Machine Learning-based Knee Pain Assessment Using the Upright CT of Patients with Patellofemoral Pain

Patellofemoral pain (PFP), a common disease in 20-30% of knee pain patients, is caused by muscle imbalance, cartilage shortening, and the abnormal alignment of joints. If the joint moves abnormally, the pressure applied to the cartilage is not evenly dispersed, causing pain, which can be said to be the main cause of PFP. The goal of this study is to develop an algorithm that can predict patellofemoral pain using only three-dimensional (3-D) computed tomography (CT) images acquired under a weight-bearing posture. A total of eight two-dimensional (2-D) or 3-D metrics that quantitatively represent the alignment of the patellofemoral joint were measured. The coordinates of the anatomical landmarks necessary to derive the corresponding quantitative metrics from the acquired CT images were extracted by three experienced evaluators. The correlation between the eight metrics and pain scale was analyzed using nine representative machine learning-based algorithms. As a result, the proposed algorithms were able to effectively predict the pain level with a high accuracy of less than 10%, and the error based on the highest performance algorithm for each biomarker was set.