Machine Learning to Predict Remission Between 6 and 24 Months in Rheumatoid Arthritis: Insights From JAK, an...

Area of Science:

• Rheumatology
• Artificial Intelligence in Medicine
• Clinical Data Science

Background:

• Predicting treatment response in rheumatoid arthritis (RA) is crucial for optimizing patient care.
• Biologic and targeted synthetic disease-modifying antirheumatic drugs (bDMARDs/tsDMARDs) are key RA therapies.
• Identifying patients likely to achieve remission early can guide treatment selection.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting remission in RA patients starting new therapies.
• To externally validate the ML model using data from international registries.
• To simplify the ML model and compare its performance against logistic regression.

Main Methods:

• Utilized data from 11 international RA registries (JAK-pot collaboration) involving 21,675 treatment courses.
• Trained an XGBoost ML model on 63 baseline variables to predict remission (CDAI ≤2.8) between 6-24 months.
• Externally validated the model using 1807 treatment courses and evaluated performance using AUC, sensitivity, specificity, PPV, and NPV.

Main Results:

• The full ML model achieved an external validation AUC of 0.797, with good negative predictive value (0.902) but moderate positive predictive value (0.454).
• A simplified ten-variable model performed comparably (AUC: 0.802), as did a logistic regression model (AUC: 0.809).
• Key predictors included patient global assessment, 28-tender joint count, prior bDMARD/tsDMARD exposure, and HAQ-DI.

Conclusions:

• Machine learning models using routine baseline data show limited ability to predict RA remission, excelling primarily at ruling out remission.
• Model complexity did not significantly improve predictive performance over simpler models or logistic regression.
• Richer predictor variables may be needed to enhance the predictive accuracy of remission in RA patients.