Development and Internal Validation of Machine-Learning Models for Short-Term Prediction of Day-to-Day Home Blood...

Day-to-day home blood pressure variability (BPV) is associated with cardiovascular risk and influenced by environmental conditions. However, it is unclear whether short-term increases in day-to-day BPV can be predicted from personal sensor data. In this study, our aim is to develop and validate a machine-learning prediction model for short-term increases in day-to-day BPV using personal sensor data on behavioral and environmental exposure.

We will conduct a 30-day monitoring study in community-dwelling adults. Participants will measure home BP twice daily, while a portable sensor and an activity tracker record environmental conditions and physical activity. The primary outcome is an episode of increased systolic day-to-day BPV, defined as a rolling 5-day coefficient of variation ≥11.0%. Candidate predictors will be derived from the preceding 5-day exposure window. We will construct window-level data, allocate participants to training and test sets, and train machine-learning models with participant-level cross-validation. We will evaluate performance using the area under the receiver operating characteristic curve, calibration, Brier score, and decision-curve analysis, and interpret the XGBoost model with Shapley additive explanations to quantify the predictor contributions.

This protocol outlines a framework for predicting short-term increases in day-to-day BPV from personally experienced environmental exposure and behaviors, supporting future personalized interventions targeting modifiable environmental and behavioral factors.