SAT0055 PREDICTION OF FLARES FOR RHEUMATOID ARTHRITIS PATIENTS ON BIOLOGIC DMARDS USING MACHINE LEARNING AND SUBSETS OF VARIABLES AVAILABLE TO PHYSICIANS, PATIENTS AND PAYERS

Background:Today approximately 50 percent of patients with RA reach sustained remission. In a specific subset of RA patients in stable remission, biological Disease Modifying Anti-Rheumatic Drugs (bDMARDs) may be successfully tapered. However, it remains challenging to predict the patients’ individual flare risk. As we have recently shown, machine learning based on extensive clinical and laboratory data could be used to estimate individual flare risk [1].Objectives:In this study we aimed to investigate the performance of machine learning models trained with variables that are typically (1) immediately available to a physician during patient visits (clinical and demographic variables without laboratory values and composite disease activity scores), (2) theoretically available to patients at home and (3) available to payers in large health-system databases.Methods:Longitudinal clinical data of RA patients on bDMARDs from the first interim analysis of the phase-3, multicentre, randomised, open, prospective, controlled, parallel-group RETRO study (EudraCT number 2009-015740-42) was used [2] to build a predictive model for estimating the flare probability within 3 months from the current patient visit. A flare was defined as a DAS-28 ESR score over 2.6. Four different models (log. regression, random forest, k-NN and naïve Bayes) were trained which output the flare probability at each patient visit. These probabilities were used as an input for a stacking logistic regression meta-classifier [3]. The final model performance expressed as the AUROC was assessed using nested cross-validation [4]. We applied this method to three variable subsets (physician, patient, payer, Table 1).Table 1.List of variables used in three subsets:Variable / RolePhysicianPatientPayerGender (m/f)xxxDisease duration (years)xxxMethotrexate co-use (yes/no)xxxOther DMARDs co-use (yes/no)xxxDrug ATC codexxxIV-administration (yes/no)xxxDose percentagexxxAgexxxBody mass indexxxxDose percentage changexxxSwollen joint countxTender joint countxVAS_GH (pat. global disease activity)xxHAQ (health assessment questionnaire)xxSmoking status (yes/no/ex)xxAlcohol consumption (yes/no)xxPrevious flares (yes/no)xResults:Data from 135 follow-ups of 41 patients were used. The measured AUROC of the best performing model using all RETRO variables was 0.802 (95%CI 0.717 – 0.887) [1]. When a subset based on demographic and clinical variables is used that is available to a physician immediately during a patient visit the AUROC drops about 5 percent points. When only variables theoretically available to patients at home are used, the performance drops about 10 percent points comparing to the original model. Similar observation holds for the variable subset typically available to payers (Figure 1).Conclusion:This study shows that predictive models for flares have the potential to support physicians in making decisions immediately during the patient visit, even though laboratory values and respective activity scores are not yet available. In the future, machine learning applications may allow fast and reliable decisions on flare prediction in RA patients. These data can guide decisions about DMARD tapering at in real time during the physician-patient contact and allow to reduce costs not only by selective treatment tapering but also by sparing additional laboratory examinations.References:[1] Vodencarevic A. et al. Arthritis Rheumatol. 2019; 71[2] Haschka J et al. Ann Rheum Dis 2016; 75:45-51.[3] Tang J et al. CRC Press 2015; 498-500[4] Cawley GC et al. J Mach Learn Res 2010; 11:2079-2107Disclosure of Interests:Asmir Vodencarevic Shareholder of: Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Employee of: Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Koray Tascilar: None declared, Fabian Hartmann: None declared, Michaela Reiser: None declared, Sara Bayat Speakers bureau: Novartis, Johannes Knitza Grant/research support from: Research Grant: Novartis, Larissa Valor: None declared, Melanie Hagen: None declared, Axel Hueber Grant/research support from: Novartis, Lilly, Pfizer, Consultant of: Abbvie, BMS, Celgene, Gilead, GSK, Lilly, Novartis, Speakers bureau: GSK, Lilly, Novartis, Arnd Kleyer Consultant of: Lilly, Gilead, Novartis,Abbvie, Speakers bureau: Novartis, Lilly, Marcus Zimmermann-Rittereiser Shareholder of: Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Employee of: Siemens Healthcare GmbH. Siemens Healthcare GmbH is a medical technology company (NOT a pharmaceutical company)., Georg Schett Speakers bureau: AbbVie, BMS, Celgene, Janssen, Eli Lilly, Novartis, Roche and UCB, David Simon Grant/research support from: Else Kröner-Memorial Scholarship, Novartis, Consultant of: Novartis, Lilly