Machine Learning Using Preoperative Patient Factors Can Predict the Severity of Pain Following Primary Total Hip...

Persistent postoperative pain is a challenge in total hip arthroplasty (THA), yet a lack of accurate predictive tools exists for this outcome. This study aimed to develop and validate machine learning (ML) models to predict long-term pain following THA and to identify key predictors.

A secondary analysis of 513 patients who underwent primary unilateral THA from 2000 to 2011 was performed. Hip pain was measured on a 0 to 10 visual analog scale at 3-years (T3) and 5-years (T5) follow-up and served as the target variable. The dataset was split into 80% training and 20% testing sets. Various ML models (linear, tree-based ensemble, and neural network approaches) were evaluated using (1) mean squared error (where lower values indicate higher accuracy), (2) buffer accuracy within ± 1 and ± 2 points of the observed pain score, and (3) accuracy in classifying pain severity as low (0 to 3), moderate (4 to 6), or high (7 to 10).

Across 13 models, performance was the strongest for nonlinear ML approaches. The best-performing model, CatBoost, achieved low prediction error (mean squared error 2.70 at T3 and 4.11 at T5) and classified pain severity with 90.1% accuracy. Age, body mass index, history of back/neck problems, and specific items from the Western Ontario and McMaster Universities Osteoarthritis Index and Intermittent and Constant Osteoarthritis Pain tools emerged as highly influential predictors across multiple ML models. These findings informed an online tool that predicts postoperative pain scores using key features.

Multiple ML models demonstrated improved performance over traditional regression approaches in predicting long-term postoperative pain after primary THA and identified key preoperative predictors. If externally validated, these models may help identify patients at higher risk of persistent pain, guide preoperative counseling, and support targeted perioperative pain strategies.

III.