Simplifying Knee OA Prognosis: A Deep Learning Approach Using Radiographs and Minimal Clinical Inputs

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Osteoarthritis Research

Background:

• Knee osteoarthritis (OA) is a degenerative joint disease requiring accurate progression prediction for timely intervention.
• Current prediction methods may lack the precision needed for effective clinical management.
• Deep learning models offer potential for enhanced diagnostic and prognostic capabilities in OA.

Purpose of the Study:

• To develop and validate a deep convolutional neural network, specifically a vision transformer-based model, for predicting knee OA progression.
• To assess the model's performance using baseline knee radiographs and clinical data.
• To evaluate the model's ability to identify patients likely to progress and require surgical intervention.

Main Methods:

• A vision transformer model was trained on 5565 knee radiographs from the Osteoarthritis Initiative (OAI) dataset.
• The model utilized baseline images and clinical data (full or essential factors, single or paired images).
• External validation was performed on 274 cases from Far Eastern Memorial Hospital, assessing metrics like AUROC, accuracy, sensitivity, and specificity.

Main Results:

• The model achieved an AUROC of 0.808 for OA progression prediction in OAI testing (simplest input: single image with essential factors).
• External validation demonstrated an AUROC of 0.709 for OA progression.
• The model showed strong predictive power for identifying surgical candidates, with an odds ratio of 23.87 in OAI and 5.92 in external validation.

Conclusions:

• The developed vision transformer model provides reliable and efficient prediction of knee OA progression.
• The model's simplicity and flexibility enhance its clinical applicability.
• This AI tool has the potential to significantly improve early intervention strategies for OA patients.