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

OBJECTIVES

Interpreting elbow radiographs in pediatric trauma cases is challenging for emergency clinicians due to anatomical peculiarities of this joint in children and the risk of complications from missed fractures. Artificial intelligence (AI), particularly deep learning algorithms, has the potential to assist in fracture detection. We aimed to assess the performance of pediatric emergency clinicians in detecting elbow fractures in children without and with the assistance of a deep learning algorithm.

PATIENTS AND

METHODS

This retrospective study included all children aged 0-15 years admitted to the emergency department of a French university hospital between January 2019 and April 2020, for whom frontal and lateral elbow radiographs were ordered following trauma. The reference standard was established by two independent experts blind to the AI algorithm results. The diagnostic performance of emergency clinicians was evaluated and compared without and with the theoretical AI assistance. Additionally, the performance of the stand-alone AI algorithm was externally tested.

RESULTS

Out of 755 children included (median age: 8 years), 352 (47 %) had an elbow fracture, joint effusion, and/or dislocation. The theoretical AI assistance improved clinician's sensitivity by 21.6 % (from 77.3 % to 98.9 %; p < 0.001), though this was accompanied by a 24.8 % decrease in specificity (from 88.3 % to 63.5 %; p < 0.001). The stand-alone AI algorithm achieved a sensitivity of 98.0 % (95 % CI: 96.0-99.0) and a specificity of 70.0 % (95 % CI: 65.3-74.2).

CONCLUSION

The AI algorithm demonstrated high performance in detecting pediatric elbow fractures, significantly improving sensitivity for emergency clinicians and helping to reduce the rate of missed diagnoses.