Recommendations on antihypertensive drugs to decrease pulse wave velocity in hypertensive patients based on machine learning algorithms: RIGIPREV Study (Preprint)

Abstract

High systolic blood pressure is a leading global risk factor for attributable deaths, but many patients with hypertension do not achieve adequate blood pressure control despite treatment. Arterial stiffness, measured by pulse wave velocity (PWV), is an independent cardiovascular risk factor. Different antihypertensive drugs have varying effects on PWV, but it is unclear if these effects depend on individual patient characteristics.

to develop a model to provide recommendations on the most suitable antihypertensive agent for reducing PWV, based on the individual characteristics of person with hypertension, using advanced ML techniques consisting of multiple Random Forest models within a multi-output regressor approach.

This study, known as the RIGIPREV study, utilized data from the EVA, LOD-DIABETES, and EVIDENT studies, involving hypertensive individuals with baseline and follow-up measurements. Ethical approval was obtained, and subjects provided informed consent. Antihypertensive drugs were categorized into groups, and primary outcome measures included carotid femoral and brachial-ankle PWV. Various covariates were collected, such as demographic factors, lifestyle characteristics, medication usage, anthropometric measurements, cardiovascular parameters, and biochemical variables. Data preprocessing addressed missing values, and a multi-output regressor with six Random Forest models was used for predicting medication effectiveness in decreasing PWV. The performance of these models was assessed using the Coefficient of Determination (R2) and Mean Square Error (MSE). A recommendation system was designed to provide personalized treatment optimization.

The predictive models exhibited high R2 values ranging from 0.78 to 0.81, with low MSE values. Variable importance analysis revealed distinct patterns for each medication model. The recommendation system demonstrated significant variability in medication recommendations compared to patients' original medications, with a matching rate of 55.3% for ARBs. A decision tree achieved an overall accuracy of 91.75% in selecting suitable antihypertensive drugs based on patient characteristics. For instance, ARBs and ACEIs were top choices based on HbA1c, with further selections based on insulin levels for ACEIs and serum creatinine levels for ARBs.

This study provides personalized recommendations for selecting the most effective antihypertensive drugs to reduce arterial stiffness based on individual patient characteristics. These recommendations can assist physicians in optimizing hypertension treatment, considering both blood pressure reduction and arterial stiffness improvement. The results underscore the importance of tailoring drug selection to individual patient profiles. However, limitations include the omission of certain variables and potential sample bias. Future studies with larger and more diverse samples are needed to validate these models' generalizability.