POS1415 A MACHINE LEARNING ANALYSIS OF FACTORS PREDICTING ORGAN DAMAGE PROGRESSION IN PATIENTS WITH SYSTEMIC LUPUS ERYTHEMATOSUS USING THE SPANISH SOCIETY OF RHEUMATOLOGY LUPUS REGISTRY (RELESSER)

Abstract

BackgroundEvidence shows that around half of the patients with systemic lupus erythematosus (SLE) develop irreversible organ damage due to the disease itself or to other factors (e.g., steroid treatment). It is essential to have a comprehensive understanding of factors that predict organ damage progression to identify at-risk patients and inform clinical decision making.ObjectivesTo develop an algorithm, using machine learning (ML) methodology, that predicts organ damage progression in SLE patients.MethodsThe Spanish Society of Rheumatology Lupus Registry (RELESSER) with patient records from 45 Rheumatology Units across Spain was used. RELESSER data were collected from 2011 to 2021 and captured demographic and comprehensive clinical information. In this analysis, a sample of 2,676 patients was used. The Systemic Lupus International Collaborating Clinics/American College of Rheumatology Damage Index (SDI) was used to measure organ damage progression between 2015 (start of the prospective data collection in RELESSER) and 2020.To predict the risk of an increase in the SDI, 102 variables were identified as potential predictors. A ML model (gradient boosting trees) was developed and validated by a simple logistic regression (LG) model. The area under the receiver operating characteristic curve (AUCROC) was used to quantify the improvement over random chance (an AUCROC of 0.5). Shapley Additive Explanation (SHAP) values were used in the ML model to identify predictors and their contribution to damage progression.ResultsOf all patients, 13% experienced organ damage progression, with 2-year patient follow-up. The ML algorithm was better at identifying these patients (AUCROC 0.68) than the LG model (AUCROC 0.63) (Figure 1). ML model performance can be contextualized using a random sample of 100 SLE patients of whom 13 suffered organ damage progression, the model would successfully identify 12. However, 66 additional patients would be incorrectly identified (True Positive 90%; False Positive 79%). The top 5 predictors of damage progression, across all patients, were patient age >49 years,Figure 1.ROC AUC plotCreatinine > 0.9 mg/dl, cardiovascular-related disease complications (> 3 complications), low hematocrit level recorded recently (<41.9 months ago), and triglycerides > 81.5 mg/dl. At the patient level, the 5 patients with the highest predicted risk had strong predictors of progression, with key predictors being patient age at study entry >49 years, age at diagnosis >40 years, cardiovascular-related disease complications, and increased creatinine >0.9 mg/dl. The 5 patients with the lowest predicted risk had strong predictors of no change with key predictors being not having a low hematocrit level recorded recently, triglycerides <81.5 mg/dl and patient age at study entry <49 years.ConclusionWe developed a machine learning model, using an exhaustive set of variables in RELESSER which successfully predicted short-term organ damage progression in SLE patients, and outperformed a standard regression model. If the model were to be used as a clinical tool, only light-touch interventions should be carried out due to high false positive rate. Further model optimizations, including exposing the model to longer follow-up data and testing it in non-Spanish patients is needed.Disclosure of InterestsJose M Pego-Reigosa Speakers bureau: Eli Lilly, GSK, Consultant of: AstraZeneca, GSK, MSD, Eli Lilly, Grant/research support from: The RELESSER registry was funded by grants from GSK, UCB, Roche and Novartis, WALID FAKHOURI Shareholder of: Eli Lilly and Company, Employee of: Eli Lilly and Company, Silvia Díaz-Cerezo Shareholder of: Eli Lilly and Company, Employee of: Eli Lilly and Company, Aidan Cooper Consultant of: Salaried employee of IQVIA while conducting the analyses for this study, Anne-Marie Saunders Consultant of: Salaried employee of IQVIA while conducting the analyses for this study, Grace Segall Consultant of: Salaried employee of IQVIA throughout conceptualization, design and execution of analysis for this study, Employee of: Eli Lilly and Company throughout finalization of analysis and abstract development for this study, Christophe Sapin Shareholder of: Eli Lilly and Company, Employee of: Eli Lilly and Company, Sebastian Moyano Employee of: Eli Lilly and Company, Iñigo Rua-Figueroa Grant/research support from: The RELESSER registry was funded by grants from GSK, UCB, Roche and Novartis