Assessing deep learning artificial intelligence support for detecting elbow fractures in the pediatric emergency department

Comprehensive Summary

This study by Da Costa et al. assesses the diagnostic performance of emergency physicians with and without the assistance of AI for detecting elbow fractures in pediatric patients. Frontal and lateral radiographs from 755 children, ranging in age from 0 to 15 years, were collected, along with demographic data and clinical details regarding diagnosis or radiologic lesions, when applicable. A reference standard was generated by a group comprising an experienced radiologist, a pediatric emergency physician specializing in traumatology, and a pediatric orthopedic surgeon to resolve disagreements. The deep learning (DL) algorithm BoneView™ was used for AI-based assessment, evaluated both as a theoretical assistive tool and as a standalone diagnostic tool. In the final assessment alongside the emergency clinician, the AI model improved the true positive diagnoses 22%, from from 77.3% to 98.9%. Overall, 95% of clinician missed diagnoses were corrected by the AI algorithm. Despite the improvement in missed diagnoses, the algorithm did result in a significant increase in false positives by almost 25% compared to the unassisted clinician. The standalone assessment achieved a similar performance to the AI-assisted assessment, with a similarly high rate of false positives compared to the unassisted emergency clinician. When AI and the clinician agreed, the error rate was below 5%, pointing to the capability of AI models to be integrated as “Second readers” to reinforce diagnostic confidence for non-specialized physicians who require an interim reader before professional consultation with a radiologist. In conclusion, the evaluated AI algorithm demonstrated robust predictive accuracy and shows potential to support emergency clinicians by enhancing clinical decision-making, with the potential to improve outcomes for pediatric patients with elbow fractures.

Outcomes and Implications

Elbow fractures account for 15–20% of all fractures in children and are particularly concerning due to their complex radiographic appearance in pediatric patients and the critical importance of accurate diagnosis, given the potential for long-term complications. The majority of emergency clinician misdiagnoses are fractures, due in part to several factors, including non-expert interpretation of radiographic data. This study’s examination of unassisted versus AI-assisted emergency clinicians suggests the potential for AI to act as a supplementary tool to better support physicians in the management of pediatric elbow fractures. Acting either as a confirming “second reader” or as an interim diagnostic tool before further examination by a specialized radiology specialist, this AI model can streamline and improve the proper diagnosis of pediatric elbow fractures. Further development of the model used in this paper would involve integration of clinical data alongside purely radiographic data to better fine-tune the diagnosis of trauma cases.