Sensitivity gains under theoretical AI assistance in paediatric elbow fracture detection: a methodological concern...

This letter discusses methodological considerations arising from the study by Da Costa et al. on deep learning-assisted fracture detection in paediatric emergency elbow imaging. While the study provides valuable external validation in a large real-world cohort and demonstrates substantial sensitivity gains using an artificial intelligence safety-net approach, its retrospective design and reliance on a theoretical OR-rule warrant cautious interpretation. The reported improvement in sensitivity reflects both the algorithm's discriminative performance and the lowered positivity threshold inherent to the combination rule, accompanied by a marked reduction in specificity and positive predictive value. Additional concerns include the prominent contribution of isolated joint effusions to the sensitivity gain, the aggregation of diagnostically distinct findings, and the handling of "doubt" outputs as abnormal. We argue that prospective reader studies with real-time AI exposure are required to determine true clinical utility, including downstream effects on immobilization, imaging, specialist consultation, emergency department workflow, and age-specific performance. Such evidence is essential before translating retrospective diagnostic gains into clinical practice.