Machine Learning Algorithm to Predict Worsening of Flexion Range of Motion After Total Knee Arthroplasty.

Background: Predicting the worsening of flexion range of motion (ROM) during the course post-total knee arthroplasty (TKA) is clinically meaningful. This study aimed to create a model that could predict the worsening of knee flexion ROM during the TKA course using a machine learning algorithm and to examine its accuracy and predictive variables. Methods: Altogether, 344 patients (508 knees) who underwent TKA were enrolled. Knee flexion ROM worsening was defined as ROM decrease of ≥10° between 1 month and 6 months post-TKA. A predictive model for worsening was investigated using 31 variables obtained retrospectively. 5 data sets were created using stratified 5-fold cross-validation. Total data (n = 508) were randomly divided into training (n = 407) and test (n = 101) data. On each data set, 5 machine learning algorithms (logistic regression, support vector machine, multilayer perceptron, decision tree, and random forest) were applied; the optimal algorithm was decided. Then, variables extracted using recursive feature elimination were combined; by combination, random forest models were created and compared. The accuracy rate and area under the curve were calculated. Finally, the importance of variables was calculated for the most accurate model. Results: The knees were classified into the worsening (n = 124) and nonworsening (n = 384) groups. The random forest model with 3 variables had the highest accuracy rate, 0.86 (area under the curve, 0.72). These variables (importance) were joint-line change (1.000), postoperative femoral-tibial angle (0.887), and hemoglobin A1c (0.468). Conclusions: The random forest model with the above variables is useful for predicting the worsening of knee flexion ROM during the course post-TKA.