Surgical Outcome Prediction in Total Knee Arthroplasty using Machine Learning

摘要

This work aimed to predict postoperative knee functions of a new patient prior to total knee arthroplasty (TKA) surgery using machine learning, because such prediction is essential for surgical planning and for patients to better understand the TKA outcome. However, the main difficulty is to determine the relationships among individual varieties of preoperative and postoperative knee kinematics. The problem was solved by constructing predictive models from the knee kinematics data of 35 osteoarthritis patients, operated by posterior stabilized implant, based on generalized linear regression (GLR) analysis. Two prediction methods (without and with principal component analysis followed by GLR) along with their sub-classes were proposed, and they were finally evaluated by a leaveone-out cross-validation procedure. The best method can predict the postoperative outcome of a new patient with a Pearson's correlation coefficient (cc) of 0.84 +/- 0.15 (mean +/- SD) and a root-mean-squared-en or (RMSE) of 3.27 +/- 1.42 mm for anterior-posterior vs. flexion/extension (A-P pattern), and a cc of 0.89 +/- 0.15 and RMSE of 4.25 +/- 1.92 degrees for internal-external vs. flexion/extension (i-e pattern). Although these were validated for one type of prosthesis, they could be applicable to other implants, because the definition of knee kinematics, measured by a navigation system, is appropriate for other implants.