# A simulation-based framework for modeling and prediction of personalized blood pressure trajectories in hypertensive patients after antihypertensive treatment

## TL;DR

This paper introduces a simulation framework to model and predict personalized blood pressure changes in hypertensive patients after treatment, enabling earlier therapy adjustments.

## Contribution

A novel simulation framework using Pharmacokinetic-Pharmacodynamic modeling and two predictive models for blood pressure steady-state prediction.

## Key findings

- The non-linear mixed effect model outperforms the Gaussian process in reducing RMSE and bias in noisy data.
- The Gaussian process is more suitable when only one day of measurements is available.
- The framework incorporates individual daily rhythms, patient characteristics, and medication effects.

## Abstract

Hypertension, a leading global cause of death, poses challenges in stabilizing blood pressure within target values despite various therapeutic options, often necessitating multiple therapy adjustments and delayed impact assessments. Recently, the first wrist-based wearable blood pressure measurement devices were introduced which allow for a continuous assessment of blood pressure trajectories. This enables the development of statistical methodology for prediction of saturated steady-state of blood pressure under treatment—and thus allowing physicians to adjust the therapy earlier. As a prerequisite for the evaluation of such models and algorithms, it is necessary to simulate reliable and realistic hypothetical patient trajectories under treatment with antihypertensive medication. In this paper, we propose a simulation framework for blood pressure profiles through Pharmacokinetic-Pharmacodynamic modeling, which incorporates individual daily rhythms, patient characteristics, and medication effects. We also propose and evaluate two models for steady-state prediction under antihypertensive therapy, a Gaussian process and a non-linear mixed effect model. When only one day of measurements is available, the Gaussian process is preferred, but in real-world situations with more data, the non-linear mixed effect model is favored. It effectively reduces RMSE and bias in noisy data, outperforming the Gaussian process regardless of sample size.