SAT0116 DYNAMIC PREDICTION OF FLARES IN RHEUMATOID ARTHRITIS USING JOINT MODELLING AND MACHINE LEARNING: SIMULATION OF CLINICAL IMPACT WHEN USED AS DECISION AID IN A DISEASE ACTIVITY GUIDED DOSE REDUCTION STRATEGY. (June 2019)

• Record Type: Journal Article
• Authors: Welsing, Paco; Broeder, Alfons den; Tekstra, Janneke; Goes, Marlies van der; Lafeber, Floris; Jacobs, Johannes W. G.; Medina, Oj; Vodencarevic, Asmir; Zimmermann-Rittereiser, Marcus; Haitjema, Saskia; van Laar, Jacob M.
• Is Part Of: Annals of the rheumatic diseases. Volume 78(2019)Supplement 2
• Page Start: 1125
• Page End: 1126
• Publication Date: 2019-06
• Subjects: Rheumatism -- Periodicals; 616.723005
• DOI: 10.1136/annrheumdis-2019-eular.2881
• Languages: English
• ISSNs: 0003-4967

Abstract

Abstract : Background: Most rheumatoid arthritis (RA) patients on a (b)iological DMARD achieve and maintain remission or low disease activity. Then, tapering the drug is safe and cost-effective 1, but with increased risk of short lived flares and joint damage progression. We developed so-called dynamic prediction models, using joint-modelling (JM) and machine learning (ML) respectively to predict an individual patients' risk of flaring. These models can make short-term (e.g. 3 month) predictions repeatedly at every clinic visit, based on routine care data, i.e. the longitudinal course of the patients' disease activity and medication, demographic and disease characteristics. 2 Objectives: To externally validate the JM and ML model and evaluate the clinical impact on flare occurrence and bDMARD use, when using predictions as decision aid for tapering bDMARDs in a protocolised disease activity guided tapering strategy. Methods: For external validation, an RCT comparing protocolised dose reduction (stepwise increase injection interval, until flare or discontinuation, n=121) with usual care (n=59) was used. 1 Both models were applied to the trial data and AUC-ROC and calibration were assessed. To simulate our dose-reduction strategy, first, we defined optimal cut-offs for predictions using Youden's index as well as a cut-off at higher/lower predicted risks. Thereafter we applied them to the dose reduction arm of the trial. Assumptions were that 1) no further tapering was performed when the cut-off was reached without any flares occurring over time in the simulation and 2) when in the trial a flare occurred for which the bDMARD dose was increased to a value higher than when tapering was stopped in simulation, this flare occurred and the higher dose was used in simulation as well. Results: AUC-ROC's in external validation were, as expected, somewhat lower than in the development cohort: 0.76 (95%CI 0.71–0.85) and 0.73 (0.72–0.74) versus 0.78/0.79 for JM/ML respectively, but still satisfactory. Calibration was better for ML, resulting in different cut-offs for the models. Results show that about 40% and 50% of flares can be prevented respectively and that the majority (66% and 76%) of the dose reduction can be maintained using the optimal cut-offs in our prediction aided dose reduction strategy over 18 months. Table 1 shows the proportion of full dose used, mean number of flares per patient, and percentage of patients experiencing a flare for several cut-offs as well as in the disease activity guided dose reduction strategy (in bold). As context values are also shown for usual care (trial control arm; in italic). Conclusion: Both models proved to be externally valid. Using them to aid decisions on biological dose reduction has the potential to reduce the occurrence of flares significantly while largely retaining the reduction in dose as obtained by a disease activity guided dose reduction strategy. This finding, and its cost-effectiveness, should be validated in a clinical study. References: [1] Kievit W., et al.. Ann Rheum Dis2016, 75:1939-1944 [2] Vodencarevic A., et al.. DATA2018: 187-192 Disclosure of Interests: Paco Welsing: None declared, Alfons den Broeder: None declared, Janneke Tekstra: None declared, Marlies van der Goes: None declared, Floris Lafeber Shareholder of: ArthroSave, Grant/research support from: FOREUM; Dutch Arthritis Society, Johannes W. G. Jacobs Grant/research support from: Roche, Consultant for: Roche, OJ Medina: None declared, Asmir Vodencarevic Shareholder of: Owning shares of Siemens Healthcare GmbH as its employee. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Employee of: An employee of Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Marcus Zimmermann-Rittereiser Shareholder of: Owning shares of Siemens Healthcare GmbH as its employee. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Employee of: An employee of Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Saskia Haitjema: None declared, Jacob M. van Laar Grant/research support from: Genentech, Consultant for: F. Hoffmann-La Roche