OBJECTIVES : Value assessment of treatments is critical for determining coverage policy and patient / provider choice of treatment. Such assessment requires reliably estimating treatment effect across multiple outcomes and in diverse sub populations based on clinical and demographic characteristics.

METHODS : We developed an end-to-end modular framework for discovering sub-populations of interest from retrospective real-world data using causal analysis. For a target disease or condition, we (1) Extract patient features and outcomes from a health records database; (2) Train causal analysis models of choice to predict the counterfactual outcomes for each patient in each optional treatment; (3) Evaluate and optimize each model for accuracy and generalizability; and (4) Identify and characterize sub-populations satisfying context-specific treatment response criteria.

RESULTS : We applied and evaluated the framework over a rheumatoid arthritis cohort extracted from IBM Truven Marketscan® Research Database, consisting of 23,100 patients who received one of 9 biologic drugs during 2010-2016. We examined 27 clinical, utilization and cost outcomes, aggregated over 12 months following treatment initiation, and 240 demographic and clinical covariates extracted from a 12 months baseline period. Doubly robust models were trained and evaluated for all treatment-outcome pairs. We systematically extracted sub-populations for two criteria: (a) Stratification by clinical or demographic characteristics (e.g. age, ethnicity, comorbidities) to differentiate treatment effect across multiple outcomes, and (b) Identification and characterization of the best responders for each treatment.

CONCLUSIONS : Our framework enables systematic comparison of competing treatments across multiple outcomes, discovering and characterizing sub-populations who might benefit from specific treatments and thus supports decision making related to matching optimal treatment to patients.