Deep learning approach to predict pain progression in knee osteoarthritis

Abstract

Purpose: Pain progression in individuals with knee osteoarthritis (OA) is poorly understood and difficult to predict. Developing improved methods for predicting the progression of knee pain could identify individuals at high risk for pain progression who would be best suited for early intervention. This study was performed to determine the feasibility of a deep learning (DL) approach to predict knee pain progression in subjects in the Osteoarthritis Initiative (OAI) through analysis of demographic and radiographic risk factors and baseline knee radiographs. Methods: The study group consisted of 2000 subjects randomly selected from the OAI with Kellgren Lawrence (KL) grades 1, 2, or 3 pre-radiographic OA or radiographic OA of the knee at baseline. 1000 subjects showed pain progression and 1000 subjects did not show pain progression over a 60-month follow-up period. Pain progression was defined according to the Foundation of the National Institute of Health OA Biomarker Consortium Project as a 9 point or greater increase in WOMAC score (0-100 scale) between 2 or more time-points from 24-month to 60-months follow-up. The study group was randomly divided into a training dataset (742 subjects with and 758 subjects without pain progression) used to train the models, a validation dataset (108 subjects with and 92 subjects without pain progression) used to choose the most optimal models, and a hold-out test dataset (150 subjects with and 150 subjects without pain progression) used to evaluate the diagnostic performance of the models. An artificial neural network (ANN) was used to create a clinical model to provide a confidence score for predicting pain progression using demographic and radiographic risk factors including baseline age, gender, race, body mass index (BMI), history of knee injury, KL grade, and tibiofemoral angle. A DL model consisted of two separate deep convolutional neural networks (YOLO and DenseNet). The first network (YOLO) was used to crop regions of interest around each individual knee joint on the standing bilateral posterior-anterior knee radiographs. The second classification network (DenseNet) provided a confidence score for predicting pain progression based on DL analysis of the cropped knee images. A combined model was created by integrating both clinical data and DL analysis of the baseline knee radiographs. The combined model used DenseNet to extract radiograph information as a feature vector which was further concatenated with the demographic and radiographic risk factors data vector. The combined feature vector was then fed to a fully connected network for joint model training. Receiver operation characteristic (ROC) and area under the curve (AUC) analysis was performed using the hold-out test dataset to evaluate the diagnostic performance of the clinical, DL, and combined clinical and DL models for predicting pain progression. Delong’s method was used to compare the AUCs of the different models. The optimal sensitivity and specificity of the models were determined using the Youden index. Results: Figure 1 shows the ROC curves for the clinical, DL, and combined clinical and DL models for predicting pain progression. Table 1 shows the AUCs and the optimal sensitivity and specificity of the models. The clinical model had the lowest AUC of 0.644 with 61.7% sensitivity and 68.9% specificity for predicting pain progression. The DL model had AUCs of 0.753, which was significantly higher (p<0.05) than the AUC of the clinical model. The DL model also had higher (p<0.05) sensitivity and specificity than the clinical model for predicting pain progression. The combined clinical and DL model had the highest AUC of 0.804 with 75.2% sensitivity and 76.2% specificity for predicting pain progression. The AUC of the combined clinical and DL model was significantly higher (p<0.001) than the AUCs of the clinical and DL model, indicating greater overall diagnostic performance for predicting pain progression. Conclusions: Our study demonstrated the feasibility of a combined clinical and DL model for predicting the progression of knee pain through analysis of demographic and radiographic risk factors and baseline knee radiographs. The combined clinical and DL model showed a significant improvement in diagnostic performance for predicting knee pain progression when compared to the clinical model and DL model alone.Tabled 1AUCs and optimal sensitivity and specificity of the models for predicting knee pain progression.AUC (95% CI)Sensitivity (95% CI)Specificity (95% CI)Clinical Model0.644 (0.587 - 0.698)61.74 (53.4 - 69.6)68.87 (60.8 - 76.2)DL model0.753 (0.700 - 0.801)65.77 (57.6 - 73.3)73.51 (65.7 - 80.4)Combined Clinical and DL model0.804 (0.754 - 0.847)75.17 (67.4 - 81.9)76.16 (68.6 - 82.7) Open table in a new tab