Blood pressure (BP) is one of the most important indicator of cardiovascular disease (e.g., coronary artery disease, stroke and heart failure) and highly correlated to health behavior (e.g. exercise, and sleep). At the same time, health wearable trackers and wireless home BP monitors become accessible in people's life. The quality and granularity of the collected data can be used for early prevention or improvement of the therapy in the hospital as well as at home. However, the potential of such data has not been fully utilized. Together with data mining and machine learning (ML) techniques, connected health brings full potentials and open problems for academia and industry. In this project, we focus on two targets to better control blood pressure through health behavior: 1) define the valuable problems in relationship between blood pressure and health behavior (e.g. exercise and sleep) and 2) providing personalized healthcare including monitoring and recommendation.

Picture2_0.jpg

System Architecture

Data Analytics and Personalized Recommendation using Wearable

Sleep and exercise are proved to be statistically correlated with BP with randomized controlled trials. However, the personalized effect of health behavior has not yet been studied. We propose to use ML techniques with individual’s historical BP and health behavior to 1. construct a personalized model to predict daily BP using an individual’s historical BP and health behavior, and estimate the effect of the individual’s health behavior on his/her BP and 2. provide personalized recommendation to users to improve their health targets.  

We propose RF with Feature Selection (RFFS) together with trend and periodicity information extracted from historical BP to enhance prediction. Using Fitbit Charge HR and Omron Evolv blood pressure monitor, our experimental results demonstrate that the proposed approach is robust to different individuals and has smaller prediction error than existing methods. Moreover, we show that the proposed personalized BP model achieves better result than the aggregated model based on a larger but non-personalized data set. We also validate the effectiveness of personalized recommendation of health behavior by showing the significant change in BP after users changed their the most significant health behavior features suggested by our model. 

PictureCH.jpg

Personalized BP model with BP estimation and effect of health behavior

In the next phase, our objective is to enhance the accuracy and effectiveness of personalized recommendation on health behavior. Therefore, we partner with doctors at UC San Diego Health and Altman Clinical and Translational Research Institute (ACTRI) for better design on patient engagement and intervention. Moreover, we aim to better utilize the high granularity of wearable data. This can be done either by raising the sampling frequency of BP or better representation of features to aggregate the wearables data. We partner with Samsung  Health USA to develop novel data collection techniques with their advanced wearable devices. On the ML perspective, to enable incoming users to have better result with limited data, we will develop online learning and transfer learning to improve the performance of personalized ML models.  

 

Associated Publications