High-dimensional Iterative Causal Forest (hdiCF) for Subgroup Identification Using Health Care Claims Data

Area of Science:

• Pharmacovigilance and Pharmacoepidemiology
• Biostatistics and Health Data Science
• Cardiovascular Disease Research

Background:

• Identifying patient subgroups that benefit differently from medications (heterogeneous treatment effects, HTE) is crucial for personalized medicine.
• Standard high-dimensional propensity score (hdPS) methods face challenges in accurately capturing complex patient characteristics.
• Novel high-dimensional approaches are needed to improve the detection of HTE in real-world data.

Purpose of the Study:

• To compare a novel high-dimensional approach with the standard hdPS method for detecting HTE.
• To identify subgroups of patients experiencing different treatment effects from sodium-glucose cotransporter-2 (SGLT2) inhibitors and glucagon-like peptide-1 receptor agonists (GLP-1 RAs) regarding heart failure risk.
• To assess the performance of these methods in a large Medicare cohort.

Main Methods:

• A novel high-dimensional approach using ordinal variables was developed and compared against the standard hdPS method (binary variables).
• The iterative causal forest (iCF) subgrouping algorithm was employed on a Medicare cohort (2015-2019) of SGLT2 inhibitors (N=8,075) and GLP-1 RAs (N=7,313).
• Conditional average treatment effects (CATEs) for 2-year risk differences in hospitalized heart failure were estimated using inverse-probability treatment weighting.

Main Results:

• The novel high-dimensional approach identified patients with ≥2 loop diuretic prescriptions as a subgroup with the largest CATE for reduced heart failure risk (aRD: -2.6%).
• The standard hdPS method identified patients with chronic kidney disease as a subgroup with a smaller CATE (aRD: -1.7%).
• Sensitivity analyses confirmed the novel approach's superior accuracy in identifying clinically relevant subgroups with HTE.

Conclusions:

• The novel high-dimensional method demonstrates enhanced capability in detecting HTE compared to the standard hdPS approach.
• This improved detection can lead to more precise identification of patient subgroups benefiting from SGLT2 inhibitors and GLP-1 RAs.
• The findings support the clinical relevance of identifying specific patient characteristics, such as loop diuretic use, for optimizing heart failure risk management.