Continuous Passive Sensing for Bayesian Diagnostics in Mobile Health

点击翻译按钮获取中文摘要

Existing healthcare systems are reactive rather than proactive. Limited time or finances often force people to postpone visits to their doctor. Even when people are able to make an appointment, clinics are overburdened and doctors have limited time with their patients. Mobile health (mHealth) is being viewed as a potential complement to in-person consultations that can relieve stress on overburdened healthcare systems. One major aspect of mHealth is the use of sensors embedded within commodity devices like smartphones and smartwatches to measure biosignals and detect symptoms. Such applications typically fall into one of two categories: (1) tools that continuously monitor a person's physiology and behaviors, and (2) tools that require explicit interaction with a device to reveal information about a symptom. These tools are often discussed independently, which is contrary to the Bayesian diagnostic process clinicians implicitly follow. Clinicians initially formulate a prior probability that their patients have a medical condition according to their medical history and self-reported symptoms, and that prior is then updated using diagnostic tests. The central idea guiding my research program is that the efficacy of diagnostic tests can be improved by incorporating continuous passive sensing. For example, a smartphone app that diagnoses influenza by automatically interpreting a visual rapid diagnostic test should be more confident in a positive result if the microphone on the user's smartphone has detected more coughing than usual or the temperature sensor on the user's smartwatch detects a fever. Although my work will be motivated by health applications, the primary contributions of this work will be made in computer science. My research program will produce new knowledge in ubiquitous computing, human-computer interaction, and various forms of applied sensing (e.g., signal processing, machine learning, computer vision). These contributions will come in the form of two major streams of research: (1) passive behavior, symptom, and physiological sensing, and (2) probabilistic models for combining the aforementioned components together. My research program will create a flexible framework that provides a holistic understanding of people's health or wellbeing. My program will produce new knowledge in multiple subfields of computer science. In ubiquitous computing, my work will uncover novel problems that can be addressed with applied sensing. In machine learning and statistics, my work will advance techniques for interpretable multimodal models and will demonstrate new ways of combining multimodal data to form a rich representation of a person's state. Lastly, in human-computer interaction, my work will elicit user-centered design considerations for future mHealth interventions.