Estimating time-to-total knee replacement on radiographs and MRI: a multimodal approach using self-supervised deep...

Area of Science:

• Biomedical Engineering
• Radiology
• Orthopedics

Background:

• Accurate prediction of time to total knee replacement (TKR) is essential for patient management and healthcare planning.
• Current models lack the precision for short-term (e.g., 3-year) time-based predictions.
• Timely surgical intervention planning and resource allocation require precise TKR prediction.

Purpose of the Study:

• To develop a survival analysis model for predicting the time-to-TKR.
• To enhance prediction accuracy using medical imaging and clinical data.
• To provide a tool for more precise patient management and healthcare resource allocation.

Main Methods:

• Utilized data from the Osteoarthritis Initiative, Multi-Center Osteoarthritis Study, and internal hospital records.
• Employed deep learning models to extract features from radiographs and MR scans.
• Integrated imaging features, clinical variables, and image assessments into a survival analysis model with Lasso Cox feature selection and a random survival forest.

Main Results:

• The proposed model effectively integrated self-supervised deep learning features with clinical and imaging data.
• Achieved strong discrimination power with an area under the curve (AUC) of 94.5 (95% CI, 94.0-95.1) for time-to-TKR prediction.
• Demonstrated high accuracy by combining multimodal data, including patient demographics, pain scores, and imaging-derived measurements like joint space narrowing and cartilage morphology.

Conclusions:

• The developed model shows significant potential for accurately predicting time-to-TKR.
• Self-supervised learning and multimodal data fusion are key to improving prediction accuracy.
• This predictive capability can assist physicians in developing personalized treatment strategies for patients.