Blood pressure (BP) is one of the essential indicators of human health and can be greatly influenced by lifestyle factors (e.g., activity and sleep). However, the degree of impact of each lifestyle factor on BP is unknown and may vary significantly between individuals. Utilizing data remotely collected by home BP monitors and wearable activity trackers, we aim to use machine learning techniques to investigate the complex relationships between BP and lifestyle factors in order to provide personalized and proactive hypertension care. In addition, we investigate continuous, non-invasive BP measurement using the photoplethysmogram (PPG) sensor which has become widely accessible in wearables and medical devices.
Data Analytics and Personalized Recommendations using Wearables
We have partnered with doctors at UC San Diego Health and the Altman Clinical and Translational Research Institute (ACTRI) in order to recruit pre-hypertensive and hypertensive patients to participate in our research. Using lifestyle data collected from Samsung Galaxy watches and BP data collected from Omron Evolv monitors, we aim to 1) provide a prediction of the patient’s BP, which will provide a quick and reliable way to understand their health condition, and 2) provide personalized and actionable insight to help the patient control their BP by adjusting their lifestyle behavior accordingly.
ML-based BP prediction and recommendation overview.
We propose Random Forest with Feature Selection (RFFS) together with trend and periodicity information extracted from historical BP to enhance prediction. 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 performance than an aggregate model trained on a larger but non-personalized dataset. Moreover, we propose a novel Online Weighted-Resampling (OWR) technique to address two main issues: 1) Since BP and health behavior data are collected and used for training sequentially, it takes significant time for new users to collect training data in order to achieve satisfying performance, which limits the usability. 2) Due to various external factors that cannot be captured by the proposed BP model (e.g., diet, stress and measurement error), non-stationary characteristics in BP and health behavior time series may deteriorate the prediction.
We utilize Shapley Value to determine which lifestyle factors have the greatest impact on the patient’s BP. Shapley Value is a model-agnostic feature interpretation method derived from game theory. Using this interpretable ML technique, our results indicate that the most important lifestyle feature ranges from patient to patient. Moreover, these results indicate that the impact of specific lifestyle factors on BP may differ across individuals. We form our personalized recommendations based on the top factors generated by Shapley Value analysis.
Left: Shapley analysis for two different patients. Right: Example of Personalized lifestyle recommendation report.
In order to determine the effectiveness of our recommendations, we conducted a micro randomized control trial (RCT) with our patient cohort. Patients in the experimental group received a report detailing their top factor and lifestyle recommendation while those in the control group did not receive any feedback. In the experimental group, 71% (5 of 7 subjects) and 86% (6 of 7 subjects) improved their mean SBP and DBP, respectively, compared to only 47% (9 of 19 subjects) and 53% (10 of 19 subjects) of the control group, respectively.
In the next phase of our study, our objective is to increase the interactivity of the ML-based recommendation system in order to improve patient engagement. Instead of providing a one-time lifestyle recommendation report, we will periodically provide feedback and encouragement to the patient according to their top lifestyle factors and BP. We aim to investigate the most effective timing of these recommendation prompts in addition to the content (text, graphs, etc.) and delivery channel (email, SMS, mobile app, etc.) to maximize patient compliance. In order to enable incoming patients to expeditiously receive actionable insights, we will develop online learning and transfer learning techniques to improve the performance of personalized ML models with limited data.
Personalized Blood Pressure Estimation using Photoplethysmography
Left: Principle of operation. Right: Transfer learning overview for PPG-based BP estimation.
For accurate diagnosis and treatment of hypertension, regular BP measurement is necessary. According to the American College of Cardiology, increased at-home BP monitoring is essential for recognizing inconsistencies in measurements taken in a medical setting. Currently, the predominant device for measuring BP is a mercury sphygmomanometer which involves attaching an inflatable cuff around the upper arm. This process requires significant user effort, which limits the frequency of BP measurements and increases the chance of measurement error. The use of an arterial catheter can provide continuous BP measurement; however, it is highly invasive and impractical for daily life. With the growing trend towards digital and virtual healthcare, there is a need for an automated, continuous and accurate BP measurement technique.
>In this research, we use the photoplethysmogram (PPG) sensor, which has become widely accessible in wearables and medical devices, and deep learning to enable non-invasive and continuous BP estimation. We propose a personalized deep learning approach that utilizes transfer learning in order to reduce the amount of personal PPG-BP data required from a new patient. We utilize the Multiparameter Intelligent Monitoring in Intensive Care III (MIMIC III) database in order to develop and test our method.
Personalized transfer learning approach for PPG-based BP estimation.
For future work, we will investigate a light-weight BP estimation model which can be implemented directly on a wearable device or embedded system while providing comparable performance to our current work. This will provide more real-time measurements and address concerns regarding data transmission and data privacy. While our method is developed and tested with ICU patient data, we plan to extend this approach to PPG data collected from consumer wearable devices. We will compare our personalized BP estimation models to cuff-based BP in order to further validate the accuracy.