MACHINE LEARNING-BASED MODEL FOR PREDICTING INFECTION ERADICATION IN PERIPROSTHETIC JOINT INFECTIONS

Aim Predicting treatment outcomes in periprosthetic joint infections (PJI) remains challenging as numerous surgical, patient-specific, and implant-related variables influence outcomes throughout the entire treatment course over several weeks. In Addition, therapy regimes may change based on pathogen identification or wound complications, impacting and altering infection eradication. This study aimed to evaluate a machine learning (ML) model for predicting infection eradication using a specific PJI database for this purpose, focusing on (a) predictive accuracy and (b) the influence of pre- and postoperative variables. Method A retrospective analysis of 649 PJI cases treated at a university hospital (2010–2023) included 218 clinical variables. After data cleaning, scaling, and preprocessing, 57 variables from 413 cases with ≥12-month follow-up were used to predict infection eradication according to the 2013 Delphi consensus criteria. Two XGBoost classifiers were trained: one using only 39 preoperative variables and one using the full 57-variable dataset. LASSO regression guided feature selection. Model performance was assessed via accuracy, area under the curve (AUC), sensitivity, and specificity. Univariate logistic regression identified significant clinical predictors for comparison with conventional statistics. Results The ML model using all variables achieved an AUC of 0.89, sensitivity of 0.87, and specificity of 0.84. Using only preoperative data, AUC remained high at 0.88 (sensitivity 0.85, specificity 0.81). Logistic regression identified key risk factors, including prior infection (OR 0.19), fistulas (OR 0.23), rifampicin use (OR 2.8), oral antibiotics (OR 0.68), surgery duration (OR 0.03), histology at reimplantation (OR 4.1), number of previous surgeries (OR 0.29), and steroid therapy (OR 0.02). All were confirmed by the ML model, which also revealed additional potential predictors. Conclusions Machine learning enabled accurate prediction of PJI resolution. Notably, preoperative data alone yielded robust predictive performance. The model validated known risk factors and uncovered further clinically plausible predictors, warranting future prospective validation.