OP0023 PREDICTION OF REMISSION FOR EACH BIOLOGICS BASED ON PATIENT’S CLINICAL INFORMATION BEFORE STARTING BIOLOGICS USING EXPLAINABLE ARTIFICIAL INTELLIGENCE: DATA FROM THE KOREAN COLLEGE OF RHEUMATOLOGY BIOLOGICS REGISTRY

Abstract

Background:Many studies have identified predictive factors of response to biologics in patients wirh rheumatoid arthritis (RA). However, there is still a lack in using them in daily clinical practice. Therefore, it is necessary to develop a method that can assist the physician in selecting effective biologics.Objectives:The purpose of this study is to establish machine learning model that predicts remission in patients treated with biologics using data of RA patients from the Korean College of Rheumatology Biologics (KOBIO) registry, and to identify the important features that have the most influence on the response to biologics using explainable artificial intelligence (AI).Methods:A total of 1,527 patients who started with biologics such as etanercept, adalimumab, golimumab, infliximab, abatacept, and tocilizumab from December 2012 to June 2019 were enrolled. Remission was predicted using 46 variables corresponding to baseline profiles at the starting of each biologics. We used five machine learning methods such as lasso, ridge, SVM, random forest, and XGBoost. For explainability of those models, we used Shapley plot to interpret the feature importance for each biologics.Results:In all machine learning methods, the accuracy and the area under the receiver operating characteristic (AUROC) were 57.2%~74.5%, 0.547~0.747, respectively (Table 1). The accuracy and AUROC of each biologics were similar between machine learning methods. Figure 2 showed interpretation of feature importance with the Shapley plot for remission. The most important feature was age in adalimumab (younger were closer to remission), daily corticosteroid dose in etanercept, golimumab, and all TNF inhibitors (using fewer doses daily were closer to remission), baseline erythrocyte sedimentation rate in infliximab (lower ESR were closer to remission), disease duration in abatacept (longer disease durations showed difficulty determining remission), baseline c-reactive protein in tocilizumab (higher CRP were closer to remission).Table.Predicting remission for all biologics in various machine learning method.MeasureLassoRidgeSVMRandom ForestXGBoostNo info rateSampleAbataceptAccuracy74.1%74.1%70.6%71.8%68.8%70.6%216AUROC0.7250.7420.7070.6770.6470.500AdalimumabAccuracy73.6%72.0%70.4%72.0%70.4%68.8%315AUROC0.7100.7290.7000.6750.6630.500EtanerceptAccuracy72.0%72.0%70.0%71.5%70.0%68.0%250AUROC0.7410.7470.7260.7190.7040.500GolimumabAccuracy71.3%68.5%66.7%68.5%68.5%68.5%138AUROC0.7460.7270.7010.6900.6550.500InfliximabAccuracy72.8%73.5%67.6%73.5%69.1%72.5%172AUROC0.6630.6830.6160.5970.5270.500TNF inhibitorsAccuracy73.9%74.5%73.9%74.2%73.6%70.3%875AUROC0.7390.7410.7260.7470.7240.500TocilizumabAccuracy62.4%63.6%62.4%59.5%57.2%59.5%436AUROC0.6330.6400.6330.6150.5470.500Figure 2.Shapley plots and SHAP values for the feature importance from clinical information in patients with RA.Conclusion:We developed machine learning models for predicting remission as a response to each biologics in active RA patients based on their clinical profiles, and found important clinical features using explainable AI. This approach may support clinical decisions to improve treatment outcomes in patients with RA.Disclosure of Interests:None declared