Using Iterative Causal Forest to Identify Heterogeneous Treatment Effects: Sodium-glucose Cotransporter-2 inhibitors versus Glucagon-like Peptide-1 Receptor Agonists on Hospitalization for Heart Failure

Background: Large trials have shown antidiabetic Sodium-glucose Cotransporter-2 inhibitors (SGLT2i) reduce the risk of hospitalization for Heart Failure (HHF), but it is unclear which patients are likely to benefit most.

Objectives: Implement iterative casual forest (iCF), a causal forest-based algorithm we developed for subgroup identification, to identify the additive interactions between SGLT2i and covariates, and subgroups with heterogeneous treatment effect (HTE) for HHF.

Methods: We implemented an active comparator, new user design identifying initiators of SGLT2i and Glucagon-like Peptide-1 Receptor Agonists (GLP1RA) using 20% random sample of fee-for-service Medicare beneficiaries aged 65+ with parts A (inpatient), B (outpatient physician services), and D (dispensed prescription drugs) coverage for ≥ 1 year from 1/1/2012 to 12/31/2015 and without prior end stage renal disease. We quantified treatment effect by the risk difference (RD) for HHF between SGLT2i and GLP1RA in initial treatment analysis over a maximum 2-year follow-up. Our cohort includes 10,343 SGLT2i initiators and 9,993 GLP1RA initiators and 380 (7.82%) were censored as they lost part A or B of Medicare coverage during follow-up, which was adjusted by inverse probability censoring weight (IPCW) based on baseline covariates. Death was treated as a competing risk by setting the risk for HHF after death to 0. We implement iCF with 2000 trees and 1000 iterations. In each identified heterogeneous group by iCF (if any), we first calculated the propensity score by logistic regression to adjust for confounding, then assessed conditional average treatment effects (CATEs) applying inverse probability treatment weights and IPCW.

Results: The aRD in the overall population was -0.8% (-1.4% to -0.2%) (crude risk 4.1% vs 6.6%). The P-value for homogeneity was 0.0003. The iCF identified subgroups defined by edema, peripheral vascular disease, loop diuretics, age 75+, and sex. Edema is one of the fundamental signs of heart failure (HF), the largest CATE (aRD of -3.9%, -6.0% to -1.9% and crude RD -5.7%, -7.8% to -3.6%) in subpopulation with edema (a proxy for HF) could be explained by SGLT2i's beneficial effect on HF.

Conclusions: Our finding is consistent with previous studies and shows our iCF algorithm can successfully identify heterogeneous subgroups and additive interactions. The iCF is a promising method to identify subgroups with HTE in large cohorts. The interpretation of the subgroups identified requires clinical and pharmacological input.