Small lesion–high risk: diagnostic performance of artificial intelligence in paediatric fractures with medicolegal impact

Background Certain paediatric fractures carry a high medicolegal risk if overlooked. While artificial intelligence (AI)-based diagnostic tools have demonstrated promising results in general fracture detection, their accuracy in identifying rare but high-risk paediatric injuries remains insufficiently studied. Objective To assess the diagnostic performance of a commercial AI-based software (SmartUrgence, Milvue, Paris, France) in detecting a predefined set of medicolegally relevant paediatric fractures. Materials and methods This retrospective, single-centre study analysed radiographs from 125 paediatric patients (ages 2 years to ≤17 years) with one of the following fracture types: lateral humeral condyle fractures, Monteggia fractures, trampoline fractures of the proximal tibia, or medial malleolar fractures. AI-generated results were compared against a reference standard established by two board-certified paediatric radiologists, using clinical, imaging, and intraoperative data. Sensitivity and specificity were calculated for each fracture type. Results Sensitivity was highest for trampoline fractures of the proximal tibia (100%) and medial malleolar fractures (78%), with specificity at the knee and ankle approaching 100%. For lateral humeral condyle fractures, sensitivity reached 73%, while specificity remained high at 90%. The most significant limitation concerned Monteggia fractures: although ulnar fractures were detected with 81% sensitivity, only 2% of associated radial head dislocations were identified correctly. Conclusion While the AI-based software demonstrated overall strong performance, its emphasis on specificity limits its utility in detecting high-risk fractures for which sensitivity is paramount. Future development should focus on enhancing sensitivity, particularly for the detection of elbow dislocations.