An Algorithmic Approach to Understanding Osteoarthritic Knee Pain

Background: Osteoarthritic knee pain is a complex phenomenon, and multiple factors, both within the knee and external to it, can contribute to how the patient perceives pain. We sought to determine how well a deep neural network could predict osteoarthritic knee pain and other symptoms solely from a single radiograph view.Methods: We used data from the Osteoarthritis Initiative, a 10-year observational study of patients with knee osteoarthritis. We paired >50,000 weight-bearing, posteroanterior knee radiographs with corresponding Knee Injury and Osteoarthritis Outcome Score (KOOS) pain, symptoms, and activities of daily living subscores and used them to train a series of deep learning models to predict those scores solely from raw radiographic input. We created regression models for specific score predictions and classification models to predict whether the modeled KOOS subscore exceeded a range of thresholds.Results: The root-mean-square errors were 15.7 for KOOS pain, 13.1 for KOOS symptoms, and 14.2 for KOOS activities of daily living. Modeling was performed to predict whether pain was above or below given pain thresholds, and was able to predict extreme pain (KOOS pain < 40) with an area under the curve (AUC) of 0.78. Notably, the system was also able to correctly predict numerous cases where the Kellgren-Lawrence (KL) grade assigned by the radiologist was 0 but patient pain was high, and cases where the KL grade was 4 but patient pain was low.Conclusions: A deep neural network can be trained to predict the osteoarthritic knee pain that a patient experienced and other symptoms with reasonable accuracy from a single posteroanterior view of the knee, even using low-resolution images. The system can predict pain and dysfunction that the traditional KL grade does not capture. Deep learning applied to raw imaging inputs holds promise for disentangling sources of pain within the knee from aggravating factors external to the knee.