Advanced machine learning for predicting individual risk of flares in rheumatoid arthritis patients tapering biologic drugs

• Main Authors: Asmir Vodencarevic, Koray Tascilar, Fabian Hartmann, Michaela Reiser, Axel J. Hueber, Judith Haschka, Sara Bayat, Timo Meinderink, Johannes Knitza, Larissa Mendez, Melanie Hagen, Gerhard Krönke, Jürgen Rech, Bernhard Manger, Arnd Kleyer, Marcus Zimmermann-Rittereiser, Georg Schett, David Simon, on behalf of the RETRO study group
• Format: Article
• Language: English
• Published: BMC 2021-02-01
• Series: Arthritis Research & Therapy
• Subjects: Rheumatoid arthritis; Machine learning; Flare prediction
• Online Access: https://doi.org/10.1186/s13075-021-02439-5

Abstract Background Biological disease-modifying anti-rheumatic drugs (bDMARDs) can be tapered in some rheumatoid arthritis (RA) patients in sustained remission. The purpose of this study was to assess the feasibility of building a model to estimate the individual flare probability in RA patients tapering bDMARDs using machine learning methods. Methods Longitudinal clinical data of RA patients on bDMARDs from a randomized controlled trial of treatment withdrawal (RETRO) were used to build a predictive model to estimate the probability of a flare. Four basic machine learning models were trained, and their predictions were additionally combined to train an ensemble learning method, a stacking meta-classifier model to predict the individual flare probability within 14 weeks after each visit. Prediction performance was estimated using nested cross-validation as the area under the receiver operating curve (AUROC). Predictor importance was estimated using the permutation importance approach. Results Data of 135 visits from 41 patients were included. A model selection approach based on nested cross-validation was implemented to find the most suitable modeling formalism for the flare prediction task as well as the optimal model hyper-parameters. Moreover, an approach based on stacking different classifiers was successfully applied to create a powerful and flexible prediction model with the final measured AUROC of 0.81 (95%CI 0.73–0.89). The percent dose change of bDMARDs, clinical disease activity (DAS-28 ESR), disease duration, and inflammatory markers were the most important predictors of a flare. Conclusion Machine learning methods were deemed feasible to predict flares after tapering bDMARDs in RA patients in sustained remission.