Transparency of reporting and methodological conduct of prognostic and diagnostic clinical prediction models developed using machine learning in total shoulder arthroplasty: A systematic review and critical appraisal

Abstract

Background: This study aimed to synthesize the applications, performance, and methodological conduct of artificial intelligence (AI) prediction models for total shoulder arthroplasty (TSA) outcomes. Methods: PUBMED, MEDLINE, EMBASE, and CENTRAL were searched on November 2, 2025 for all articles that utilized AI to develop prognostic or diagnostic prediction models utilizing TSA populations. Methodological quality was assessed using the TRIPOD statement and PROBAST tool. Results: Twenty-four studies comprising outcomes of 497,365 patients (35.6% female; 69.6 ± 0.9 years) were included. Of these patients, 31.0% underwent rTSA, 29.1% aTSA, and 2.8% hemiarthroplasty. The remaining patients received a mix of aTSA and rTSA, but the exact proportions were not reported in their respective studies. Nine studies applied AI to clinical outcomes (AUC 0.85, 0.65–0.96), seven to adverse events (AUC 0.73, 0.52–0.92), and six to resource utilization (AUC 0.78, 0.59–0.91). All twelve studies comparing AI to traditional regression reported that AI models demonstrated superior performance. The need and caution for external validation was reported in 15 studies (62.5%). The mean TRIPOD adherence was 11.6 items (range 9–15); 19 studies (82.3%) met >50% of criteria, and 6 (26.1%) met >66%. PROBAST rated 16 studies (66.7%) at high risk of bias. Conclusion: AI prediction models in TSA show poor methodology, especially in calibration, sample size, missing data, and validation, warranting cautious interpretation and clearer direction for future research. Level of evidence: IV, systematic review of level I-IV studies