Abstract Background Identifying patients unlikely to achieve meaningful improvement following total knee arthroplasty (TKA) supports more effective shared decision-making (SDM). This study aimed to develop and validate machine learning (ML) models that preoperatively predict insufficient clinical improvement one year after TKA using Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) subscales and total scores, and to assess the important predictive variables. Methods A retrospective analysis was performed on consecutive primary TKA patients from 2004 to 2022 at a single tertiary hospital was conducted. Insufficient clinical improvement was defined as not achieving the minimal clinically important difference (MCID) for each WOMAC subscale and total. Candidate preoperative variables included demographics, comorbidities, knee range of motion, radiologic variables, and WOMAC scores. A variety of ML models were evaluated using performance metrics for calibration and discrimination, as well as decision curve analysis and Shapley additive explanations. Results Among the 3,810 TKAs included, the ExtraTrees model performed best for WOMAC pain, stiffness, function, and total scores, achieving AUCs of 0.92, 0.90, 0.87, and 0.89; recall rates of 0.79, 0.86, 0.70, and 0.83; and Brier scores of 0.09, 0.10, 0.11, and 0.06, respectively, along with demonstrating good calibration curves and net clinical benefit. Shapley additive explanations identified better preoperative WOMAC scores, osteoporosis, diabetes mellitus, older age, malignancy, and coronary artery disease as important predictors of insufficient clinical improvement. Conclusions The ML models demonstrated good performance in preoperatively predicting insufficient clinical improvement at 1 year after TKA based on WOMAC. These models have the potential to enhance SDM and perioperative patient management by preoperatively identifying approximately 70% to over 80% of patients likely to experience insufficient clinical improvement, with a specificity of about 80%, and by providing explanations regarding associated factors.
Gimm et al. (Wed,) studied this question.