PREDICTING FUNCTIONAL OUTCOMES AFTER TOTAL KNEE ARTHROPLASTY USING MACHINE LEARNING ALGORITHMS

초록

Background: Total knee arthroplasty (TKA) is an effective treatment for patients with end-stage knee osteoarthritis (OA). After TKA, most patients show improved physical function (e.g., walking speed and endurance), however, 10–30 % of patients continue to have functional limitations more than a year later. Objectives: This study aims to develop a predictive model based on preoperative variables to predict functional outcomes three months after TKA using machine learning algorithms. Methods: This retrospective cohort study involved 313 patients who underwent TKA between September 2013 and January 2022. The analyzed data consisted of demographic & anthropometric variables, clinical characteristics, and functional assessments (Western Ontario and McMaster Universities Osteoarthritis Index [WOMAC] total score, Visual Analogue Scale [VAS] score, EQ-5D total score, Timed Up and Go [TUG], 6-minute walking test [6MWT], stair climbing test [SCT] ascending and descending, peak torque (PT) of knee extensor strength and knee flexor strength. These measurements were taken both preoperatively and 3 months postoperatively. Five machine learning algorithms were used to predict postoperative functional outcomes including Linear Regression, Poisson Regressor, Tweedie Regressor, Passive Aggressive Regressor, and ARD Regression. External validation was performed using new external data from 7 individuals. Feature importance was assessed using bar plots to visualize the contribution of each variable to the model`s predictions. SHAP (SHapley Additive exPlanations) values were used to interpret model predictions and understand the impact of each feature on the outcome. The 10 most important variables analyzed in each model`s SHAP and feature importance, as well as the variables that showed statistically significant results in the multivariate analysis, were comprehensively considered and selected. Results: The average age of a total of 313 patients was 71.80 (± 5.91) years and 49 (15.65% of the total) patients were males (Table 1). In the analysis conducted to identify key preoperative factors influencing postoperative outcomes, different machine learning models demonstrated the best performance for each factor. Significantly important variables for predicting postoperative functional outcomes were TUG, SCT ascending and descending, and PT of knee flexors and knee extensors. Detailed findings are listed in Table 2. Overall, the predictive model incorporating these variables demonstrated strong explanatory power in predicting postoperative functional recovery. Conclusion: The machine learning models incorporating preoperative variables demonstrated strong explanatory power in predicting postoperative functional recovery after TKA. Therefore, measuring these key preoperative variables, particularly clinical characteristics and functional assessments, would be crucial for predicting postoperative prognosis and recovery.