Predicting knee osteoarthritis severity: comparative modeling based on patient's data and plain X-ray images

Area of Science:

• Orthopedics
• Medical Imaging
• Machine Learning

Background:

• Knee osteoarthritis (KOA) significantly impacts knee function and causes pain.
• Radiologists assess KOA severity using the Kellgren and Lawrence (KL) grading scale (0-4) on X-ray images.
• Subjectivity in KL grading remains a challenge in KOA assessment.

Purpose of the Study:

• To develop and compare predictive models for KOA severity using distinct data types.
• To evaluate the performance of machine learning models trained on X-ray images versus patient assessment data.
• To identify potential explanatory variables for early KOA monitoring.

Main Methods:

• Developed a Convolutional Neural Network (CNN) model using knee X-ray images.
• Built Elastic Net (EN) and Random Forests (RF) models utilizing patient assessment data (symptoms, medication use).
• Employed Linear Mixed Effect Models (LMM) to account for within-subject correlation between knees.

Main Results:

• CNN model achieved a Root Mean Squared Error (RMSE) of 0.77, outperforming EN (0.97) and RF (0.94).
• LMM demonstrated comparable prediction accuracy to EN regression while providing more reliable inference by handling data hierarchy.
• Identified significant explanatory variables from patient data for potential use in pre-imaging monitoring.

Conclusions:

• Machine learning models can effectively predict KOA severity using either X-ray images or patient data, achieving comparable results.
• The CNN model shows superior performance in predicting KOA severity from imaging data.
• Further research is needed to address the inherent subjectivity in KL grading.