Machine Learning-Based prediction of complications and residual pain after total knee arthroplasty

Accurate risk adjustment is critical for outcome prediction and quality improvement in total knee arthroplasty (TKA). While machine learning (ML) offers promising capabilities, most models rely solely on patient demographics and comorbidities. The American Association of Hip and Knee Surgeons (AAHKS) has proposed a set of nine risk factors to enhance current models. This study aimed to evaluate these factors using a machine learning model based on eXtreme Gradient Boosting (XGBoost).

We retrospectively analyzed 783 patients who underwent primary TKA at a single academic center between January 2020 and December 2022. Preoperative clinical data and AAHKS-defined risk factors were used to train and evaluate an XGBoost model. The primary outcome measures were: (1) major complications requiring revision, (2) any complication (major or minor), and (3) residual pain at one year (Visual Analog Scale ≥4). Model performance was assessed using area under the curve (AUC), sensitivity, specificity, and accuracy. Feature importance was determined using SHapley Additive exPlanations (SHAP).

The model achieved moderate predictive accuracy for major complications (AUC = 0.68) and any complication (AUC = 0.65), but performed poorly in predicting residual pain (AUC = 0.53). Among AAHKS-defined risk factors, only "smoking" and "previous open reduction and internal fixation (ORIF) of the knee" showed high predictive value. Other proposed variables, such as angular deformity >15°, had limited impact.

An XGBoost-based ML model incorporating AAHKS-defined risk factors showed moderate effectiveness in predicting postoperative complications following TKA. However, the model was unable to reliably predict residual pain. These findings underscore the need for broader inclusion of joint-specific variables and imaging data in future risk adjustment frameworks to enhance personalized care in knee arthroplasty.