Development of Machine Learning Models to Predict Improvement in Knee Injury and Osteoarthritis Outcome Score After Total Knee Arthroplasty.

This study was aimed to create an artificial intelligence model to predict the Knee injury and Osteoarthritis Outcome Score (KOOS) in the early preoperative phase of TKA and implement it as an application.This study was aimed to create an artificial intelligence model to predict the Knee injury and Osteoarthritis Outcome Score (KOOS) in the early preoperative phase of TKA and implement it as an application.

Prospectively collected data from 236 patients who underwent TKA were used. The objective variables were whether the patient would achieve the minimum clinically important difference (MCID) of the following 1-6 items from preoperatively to 1 year postoperatively (1. total KOOS score, 2. subitem symptoms (S), 3. pain (P), 4. Activities of daily living (A), 5. sports and recreation (SP), 6. quality of life (Q)). The predictive variables were patient background (age, gender, body mass index), preoperative physical function (knee flexion angle, extension angle, the quadriceps and hamstrings strength, comfortable walking speed, pain visual analogue scale (VAS) during walking) and preoperative KOOS (total score and 5 subitems). For each objective variable, several machine learning algorithms ((1) Random Forest, (2) eXtreme Gradient boosting (XGboost), (3) Neural Network, (4) Logistic Regression) were trained to create a prediction model with maximum accuracy. The area under the ROC curve (AUC) was used to assess prediction accuracy.

The highest AUC was 0.86 for KOOS total score by random forest, 0.94 for KOOS-S, 0.95 for KOOS-P, 0.88 for KOOS-A, 0.73 for KOOS-SP and 0.98 for KOOS-Q by XGboost. These prediction models created was made into an application. The main predictive variables for predicting the KOOS total score were preoperative KOOS total score, KOOS-P, knee flexion angle, gait speed, age and pain VAS.

Our application can predict with a high accuracy whether the MCID can be achieved one year after TKA, by performing an assessment before TKA and inputting the results. In the future, prospective accuracy validation, including external validation, should be carried out.

This application can be used for explanation and setting goals for the patients. Furthermore, it can be useful in providing tailored rehabilitation focused on improving physical functions that are predicted not to be achieved in the early preoperative phase of TKA.