Current Applications of Artificial Intelligence in Orthopaedic Trauma: A Narrative Review
Abstract
Background: Deep learning, and in particular the convolutional neural network (CNN), has moved from proof of concept to commercial deployment in orthopaedic trauma within a single decade. Practising surgeons increasingly encounter these tools without a clear account of what the evidence does and does not support. Objective: To synthesise current evidence on artificial intelligence across the orthopaedic trauma pathway, covering fracture detection, classification, outcome prediction, operative support and generative language applications; to appraise the methodological quality of that evidence; and to define what is required for safe adoption in India and other low- and middle-income countries. Methods: PubMed (MEDLINE), Embase and Scopus were searched for English-language, peer-reviewed studies published between January 2015 and April 2026. Titles, abstracts and full texts were screened against pre-specified inclusion and exclusion criteria, and findings were synthesised narratively. Methodological quality was appraised against the domains of QUADAS-2 for diagnostic-accuracy studies and PROBAST for prediction-model studies. Key findings: Pooled sensitivity and specificity for fracture detection on plain radiographs clustered around 0.87 to 0.91, comparable with specialist clinicians, and clinician sensitivity rose to 0.97 when the algorithm was used as an adjunct. Site-specific networks matched or exceeded expert performance at the hip, distal radius, scaphoid, proximal humerus, ankle, ribs and spine. Classification remained weaker (accuracy 60% to 81% for the distal radius), and outcome-prediction models gained little over conventional regression (area under the curve 0.80 versus 0.79 for 30-day mortality after hip fracture). Generative language models showed early utility in documentation and patient education but did not reach resident-level performance and produced erroneous content at clinically relevant rates. Risk of bias was high in about half of the diagnostic studies, and only about one model-development study in nine reported external validation. Conclusion: Artificial intelligence has achieved specialist-level accuracy for fracture detection but has not yet demonstrated patient-level benefit. The priorities for India are representative multicentric datasets, external validation, CLAIM- and TRIPOD+AI-compliant reporting, and prospective clinical-impact trials rather than further internal-validation accuracy studies. Keywords: Artificial intelligence; Deep learning; Convolutional neural network; Orthopaedic trauma; Fracture detection; Large language models.