Machine Learning Better Predicts Mortality in Patients Undergoing Revision Total Hip Arthroplasty for Periprosthetic Joint Infection

Abstract

Periprosthetic joint infection (PJI) following total hip arthroplasty (THA) is associated with a 25% 5-year mortality. Synthetic Minority Oversampling Technique (SMOTE) is a novel machine learning technique used in other surgical specialties to better predict outcomes, but has never been utilized in the field of orthopedics. This is the first study utilizing machine learning (ML) and SMOTE to predict the greatest risk factors for mortality following revision THA for PJI. This retrospective study utilized the Nationwide Readmissions Database to identify adult patients undergoing revision THA for PJI. Model performance was assessed using AUROC, Brier score, and F1 score, with SHapley Additive exPlanation (SHAP) analysis identifying key predictors. Among 19,099 patients undergoing revision THA for PJI, there was a 0.8% mortality rate. Logistic regression (AUROC 0.855) and gradient boosting (AUROC 0.862) outperformed random forest (AUROC 0.765), with gradient boosting demonstrating the highest F1 score (0.257). Following the application SMOTE, all models demonstrated improved AUROC, model calibration, and enhanced F1. SHAP analysis identified fluid and electrolyte disorders, advancing age, and cardiac arrhythmia as key predictors of mortality. This study is the first in the field of orthopedics to utilize SMOTE and subsequently introduces a novel, streamlined risk score that identifies high-risk patients undergoing revision THA for PJI. Fluid and electrolyte abnormalities were the strongest predictors of mortality. Given the established link between high-dose antibiotic-loaded bone cement spacers and calcium sulfate antibiotic impregnated beads with AKI and subsequent electrolyte disturbances, optimizing electrolyte levels perioperatively may be a mortality-reducing intervention.