POS0641 MACHINE LEARNING AND EXPLAINABLE AI METHODS CAN HELP TO PREDICT THE INEFFECTIVENESS OF INDIVIDUAL BIOLOGICAL DISEASE MODIFYING ANTIRHEUMATIC DRUGS (bDMARDS) – DATA FROM THE AUSTRIAN BIOLOGICAL REGISTRY BIOREG

Abstract

BackgroundThe continuous development of biological disease modifying antirheumatic drugs (bDMARDs) in recent years has significantly improved the treatment options for patients suffering from rheumatoid arthritis. Selecting the most effective biologic remains a challenge, since therapy-response is highly individual depending on the patient history, laboratory values and demographics.ObjectivesThe aim of this study was to investigate, if non-responders can be detected before therapy start using machine learning models and explainable artificial intelligence to provide a probability of non-response and to identify the most impactful contributing factors to the model output.MethodsData from the Austrian Registry for bDMARDs and tsDMARDs in Rheumatic Diseases – BIOREG were obtained. BIOREG provides a real-world data set, which covers rheumatology hospitals and practices throughout Austria. According to EULAR-guidelines the observation time window for treatment response is 6 months and the prediction time horizon was set at 6 months as well.Different machine learning models were trained for Abatacept (ABA), Adalimumab (ADA), Certolizumab (CERT), Etanercept (ETA) and Tocilizumab (TOC) to predict the risk of non-response per treat to target (ttt)-course. Nested cross-validation and hyperparameter tuning (iteration over fixed parameter grid) were applied. To evaluate the prognostic quality per model the area under the receiver operating characteristic (AUROC) was collected and the model with the highest score was selected for further evaluation. By applying the Explainable AI method SHAP (SHapley Additive exPlanations; a game-theoretic approach to evaluate variable importance) to each final model, the most contributing factors and direction of impact were evaluated.ResultsData from 1397 patients, 2004 (baseline) visits and 22 variables (19 after cleaning) with at least 100 ttt-courses per drug were included in the study.The best models per biologic achieved an AUROC-score of: CERT: 0.76 (95% CI, 0.67–0.86). TOC: 0.72 (95% CI, 0.69–0.79), ABA: 0.71 (95% CI, 0.65–0.77), ADA: 0.67 (95% CI, 0.62–0.76), ETA: 0.68 (95% CI, 0.53–0.85).The explainable AI interpreted visual analytic scores (VAS) as most important variables for ABA, ETA and TOC. High scores were associated with high risk of non-response for these drugs. For ADA, co-therapy with glucocorticoids was the most important and risk-increasing factor. For CERT, the dosage of the prescribed drug was ranked as the most influential variable; high dosages were associated with lower risk of non-response. Interestingly, some variables displayed opposite impacts in different drugs: Male gender was interpreted as risk-increasing for ABA and risk-decreasing for ETA. Moreover, negative rheumatoid-factor was interpreted as risk-decreasing for ABA/ETA, but risk-increasing for ADA/CERT.ConclusionThe results of our study show that non-responders of biological drugs can be detected with moderate to even good prognostic qualitybeforestarting a ttt-course, comparable to similar research with different prediction time horizons [1].The opposite impact of some variables in different bDMARDs as well as the difference in variable importance per bDMARD indicate, that selecting the right drug is highly dependent on the individual patient characteristic. Machine Learning could be of additional support for rheumatologists and patients by providing not only a prediction of ineffectiveness per drug, but also an explanation for the prediction.Reference[1]Koo, B.S., Eun, S., Shin, K. et al. Machine learning model for identifying important clinical features for predicting remission in patients with rheumatoid arthritis treated with biologics. Arthritis Res Ther 23, 178 (2021).https://doi.org/10.1186/s13075-021-02567-yAcknowledgements:NIL.Disclosure of InterestsDubravka Ukalovic Employee of: Siemens Healthineers. Siemens Healthineers is a medical technology company (NOT a pharmaceutical company), Burkhard Leeb Speakers bureau: AbbVie, Roche, MSD, Pfizer, Actiopharm, Boehringer-Ingelheim, Kwizda, Celgene, Sandoz, Grünenthal, Eli-Lilly, Consultant of: AbbVie, Amgen, Roche, MSD, Pfizer, Celgene, Grünenthal, Kwizda, Eli-Lilly, Novartis, Sandoz, Grant/research support from: TRB, Roche, Bernhard Rintelen Speakers bureau: BMS, Eli-Lilly, Pfizer, TRB-Chemedica, UCB, Wyeth, Consultant of: Abbott, Abbvie, Amgen, Gileat, Novartis, Pfizer, Roche, TRB-Chemedica, UCB, Wyeth, Grant/research support from: Abbott, Aesca, Amgen, Centocor, Eli-Lilly, Servier, UCB, Gabriela Eichbauer-Sturm Speakers bureau: AbbVie, Astro-Pharma, Grünenthal, Jansen, Eli-Lilly, Menarini, MSD, Novartis, Pfizer, Roche, TRB, UCB, Fresenius Kabi, Peter Spellitz: None declared, Rudolf Puchner Speakers bureau: AbbVie, BMS, Janssen, Kwizda, MSD, Pfizer, Celgene, Grünenthal, Eli-Lilly, Consultant of: AbbVie, Amgen, Pfizer, Celgene, Grünenthal, Eli-Lilly, Manfred Herold: None declared, Miriam Stetter: None declared, Vera Ferincz: None declared, Johannes Resch-Passini: None declared, Marcus Zimmermann-Rittereiser Employee of: Siemens Healthineers. Siemens Healthineers is a medical technology company (NOT a pharmaceutical company), Ruth Fritsch-Stork Speakers bureau: AbbVie, Astra Zeneca, Astropharm, Novartis.