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

Area of Science:

• Health Informatics
• Machine Learning
• Real-World Evidence

Background:

• Predefined covariates in machine learning algorithms may miss crucial features for accurate subgroup identification.
• Identifying patient subgroups with heterogeneous treatment effects (HTEs) is vital for personalized medicine.

Purpose of the Study:

• To develop and validate a semi-automatic subgrouping algorithm, hdiCF, for enhanced feature recognition in claims data.
• To identify patient subgroups with HTEs for hospitalized heart failure incidence using the hdiCF algorithm.

Main Methods:

• The hdiCF algorithm uses high-dimensional feature identification from medical codes (diagnoses, procedures, prescriptions) and propensity score methodology.
• Features are created based on the frequency of occurrence, followed by propensity score trimming and preparation.
• Iterative causal forest (iCF) is implemented to identify subgroups with HTEs.

Main Results:

• Application of hdiCF in Medicare beneficiaries initiating SGLT2i or GLP-1 RAs identified subgroups with HTEs for hospitalized heart failure.
• Findings aligned with existing research, suggesting SGLT2i benefits patients with pre-existing heart failure or chronic kidney disease.
• The algorithm successfully identified subgroups with markers for potential HTEs without prior hypotheses.

Conclusions:

• The hdiCF algorithm offers a robust method for identifying HTEs in real-world evidence studies.
• It overcomes limitations of predefined covariates by adapting high-dimensional propensity score methods for feature discovery.
• hdiCF enhances the identification of patient subgroups for targeted therapies, particularly in settings with potential unmeasured confounding.