# Physical Activity Recommendations Tailored by a Predictive Model for Adults With High Blood Pressure: Observational Study

## TL;DR

This study shows that personalized physical activity recommendations for people with high blood pressure can improve survival by tailoring patterns to individual characteristics.

## Contribution

A machine learning model predicts optimal physical activity patterns for adults with high blood pressure based on individual characteristics.

## Key findings

- The model predicted active regular, active light, and active weekend warrior patterns as optimal for different individuals.
- Inconsistent physical activity patterns with predicted optimal ones increased all-cause mortality risk by 28%.
- Stroke history, age, sex, BP class, and medication were key factors in determining optimal patterns.

## Abstract

Whether the benefits of identical physical activity (PA) patterns for adults with high blood pressure (BP) vary according to an individual’s characteristics has not been adequately studied.

This study aimed to investigate whether an individual’s characteristics modify the associations between PA patterns and mortality rate.

Four PA patterns were derived from accelerometer-based data: active weekend warrior, active regular, active light PA, and baseline PA. The main outcome was all-cause mortality. A machine learning model to predict the optimal PA pattern for individual patients was trained in the UK Biobank (UKB) cohort and externally validated in the National Health and Nutrition Examination Survey cohort, which was subsequently integrated into a web-based application. The potentially optimal PA pattern within patients was identified as the one leading to the highest predicted survival probability. Multivariable Cox models were used to estimate hazard ratios and 95% CIs for all-cause mortality corresponding to the inconsistency of the current PA pattern with the predicted optimal PA pattern.

A total of 71,637 UKB adults and 5104 National Health and Nutrition Examination Survey individuals were enrolled. External validation demonstrated that the area under the receiver operating characteristic curve of our model for predicting mortality at 10 years of follow-up was 86.4% (95% CI 85.1%‐87.7%). The predicted optimal PA patterns in the UKB cohort were active regular PA for 26,643 (37.2%) participants, active light PA for 22,606 (31.6%) participants, and active weekend warrior for 21,749 (30.4%) participants. Stroke history, age, sex, BP class, and antihypertension medication were key factors driving heterogeneity in individuals’ optimal PA patterns. Cox regression analysis suggested that individuals in the UKB cohort whose current PA patterns were inconsistent with the predicted optimal patterns may be associated with a 28% increase in all-cause mortality risk on average (hazard ratio 1.28, 95% CI 1.20‐1.38) compared to those with consistent patterns.

Our findings may help patients with high BP obtain individualized recommendations for PA patterns based on their specific characteristics, thereby improving their prognosis.