Alejandro Schuler, Graduate Student
Shah Lab
Stanford University

Abstract:
Hypertension (high blood pressure) is a hugely prevalent risk factor for cardiovascular diseases. Several drug classes effectively treat hypertension, but there is evidence that particular patient groups may respond favorably only to some of these drug classes. We sought to discover and characterize patient groups that respond more favorably to ACE-inhibitors or beta-blockers (AB) than to calcium channel blockers and diuretics (CD). Using data from Stanford’s electronic health record (EHR), we built and new-user cohort of hypertensive patients who were prescribed AB or CD drugs. We matched patients from the two treatment groups using a regularized regression propensity score model trained with over 6,000 clinical features. We discovered that use of the one-standard-error rule to select the propensity score model dramatically improved covariate balance in the matched cohort, contrary to colloquial advice that propensity score models should be overfit. Using the matched cohort, we trained a gradient boosted tree model to predict each patient’s post-treatment blood pressure drop given the observed treatment (AB or CD) and the same 6,000 features. We used this model to estimate the counterfactual outcomes for each patient and found no estimated treatment differences for any patient groups.