EP8.38 Traditional Statistics and Machine Learning for Prediction of Post-Operative Outcomes in Hip Preservation: A Systematic Review and Narrative Synthesis

Abstract Purpose Prediction of post-operative outcomes after hip preservation is important for treatment decision-making. Machine Learning (ML) analyses may capture complex, non-linear relationships between demographic, imaging, and symptom factors. Statistical and ML models have been used to identify pre- and intra-operative predictors for prognosis. This review aims to 1) identify predictors consistently associated with post-operative outcomes, 2) summarize ML model performance across studies. Methods A systematic search of PubMed, Embase, and Web of Science was performed on articles through February 2026. Studies evaluating pre- or intra-operative predictors of post-operative outcomes following arthroscopy and periacetabular osteotomy (PAO) procedures were included. Exclusion criteria included revision populations, skeletally immature patients, no ≥1-year post-operative outcomes, or no pre-/intra-operative predictors. Of 1126 studies, 625 remained after duplicates, and 109 after full-text review. Data extracted included study design, procedure type, outcomes, and predictors for narrative synthesis. Studies were classified as ML only if they included a true predictive framework (train–test split, cross-validation, etc.). Studies using ML algorithms for exploratory analysis without proper prediction methology were categorized as statistical approaches. For ML studies, the highest reported area under the curve (AUC) for the best-performing model was extracted for comparison. Results Of 109 studies included, 100 (92%) used traditional statistical methods and only 9 (8%) used ML models. Most studies investigated arthroscopy (76%), followed by PAO (21%) and combined populations (3%). The most common outcomes were patient-reported outcome measures (PROMs) (61%), total hip arthroplasty (THA) conversion rate (9%), revision rate (7%), nonunion complications (2%), and return to sport (2%). The most common predictors were age (13%), pre-operative PROMs (12%), BMI (7%), sex (6%), mental health (5%), cartilage damage (5%), and symptom duration (5%). All ML studies evaluated arthroscopy populations and primarily focused on predicting PROMs, with limitations in consistency, external validity, and variable model performance (AUC 0.67-0.99). Conclusion Most ≥1-year outcome prediction studies use statistical methods, while ML performance ranges from moderate to excellent. Across both approaches, consistent predictors included patient demographics, baseline PROMs, mental health, and cartilage damage. Few studies applied true ML frameworks, highlighting the need for more consistent methodology to build advanced ML models in the future.