POS1560 USING SURVIVAL CLASSIFICATION AND REGRESSION TREES AND SURVIVAL RANDOM FOREST TO INVESTIGATE ASSOCIATIONS AMONG FACTORS FOR CANCER PROGRESSION AND ARTHRITIS CONTROL IN A MULTI-CENTER RETROSPECTIVE COHORT OF PATIENTS WITH IMMUNE CHECKPOINT INHIBITOR ASSOCIATED INFLAMMATORY ARTHRITIS

Abstract

BackgroundIt is unknown whether particular presentations of immune checkpoint inhibitor-associated inflammatory arthritis (ICI-IA) are associated with better cancer responses or faster time to arthritis control. Machine learning methods have the ability to determine important factors in datasets where there are often many variables.ObjectivesTo identify variables associated with arthritis control and cancer progression among persons with ICI-IA.MethodsThis study is ancillary to a retrospective observational study of 147 DMARD-treated (TNFi, IL-6Ri, or methotrexate) patients with ICI-IA seen at 6 U.S sites. Variables from the medical record included demographics, patterns of swollen joints, medications, lab values, and concomitant irAEs. Arthritis control and cancer progression were the two outcomes analyzed. survival Classification and Regression trees (sCART) were created using the rpart and partykit R packages[1]. Random Survival Forest (RSF) was performed using the R package randomForestSRC[2]. Variable importance (VI) for sCART and Relative Importance Score (RIS) were computed to assess influential variables..ResultsWe analyzed 147 people with ICI-IA. 69% received PD-1/PD-L1 monotherapy, and 43% had melanoma. ICI-A treatment was a TNFi in 17%, IL6Ri in 26%, methotrexate in 33%, and 24% received >1 DMARD. Median time to cancer progression was 333 (IQR 110, 811) days for the 26% that progressed. Median time to arthritis control was 109 days (IQR 32, 287) for the 93% that achieved control.For cancer progression the following were identified by both sCART and RSF as important: steroid duration, total joint count (TJC), study site, maximum steroid dose, ICI type, shoulder arthritis, >1 DMARD and number of IRAEs. For classifying arthritis control, the following variables were found to be important in both sCART and RSF: steroid duration, >1 DMARD, elbow arthritis, age, cancer type, TJC and first DMARD (Table 1). The Figure 1 shows the sCART for arthritis control.ConclusionBoth methods, SCART and RSF, demonstrated the important influence of steroid duration on arthritis control and cancer progression. Machine learning methods demonstrated the potential prognostic importance of specific joint involvement for each outcome- knee for time to arthritis control and shoulder and wrist for cancer progression.