SAT0119 CREATE RISK PREDICTION MODELING AND DRUG WITHDRAW ROAD MAP THROUGH PATTERN EXTRACTION AND DATA MINING: A MASTER ALGORITHM DEVELOPMENT FROM THE SMART SYSTEM OF DISEASE MANAGEMENT (SSDM)

Background: Combination therapy with DMARDs for treating RA is standard of care. However, certain rates of adverse events (AEs) are unavoidable. The stigmas are how to predict the risk and how to define drug withdraw sequence if AEs persist for optimal risk reductions. The decisions made are always empirically. Objectives: To develop a risk prediction model and an algorithm for drug withdraw sequence based on data mining from the SSDM. Methods: SSDM is an interactive mobile disease management tool, including two application systems (APPs) for both the doctors and the patients. The patients can input medical records (including medication and laboratory test results) and perform self-evaluation (DAS28, HAQ) via App. The data synchronizes to mobiles of authorized rheumatologists through cloud and advices could be delivered. In previous studies, we demonstrated that patients could master SSDM after training. In order to develop a prediction model and the master algorithm, abnormal white blood cell counts (WBC) and alanine aminotransferase (ALT) elevation were targeted. Data was collected, extracted, validated, and Bayesian networking, data mining, modeling were performed. WBC under 4k/ml is defined as leukocytopenia (LP), over 10k/ml as infection predisposing (IP), and ALT > 40 U/L as ALT elevation. Results: From Jun 2014 to Jan 2019, 44,533 RA patients from 587 centers registered in SSDM. 135 different drugs and 882 combination therapies are identified. LP happens at 317 and IP at 286, ALT at 322 cases in 641 treatment regiments. Among them, MTX based regiments are 257 types, and the risk ratio (RR) are profiled as prediction model by comparing each AE rate of combination regiment with that of MTX monotherapy (Fig 1). The RR ranges from 0.28 to 6.28. The highest risk combination of prednisone (Pred), leflunomide (LEF), methotrexate (MTX), hydroxychloroquine (HCQ) and Celecoxib is selected (RR=6.28) to develop a master algorithm. Figure 2 shows Bayesian network, in which, quartet correlaties with 31 different regiments. Based on Bayesian method, the probabilities of LP, IP and ALT are plotted through 64 modeling, and the algorithm for drug withdraws strategies is generated. Drug withdrawing sequence for LP is HCQ, then Cel, then LEF, then Pre, the risks of LP are reduced by 41%, 22% 36% and 15%, respectively. For IP, withdraw sequence is Pred, then LEF, then Cel, then HCQ, the risks of IP are reduced by 45%, 28%, 23% and 4%, respectively, For ALT, withdraw sequence is MTX, then Pred, LEF, then Cel, the risks of ALT are reduced by 48%, 8%, 7%and 6%. Conclusion: Through patterns extraction, data mining, modeling, and Bayesian networking, a risk prediction model and a master algorithm for drug withdraw strategy in reduction of AEs are developed, which are expendable and replicatablei. Via continuing data inputs and machine leaning, an artificial intelligent system in assisting clinical forecast and decision-making may be achieved with SSDM. Disclosure of Interests: None declared