Abstract

Hypertension is a leading cause of cardiovascular disease and all-cause mortality. Although effective anti-hypertensive drugs (AHD) are available, many patients do not respond to treatment. Multiple molecular actors are involved in the short- and long-term blood pressure (BP) regulation, with kidney and adrenal glands playing a major role.

In the kidney, renin activity initiates a cascade of reactions that produces angiotensin II, which in turn stimulates the adrenal gland to produce the corticosteroid hormone aldosterone. Renin-Angiotensin-Aldosterone System (RAAS)-targeting drugs are a cornerstone in the treatment of hypertension and related cardiovascular conditions. Besides aldosterone, other biologically connected corticosteroids have been studied individually with respect to BP regulation, but joint modelling of the entire system in a unified manner is lacking to date. Modern techniques allow the simultaneous and accurate quantification of both RAAS molecules and corticosteroids in large samples within a short time frame. This provides a valuable resource to characterize the underlying biology of BP control as well as response to AHD treatment.

In this talk we present a causal learning approach to study adrenal gland metabolism alterations and their downstream effect leading to hypertension, using data from the Cooperative Health Research in South Tyrol (CHRIS) study in Italy. Analytic challenges include measurement instrument limitations, high dimensionality of the predictors, missingness, non-normality, and complex correlations between multiple biomarkers that represent biological chains of events. Our goal is to develop a comprehensive, biologically interpretable prediction algorithm to define the main and conditional effects of RAAS and corticosteroids and identify causal pathways regulating blood pressure according to AHD intake, while controlling for environmental, behavioral, biological and genetic variables that could play either an agonist or a confounding role.

We propose a causal counterfactual framework to estimate the average treatment effect (ATE) using two approaches: (1) unadjusted, defined as the observed difference in mean systolic (SBP) or diastolic BP (DBP) between treatment groups without adjusting for patient characteristics, and (2) Targeted Maximum Likelihood Estimation (TMLE), defined as the covariate-adjusted mean difference estimated using outcome and treatment-assignment models to account for measured confounders. Additionally, we develop a causal tree-based analysis to investigate treatment-effect heterogeneity by identifying participant subgroups with potentially different responses to AHD treatments. Within each identified subgroup, the conditional average treatment effect (CATE) is estimated using TMLE to quantify subgroup-specific differences in mean SBP or DBP between treatment groups.

Finally, we propose an extension of the above methodology to jointly model SBP and DBP within a multivariate causal inference framework, accounting for potential correlation between outcomes. This approach may provide a more comprehensive assessment of AHD treatment effects and their heterogeneity across patient subgroups.