From Callus to Code: A systematic review of early evidence on artificial intelligence in fracture non-union.

BACKGROUND: Failure of fracture healing continues to pose a major challenge in orthopaedic practice, occurring in approximately 5-10% of long bone injuries and contributing to prolonged disability and increased healthcare utilisation. Traditional prediction methods rely on clinical and radiographic assessment but are limited by variability and inability to capture complex, non-linear interactions between risk factors. Advances in artificial intelligence (AI), particularly machine learning and computational modeling, offer new opportunities for improving the prediction and management of fracture healing outcomes. METHODS: This systematic review was performed in accordance with PRISMA 2020 recommendations and registered in the PROSPERO database (CRD420261371644). A structured search of PubMed, Embase, Scopus, and Web of Science was conducted up to December 2025. Studies evaluating AI-based models for the prediction, diagnosis, or management of fracture non-union were included. Data extraction focused on study characteristics, AI model type, input variables, and performance metrics such as area under the curve (AUC), sensitivity, and accuracy. Risk of bias was assessed using the PROBAST tool. RESULTS: Three studies involving 785 patients satisfied the inclusion criteria. Machine learning algorithms, including Random Forest, XGBoost, and CatBoost demonstrated strong predictive ability, with AUC values between 0.83 and 0.86. A biomechanical simulation model accurately classified 23 healed fractures and identified 4 of 6 non-union cases. Considerable variability in study design, absence of external validation, and predominantly retrospective methodologies limited comparability. CONCLUSION: Current evidence regarding AI for fracture non-union prediction remains limited and heterogeneous, with findings derived from a small number of predominantly retrospective studies. Although machine learning approaches have demonstrated encouraging predictive performance, substantial methodological limitations and lack of external validation preclude routine clinical implementation. Further large-scale prospective studies with standardized methodologies and robust validation are required.