Machine learning algorithms do not outperform preoperative thresholds in predicting clinically meaningful improvements after total knee arthroplasty.

Abstract

Patient-reported outcome measures (PROMs) are important measures of success after total knee arthroplasty (TKA) and being able to predict their improvements could enhance preoperative decision-making. Our study aims to compare the predictive performance of machine learning (ML) algorithms and preoperative PROM thresholds in predicting minimal clinically important difference (MCID) attainment at 2years after TKA. Prospectively collected data of 2840 primary TKA performed between 2008 and 2018 was extracted from our joint replacement registry and split into a training set (80%) and test set (20%). Using the training set, ML algorithms were developed using patient demographics, comorbidities and preoperative PROMs, whereas the optimal preoperative threshold was determined using ROC analysis. Both methods were used to predict MCID attainment for the SF-36 PCS, MCS and WOMAC at 2years postoperatively, with predictive performance evaluated on the independent test set. ML algorithms and preoperative PROM models performed similarly in predicting MCID for the SF-36 PCS (AUC: 0.77 vs 0.74), MCS (AUC: 0.95 vs 0.95) and WOMAC (AUC: 0.89 vs 0.88). For each outcome, the most important predictor of MCID attainment was the patient's preoperative PROM score. ROC analysis also identified optimal preoperative threshold values of 33.6, 54.1 and 72.7 for the SF-36 PCS, MCS and WOMAC, respectively. ML algorithms did not perform significantly better than preoperative PROM thresholds in predicting MCID attainment after TKA. Future research should routinely compare the predictive ability of ML algorithms with existing methods and determine the type of clinical problems which may benefit the most from it. II.