OBJECTIVES:

Assessing variation in treatment effects, is vitally important to achieving the vision of personalized medicine and patient-centric outcomes, wherein treatment is tailored to achieve the best response and safety margin for an individual patient’s care¹. It has also been identified as an important factor in assessing reliability and fitness-for-use of real-world evidence (RWE) observational research². In their framework for RWE reliability, Mahendraratnam, et. al., stress the importance of standardized verification checks for heterogeneity as a first step to assessing reliability across data sources.³ In this study we evaluate a Bayesian heterogeneity of treatment effect (HTE) assessment⁴ in a Congestive Heart Failure population (CHF), versus frequentist sub-group analysis.

METHODS:

Analysis was conducted in a nationally representative linked EMR-Claims dataset. Data was converted into OMOP CDM version 5. Eligibility criteria were two (2) CHF diagnoses (ICD-10 I50.x) with at least 6 months pre-index and 1-year post-index continuous activity. Outcomes assessed were time to first instance of hospitalization as well as myocardial infarction 1-year from baseline. Bayesian HTE was conducted with various models, including regression, basic shrinkage, Dixon and Simon⁵, and its extended version using the beanz package⁶. The goodness-of-fit measures were employed to determine model selection. Results were compared with the frequentist approach of subgroup analysis.

RESULTS:

Variation in information criteria was observed across the Bayesian models explored. Basic shrinkage model appeared to fit better than simple regression, followed by Dixon and Simon then its extended version. Despite that fact, interpretability and model complexity must also be considered in model selection. Bayesian approach was shown to address the gap in the traditional approach by accounting the treatment effects on across a population instead of exclusively to a given subgroup.

CONCLUSIONS:

Bayesian HTE shows potential advantages over frequentist subgroup analysis for both patient-centric outcomes analysis as well as ascertaining reliability and generalizability of secondary observational research.