The prognostic ability of machine learning models for the treatment of femoroacetabular impingement: a systematic review and narrative synthesis

Objective: To assess the ability of machine learning (ML) models to predict outcomes following arthroscopy for femoroacetabular impingement (FAI). Design: Systematic review with narrative synthesis of clinical outcome studies, following the PRIMSA guidelines. Literature Search: EMBASE, MEDLINE, and PUBMED databases were searched. Study Selection Criteria: Included studies examined FAI cases undergoing arthroscopy, using ML models to predict surgical outcomes. Reviews, editorials, conference abstracts, and studies with patients under 18 were excluded. Data Synthesis: Risk of bias was assessed using the PROBAST tool. The certainty of evidence was assessed using Grading of Recommendations, Assessment, Development and Evaluations (GRADE). Due to methodological heterogeneity, data were synthesised narratively according to the SWiM guidelines. The area under the curve (AUC) was selected as the common metric extracted from each study. Results: Ten studies comprising of 19 146 arthroscopy cases were included. Seven used ML to predict minimal clinically important difference (MCID) achievement, two to predict further surgery and one used ML to predict loss to follow up. Eight studies had high risk of bias and two had unclear bias. AUC values for MCID prediction ranged from 0.42 to 0.90, 0.54–0.808 for further surgery, and 0.76 (.701–.848) for lost follow-up. The certainty of evidence was downgraded to ‘very low’ due to risk of bias, inconsistency and imprecision. Conclusion: ML models predicting outcomes after hip arthroscopy for FAI have high bias, low certainty of evidence and lack external validation. The AUC values, which assess the clinical utility of these models, varied significantly, sometimes nearing random chance. In addition, the lack of reported calibration metrics further limits interpretation of clinical reliability. Current evidence does not support the reliable prognostic ability of ML models in hip arthroscopy outcomes.