Non-invasive wearable sensing systems for continuous health monitoring and long-term behavior modeling

Abstract

Deploying new healthcare technologies for proactive health and elder care will become a major priority over the next decade, as medical care systems worldwide become strained by the aging populations. This thesis presents LiveNet, a distributed mobile system based on low-cost commodity hardware that can be deployed for a variety of healthcare applications. LiveNet embodies a flexible infrastructure platform intended for long-term ambulatory health monitoring with real-time data streaming and context classification capabilities. Using LiveNet, we are able to continuously monitor a wide range of physiological signals together with the user's activity and context, to develop a personalized, data-rich health profile of a user over time. Most clinical sensing technologies that exist have focused on accuracy and reliability, at the expense of cost-effectiveness, burden on the patient, and portability. Future proactive health technologies, on the other hand, must be affordable, unobtrusive, and non-invasive if the general population is going to adopt them.

(cont.) In this thesis, we focus on the potential of using features derived from minimally invasive physiological and contextual sensors such as motion, speech, heart rate, skin conductance, and temperature/heat flux that can be used in combination with mobile technology to create powerful context-aware systems that are transparent to the user. In many cases, these non-invasive sensing technologies can completely replace more invasive diagnostic sensing for applications in long-term monitoring, behavior and physiology trending, and real-time proactive feedback and alert systems. Non-invasive sensing technologies are particularly important in ambulatory and continuous monitoring applications, where more cumbersome sensing equipment that is typically found in medical and clinical research settings is not usable. The research in this thesis demonstrates that it is possible to use simple non-invasive physiological and contextual sensing using the LiveNet system to accurately classify a variety of physiological conditions. We demonstrate that non-invasive sensing can be correlated to a variety of important physiological and behavioral phenomenon, and thus can serve as substitutes to more invasive and unwieldy forms of medical monitoring devices while still providing a high level of diagnostic power.

(cont.) From this foundation, the LiveNet system is deployed in a number of studies to quantify physiological and contextual state. First, a number of classifiers for important health and general contextual cues such as activity state and stress level are developed from basic non-invasive physiological sensing. We then demonstrate that the LiveNet system can be used to develop systems that can classify clinically significant physiological and pathological conditions and that are robust in the presence of noise, motion artifacts, and other adverse conditions found in real-world situations. This is highlighted in a cold exposure and core body temperature study in collaboration with the U.S. Army Research Institute of Environmental Medicine. In this study, we show that it is possible to develop real-time implementations of these classifiers for proactive health monitors that can provide instantaneous feedback relevant in soldier monitoring applications. This thesis also demonstrates that the LiveNet platform can be used for long-term continuous monitoring applications to study physiological trends that vary slowly with time.

(cont.) In a clinical study with the Psychiatry Department at the Massachusetts General Hospital, the LiveNet platform is used to continuously monitor clinically depressed patients during their stays on an in-patient ward for treatment. We show that we can accurately correlate physiology and behavior to depression state, as well as to track changes in depression state over time through the course of treatment. This study demonstrates how long-term physiology and behavioral changes can be captured to objectively measure medical treatment and medication efficacy. In another long-term monitoring study, the LiveNet platform is used to collect data on people's everyday behavior as they go through daily life. By collecting long-term behavioral data, we demonstrate the possibility of modeling and predicting high-level behavior using simple physiologic and contextual information derived solely from ambulatory mobile sensing technology.