements) classified each exam as normal or abnormal. 352/755 (46.6%) of the exams were abnormal. The AI system, an FDA-cleared and CE-marked fracture-detection tool, analyzed radiographs through the ED’s existing imaging software and labelled them as negative, positive, or doubt based on the system’s confidence thresholds (‘doubt’ was treated as abnormal to maximize sensitivity). Unassisted emergency clinicians correctly identified 272/352 abnormal exams (sensitivity 77.3%) and 356/403 normal exams (specificity 88.3%), with an overall accuracy of 83.2%. In a simulated clinician with AI scenario, where an exam was considered abnormal if either flagged it, sensitivity rose to 348/352 (98.9%). However, specificity fell to 256/403 (63.5%), accuracy 80.0%, and NPV 256/260 (98.5%). Standalone AI achieved a sensitivity of 345/352 (98.0%), specificity 282/403 (70.0%), PPV 345/466 (74.0%), NPV 282/289 (97.6%), and accuracy 627/755 (83.0%). Of the 80 abnormal exams initially missed by clinicians (54 isolated effusions and 26 fractures/dislocations), the AI correctly reclassified 76/80 (95%) of them as abnormal. Conversely, the AI correctly identified 26 clinician false positives as normal, potentially avoiding unnecessary immobilization. Overall, 13.5% of children (102/755) might have had their management changed with AI support, through either preventing missed injuries or reducing unneeded casting and slings.
Outcomes and Implications
Interpreting pediatric elbow radiographs is challenging due to complex, age-dependent ossification centers and the need to recognize indirect signs such as joint effusion. Missed injuries can lead to delayed treatment and prolonged symptoms. The findings showed that BoneView functioned effectively as a high-sensitivity “second reader,” approaching near-perfect sensitivity and substantially reducing missed fractures and effusions compared to emergency clinicians alone. However, this improvement came at the cost of reduced specificity and more false positives. Performance also varied by age, with all AI false negatives occurring in children under five and involving either effusions or fractures in that younger age group. In practice, integrating tools like BoneView as an adjunct in pediatric emergency workflows could help standardize image interpretation, prioritize radiologist review, and reduce the rate of missed elbow injuries while still relying on clinician judgment to balance false positives and avoid overtreatment.