(A-172) Iterative Causal Survival Forest: A Novel Subgrouping Algorithm for Survival Data Allowing for Censoring

Background: Our recently developed iterative causal forest (iCF) and its high-dimensional (HD) version (hdiCF) machine learning algorithms could identify subgroups with heterogeneous treatment effects (HTEs) using predefined or HD variables from claims. However, they do not support survival data or account for censoring due to disenrollment when splitting population.

Objectives: Develop a new subgrouping algorithm, iCSF, to handle informative censoring in survival data.

Methods: Based on the iCF framework, we replaced causal forest with causal survival forest (CSF, which estimate HTE in survival data by survival time or survival probability (i.e., 1-risk) to develop iCSF in 3 steps: at each depth (2, 3, 4, 5), 1) iteratively grows CSF and votes to select the most stable best tree (best representing HTE signal) structure from each CSF as a subgroup (SG) decision; 2) builds models to predict treatment effect (via doubly robust estimation of survival probability as the transformed outcome, accounting for informative censoring) using SGs obtained; 3) finds the cross-validated SG that best predicts the treatment effect as final SG. We included 8075 sodium-glucose cotransporter-2 (SGLT2i) initiators and 7313 glucagon-like peptide-1 receptor agonists initiators from 4/1/2013 to 12/31/2015 after a washout of 12 months with continuous parts A, B, and D coverage and without severe renal disease from a 20% random sample of fee-for-service Medicare beneficiaries. We followed the cohort until the hospitalized heart failure (HHF) outcome, Part A/B disenrollment, 12/31/2017, death, or a maximum of 2-years. We ran iCSF (1000 trees, 100 iterations) to obtain SGs. In each identified group, we computed the inverse probability treatment and censoring weight by 80 predefined covariates to assess conditional average treatment effects (CATEs) by the adjusted risk difference (aRD%) at 2-year for HHF in intention-to-treat analyses. Death was handled as a censoring event in iCSF and competing event when estimating aRDs.

Results: At 2-year, the aRD% in the overall population was -0.4 (-1.1, 0.2) (crude RD% -1.8 (-2.4 to -1.1), risk% 3.4 v 5.1, censoring rate% 1.5 v 1.4). The iCSF identified subgroups defined by loop diuretics and lipid test. Loop diuretics is a key treatment for heart failure (HF), the largest CATE (N 1416 vs 1938, aRD% -2.1 [-4.3, 0], crude RD% -2.4 [-4.6, 0.3], crude risk% 10.2 v 12.8, censoring rate% 1.0 v 1.4) in subpopulation with loop diuretics (a proxy for HF) could be explained by SGLT2i’s beneficial effect in patients with HF.

Conclusions: Our finding is consistent with previous studies and shows iCSF can successfully identify subgroups with HTE in survival data. Further studies are needed to test the performance of iCSF in as-treated analyses or settings with different censoring.