Prediction of progression rate and fate of osteoarthritis: Comparison of machine learning algorithms

Abstract

Appropriate prediction models can assist healthcare systems in delaying or reversing osteoarthritis (OA) progression. We aimed to identify a reliable algorithm for predicting the progression rate and fate of OA based on patient-specific information. From May 2003 to 2019, 83,280 knees were collected. Age, sex, body mass index, bone mineral density, physical demands for occupation, comorbidities, and initial Kellgren-Lawrence (K-L) grade were used as variables for the prediction models. The prediction targets were divided into dichotomous groups for even distribution. We compared the performances of logistic regression (LR), random forest (RF), and extreme gradient boost (XGB) algorithms. Each algorithm had the best precision when the model used all variables. XGB showed the best results in accuracy, recall, F1 score, specificity, and error rates (progression rate/fate of OA: 0.710/0.877, 0.542/0.637, 0.637/0.758, 0.859/0.981, and 0.290/0.123, respectively). The feature importance of RF and XGB had the same order up to the top six for each prediction target. Age and initial K-L grade had the highest feature importance in RF and XGB for the progression rate and fate of OA, respectively. The XGB and RF machine learning algorithms showed better performance than conventional LR in predicting the progression rate and fate of OA. The best performance was obtained when all variables were combined using the XGB algorithm. For each algorithm, the initial K-L grade and physical demand for occupation were the greatest contributors with superior feature importance compared with the others.

Keywords: fate; logistic regression; machine learning; osteoarthritis; prediction; progression rate.