Close

Presentation

CovIdentify - Developing Predictive Models for COVID-19 with Wearables Data
Author
Event Type
Oral Presentations
TimeFriday, April 1612:50pm - 1:10pm EDT
LocationDigital Health
DescriptionThe COVID-19 pandemic has resulted in over 9M infections and over 229k deaths in the US alone as of October 30, 2020. To ensure that new infections and clusters are identified before they promote viral spread, case-finding tools are needed to target diagnostic testing of individuals suspected to be infected. Our previous research has demonstrated that there are changes in physiologic and behavioral parameters measured by wearables in the setting of influenza infection, including high resting heart rate (HR), low HR variability, decreased blood oxygen saturation, disturbed sleep, decreased physical activity, and changes in wear habits. Together, these “digital biomarkers” form a signature of infection. Very recently, researchers have used these methods to successfully detect COVID-19 from wearable devices. To test whether this is possible, we launched CovIdentify in March 2020, a platform that integrates information from widely used wearables with simple daily self-reports on symptoms and social distancing, for up to 12 months. CovIdentify’s overarching objective is to implement existing digital biomarkers and establish new digital biomarkers by using our newly designed platform to develop, validate, and translate CovIdentify as a continuous screening tool. Since September 21, 2020, we have collected data from over 5,500 individuals.

However, while the data grows, the impact of this data on predicting altered physiology has been limited by two factors. The first factor is the limited infrastructure for analyzing large-scale wearable device data. While data pipelines to organize and condense data exist for the commercial wearable device manufacturers, the methodologies are proprietary and inaccessible to researchers. Organizations such as Open mHealth have created open-source database schemas for conducting research with wearable data, but their lack of scalability and simplification of the data results in a structure that cannot grow or handle data variations. Further, this limits the ability for real-time analytics to support just-in-time interventions. Therefore, our team has developed the infrastructure to connect survey and wearable device data from REDCap and commercial manufacturers, respectively, in the Microsoft Azure environment.

The second factor limiting the utility of wearable sensor data is a lack of participant interaction. Our study relies on a bring-your-own-device system to rapidly collect wearable data. With a barrier to entry for participants without a wearable device and a lack of financial incentives, our team came up with creative methods to improve both participant adherence to surveys and participant sign-ups. Our study launched an iOS application several months after our study began for users to not only fill out their daily symptom surveys but also receive information about testing sites near them and other information relevant to the virus. We also provided free wearable devices to local community members, in hopes of developing a study population that was more representative of the U.S. population.

By developing the infrastructure to work with large-scale wearable device data and improving our approach to engage users, we plan to create a model that will incorporate the different features measured and provide real-time feedback through our mobile application.