Offline and Online Learning Techniques for Personalized Blood Pressure Prediction and Health Behavior Recommendations

Po-Han Chiang,Sujit Dey

doi:10.1109/access.2019.2939218

Po-Han Chiang, Sujit Dey

Open Access

https://doi.org/10.1109/access.2019.2939218

Copy DOI

Abstract

Blood pressure (BP) is one of the essential indicators of human health and highly correlated to health behavior (e.g., exercise and sleep). However, the degree of impact of each health behavior on BP is unknown and may vary significantly between individuals. In this paper, we investigate the relationship between BP and health behavior using data collected from off-the-shelf wearable devices and wireless home BP monitors. We propose a personalized BP model based on random forest (RF), which can predict individual's BP using health behavior and historical BP, and identify the most important factors in predicting an individual's BP. The latter can be used to provide personalized health behavior recommendations to improve and manage BP. We propose RF with Feature Selection (RFFS), which performs RF-based feature selection to enhance the prediction. Furthermore, since BP and health behavior data are collected and learned sequentially, the performance of prediction is prone to the existence of concept drifts and anomaly points. To solve this problem, we propose an Online Weighted-Resampling (OWR) technique to enhance RFFS in an online learning scenario. To show the effectiveness of RFFS and OWR, we use existing machine learning methods on the proposed dataset as a comparison. Our experimental results show that the proposed approach achieves the lowest prediction error. We also validate the effectiveness of personalized recommendations based on feature importance, which influences the user to change lifestyle to improve his/her BP.

Highlights

Hypertension, or high blood pressure (BP), one of the most prevalent chronic disease in the world, affects 30% of American adults and contributes to over 410,000 deaths per year [1]
To extend RF with Feature Selection (RFFS) with sequential and limited data examples, we propose a novel Online Weighted-Resampling (OWR) technique to alleviate the negative effect of concept drifts and anomaly points and thereby decrease the necessary time for data collection
We will validate the effectiveness of personalized recommendations of health behavior generated by our personalized BP model using RFFS

Summary

Introduction

Hypertension, or high blood pressure (BP), one of the most prevalent chronic disease in the world, affects 30% of American adults and contributes to over 410,000 deaths per year [1]. The relationship between health behavior and BP is only studied through clinical trials in ambulatory settings, whose scope is limited in terms of trial population and duration. Wearables such as Fitbit, Apple Watch, and Samsung Galaxy Watch collect a large amount of high-granularity health behavior data such as the duration and quality of activities and sleep. The primary metrics used to measure BP are systolic (SBP) and diastolic blood pressure (DBP), which are defined as the maximum and minimum blood pressure, respectively, during a pulse. They are measured in millimeters of mercury (mmHg). There has been great attention to automatic and continuous BP estimation using electrocardiography and/or photoplethysmography (PPG) signal [5], the accuracy and cost limit the viability of such methods

Objectives

Methods

Results

Conclusion