Preoperatively predicting failure to achieve the minimum clinically important difference and substantial clinical benefit for total knee arthroplasty patients using machine learning.

Background: A clear understanding of minimal clinically important difference (MCID) and substantial clinical benefit (SCB) is essential for effectively implementing patient-reported outcome measurements (PROMs) as a performance measure for total knee arthroplasty (TKA). Since not achieving MCID and SCB may reflect suboptimal surgical benefit, the primary aim of this study was to use machine learning to predict patients who may not achieve the threshold-based outcomes (i.e., MCID and SCB) on the Knee Injury and Osteoarthritis Outcome Score for Joint Replacement (KOOS JR) following TKA.

Methods: Data from 1064 patients who underwent TKA at a single academic medical center between 2016 and 2022 contained 81 preoperative variables, including routinely collected measures and PROMs (KOOS JR and Patient-Reported Outcomes Measurement Information Systems [PROMIS-10]). Several machine-learning models were developed, which include penalized logistic regression as a linear model, support vector machine with polynomial and radial kernels as nonlinear models, and random forest and extreme gradient boosting as nonparametric models. These models predicted both distribution- and anchor-based MCIDs and SCB. In addition, logistic regression models were used to identify relevant risk factors for failing to meet these thresholds.

Results: The random forest models and the penalized logistic regression models achieved acceptable area under the receiver operating characteristic curve (AUC) close to or above 0.7 for all the outcomes. Furthermore, the logistic regression models identified shared risk factors for the three outcomes: preoperative PROMs (i.e., KOOS JR score, PROMIS-10 global physical T-score, and PROMIS-10 general mental health), antidepressant medication history, age, and Kellgren-Lawrence grade.

Conclusions: Machine-learning models were able to identify patients at risk of failure to achieve the threshold-based metrics and relevant preoperative factors. As such, these models may be used to both improve shared decision-making and help create risk-stratification tools to improve quality assessment of surgical outcomes.