Machine learning and artificial intelligence for clinical prediction in orthopaedic surgery: a bibliometric analysis of 1,739 publications, 2015–2026

Background. Machine learning (ML) and artificial intelligence (AI) are now routinely applied to clinical prediction in orthopaedic surgery, but the structure of this literature — who produces it, where it is published, and which intellectual traditions it draws on — has not been mapped. Methods. We searched the Web of Science Core Collection (SCI-EXPANDED, SSCI, ESCI) on 4 August 2026 for English-language articles published 2015–2026 combining three concept blocks: orthopaedic domain, AI/ML, and prediction or prognostic modelling. Full records with cited references were exported and analysed with Python 3.11. We quantified annual production, journal and source concentration (Bradford's law), country and institutional collaboration, author productivity (Lotka's law), citation impact, document co-citation with Louvain community detection, and keyword co-occurrence with Callon centrality–density thematic mapping. Reporting follows BIBLIO [1]. Results. The corpus comprised 1,739 articles. Annual output rose from 14 in 2015 to 438 in 2025, a compound annual growth rate of 41.1%. The corpus attracted 28,509 citations (mean 16.39 per document; h-index 66), with 132 documents (7.6%) uncited. Contributions came from 72 countries; 26.4% of articles were internationally co-authored. China led corresponding-author output (513 articles) but showed low international collaboration (11.1% multi-country publications) and low network betweenness (0.027), whereas the USA (466 articles, 23.0%, betweenness 0.203) and the United Kingdom (55 articles, 54.5%) occupied the brokering positions. Bradford's core zone comprised 22 journals. Of 10,266 authors, 86.3% contributed a single article. The intellectual base was predominantly methodological rather than clinical: the ten most co-cited references were dominated by general machine-learning tooling and architectures, while the most co-cited journals were clinical orthopaedic titles. Thematic mapping identified 11 themes; machine learning, deep learning and prediction modelling occupied the basic/transversal quadrant, biomechanics was a motor theme, and spine and risk prediction remained niche. Conclusions. Orthopaedic prediction research is expanding rapidly but remains methodologically imported and structurally fragmented: it borrows its analytical apparatus from computer science while publishing its evidence in clinical journals, and it is dominated by authors who contribute once and do not return. Prediction-model reporting standards appear in the co-citation core but are not yet central.