Development of a novel machine learning model for the prediction of periprosthetic joint infection following primary total knee arthroplasty: a 23-year retrospective study

Abstract

<p>Introduction: Periprosthetic joint infection (PJI) is a significant complication of primary total knee arthroplasty (TKA). A prediction tool to assist clinical preoperative risk assessment is important. However, no such model is tailored for Hong Kong patients. This study aimed to develop a machine learning (ML)-based model for predicting PJI following primary TKA in Hong Kong.<br>Methods: A retrospective analysis was conducted in a local teaching hospital on 3,483 primary TKA (81 with PJI) from 1998 to 2021. We gathered 61 features, encompassing patient demographics, operation-related variables, laboratory findings and comorbidities. Six of them were selected by univariate and multivariate analysis. We trained an Easy Ensemble classifier with Random Forest as the base estimator using stratified 10-fold cross-validation and compared it with Logistic Regression to verify ML performance.<br>Results: The ML model demonstrated stable and robust performance across ten folds, with average metrics of 0.913 for area under the receiver operating curve, 0.831 for balanced accuracy, 0.839 for sensitivity, and 0.822 for specificity, outperforming the logistic regression model (AUC=0.739). The significant risk factors identified were long operative time (HR= 9.15; p=0.017), male (HR=3.09; p<0.001), ASA>2 (HR=1.65; p=0.028), history of anaemia (HR=2.18; p=0.023) and history of septic arthritis (HR=4.38; p=0.029). Spinal anaesthesia (HR=0.55; p=0.023) was a significant protective factor.<br>Discussion and Conclusion: We developed the first ML-based model for predicting PJI following primary TKA in Hong Kong, demonstrating its superiority over statistical methods. It may assist the preoperative treatment decision-making and patient health optimization.</p>