Identifying who are unlikely to benefit from total knee arthroplasty using machine learning models

Abstract

Identifying and preventing patients who are not likely to benefit long-term from total knee arthroplasty (TKA) would decrease healthcare expenditure significantly. We trained machine learning (ML) models (image-only, clinical-data only, and multimodal) among 5720 knee OA patients to predict postoperative dissatisfaction at 2 years. Dissatisfaction was defined as not achieving a minimal clinically important difference in postoperative Knee Society knee and function scores (KSS), Short Form-36 Health Survey [SF-36, divided into a physical component score (PCS) and mental component score (MCS)], and Oxford Knee Score (OKS). Compared to image-only models, both clinical-data only and multimodal models achieved superior performance at predicting dissatisfaction measured by AUC, clinical-data only model: KSS 0.888 (0.866–0.909), SF-PCS 0.836 (0.812–0.860), SF-MCS 0.833 (0.812–0.854), and OKS 0.806 (0.753–0.859); multimodal model: KSS 0.891 (0.870–0.911), SF-PCS 0.832 (0.808–0.857), SF-MCS 0.835 (0.811–0.856), and OKS 0.816 (0.768–0.863). Our findings highlighted that ML models using clinical or multimodal data were capable to predict post-TKA dissatisfaction.