Artificial intelligence-assisted triage of pediatric supracondylar humerus fractures in emergency departments: A single-centre validation study.

Abstract

Pediatric supracondylar humerus fractures (SCHF) are the most common elbow fractures in children. Accurate classification using the Gartland system is essential for treatment decisions, yet considerable inter-observer variability persists, particularly among non-specialist emergency physicians. Artificial intelligence (AI) has potential to support triage in emergency settings where specialist orthopaedic expertise may be limited. The aim of this study was to develop and validate an AI model capable of assisting in the classification of paediatric SCHF and triaging cases into operative versus non-operative management categories. We retrospectively analysed 1811 annotated radiographs (anteroposterior and lateral views) of paediatric SCHF from a single tertiary Paediatric Orthopaedic centre (2010-2017). Fractures were dichotomised into operative (Gartland IIB, III, flexion type) and non-operative (normal, I, IIA). Two senior Paediatric Orthopaedic surgeons performed independent labelling, with consensus resolution of discordance. A convolutional neural network model was trained with transfer learning and tested on an independent set of 215 radiograph pairs. Test-set composition: 128 non-operative (59.5 %), 87 operative (40.5 %). Model performance was assessed using accuracy, sensitivity, specificity, predictive values, AUROC, and Cohen's kappa. The model achieved overall accuracy of 77 % (95 % CI 71-82 %) and AUROC of 0.84 (95 % CI 0.78-0.89). Specificity (82 %, 95 % CI 74-88 %) exceeded sensitivity (69 %, 95 % CI 58-78 %). Inter-observer agreement between the AI model and surgeon consensus was κ = 0.51 (95 % CI 0.39-0.63), reflecting diagnostic consistency comparable to individual clinician interpretation. Grad-CAM visualisations localised to the supracondylar region in 83 % of predictions. This validation study demonstrates that AI can feasibly assist as a decision-support filter in emergency departments, achieving diagnostic consistency within the range of documented human inter-observer variability. All high-probability predictions require mandatory surgeon review before operative decisions. This proof-of-concept supports further development through multi-centre external validation, expanded governance frameworks, and deployment safety protocols before implementation in diverse emergency settings. II.