A Novel High Dimensional Approach to Assess Treatment Heterogeneity in Claims Data: Sodium-glucose Cotransporter-2 inhibitors (SGLT2i) vs Glucagon-like Peptide-1 Receptor Agonists (GLP1RA) on Hospitalized Heart Failure Risk

Background: We recently developed a machine-learning algorithm (MLA), iterative causal forest (iCF), to identify subgroups with heterogeneous treatment effects (HTE), and its high dimensional (HD) version, hdiCF, which is not dependent on the pre-specification of heterogeneity candidates. Whereas the HD propensity score (hdPS) generates 3 binary variables per code, low, medium, and high frequency of occurrence for confounding control, we proposed to create a single ordinal variable per code serving as potential splitters by frequency 0, 1, sporadically (2 to 75th percentile), or frequent (> 75th percentile) for hdiCF, resulting in explainable subgroup decisions.

Objectives: To apply the HD approach for hdiCF in an empirical example using various MLA to assess HTE.

Methods: We created a cohort of SGLT2i vs GLP1RA (8075 vs 7313) initiators from 20% random sample of fee-for-service Medicare beneficiaries 2015-2019 with parts A, B, D for ≥ 1 year and without chronic kidney disease stage 4 or 5, then identified 591 HD features (1 ordinal variable per code) by 3-digit ICD-10, 5-digit CPT, and 4th level ATC codes for top 200 most prevalent codes with a cutoff value of >1 & < 99% in each of the 5 dimensions (in/outpatient diagnosis & procedures, prescriptions) during 1-year baseline period. We ran 5 MLA including 3 tree-based methods, hdiCF, CausalTree (CT), aVirtualTwin (VT), as well as ROWSi and FindIt, to assess HTE. In each identified subgroup by tree-based MLA, we computed the PS by 80 predefined variables to assessed conditional average treatment effects (CATE) by the adjusted 2-year risk difference (aRD) for hospitalized heart failure (HHF) outcome using IPTW in initial treatment analysis.

Results: The aRD for SGLT2i vs. GLP1RA in the overall population is -0.5% (-1.1%, 0.2%) (crude risk 3.4% v 5.1%). The identified largest CATE subgroups, along with their sample size, crude risk, and aRD are: 1) hdiCF: 2+ loop diuretics (2786, 10.5% v 13.6%, -2.6% [-5.0%, -0.2%]); 2) CT: ischemic heart disease (IHD) (1141, 9.3% v 16.7%, -5.1% [-8.9%, -1.3%]); 3) VT: IHD and no other (unspecified) cognitive symptoms (884, 8.5% v 17.1%, -7.0% [-11.3%, -2.7%]). The strongest interactions identified are: 1) ROWSi: SGLT2i-correcting fluid or electrolyte balance; 2) FindIt: SGLT2i-Aphagia/dysphagia-cardiomyopathy.

Conclusions: Our HD setting helps to identify HTE without prior knowledge about the biology or healthcare system, and most of the HTE identified by various MLA aligns with clinical evidence that SGLT2i have a larger beneficial effect on HHF risk in patients with Cardiovascular-Kidney-Metabolic syndrome relative to GLP1RA. Careful interpretation of results is required and further studies are needed to test the performance of confounding control in various subgroups and settings.