Jessilyn Dunn, a biomedical engineer from Duke University, always wears a smartwatch, but you’ll never catch her wearing the same one twice. Dunn doesn’t use the watches to count her steps; she wants to know which ones can predict if the wearer is getting sick.
Dunn’s research team recently analyzed biometric data — heart rate, temperature, and skin conductance — from 39 health monitor wearers who were voluntarily infected with H1N1 influenza or rhinovirus (1). All of the participants wore the same type of monitor, an E4 wristband, for a week after they were infected. Dunn’s team predicted if a participant was falling ill within 24 hours and foresaw how severe or moderate the infection would become with nearly 90% accuracy.
Jessilyn Dunn is a bioengineer at Duke University who uses data from wearable health monitors to identify predictive biomarkers for infection.
Credit: Jessilyn Dunn
Gathering biometric data from smartwatch users isn’t usually this controlled; it’s not every day that forty people volunteer to get the flu. But for further exploration, Dunn recruits people to self-report their vitals daily for a year using a variety of digital health monitors, including smartwatches and phones, with the hope of finding a way to predict COVID-19 infection; the project is aptly named COVIDENTIFY. The data are still being collected, but Dunn is excited to develop an algorithm that could catch COVID-19 before a patient even feels the first symptom.
What led you to focus on wearable health monitors?
In my early research days, I analyzed biomolecular datasets to address research questions in genomics, transcriptomics, and epigenomics. I started with small datasets back when the one gene one protein hypothesis was prevalent. As -omics gained popularity, I started working with larger datasets. I really enjoyed working with these datasets, but I wanted to conduct research that had a more immediate impact on healthcare and public health.
During my postdoctoral fellowship at Stanford University, I did a side project using wearable health monitors in combination with other -omics techniques. In that process, I realized that this data was really cheap and easy to collect. The biometric data give us a lot of information about someone’s health, and it can be used immediately to develop disease interventions or testing methods. I found it fascinating, so I pivoted my research focus to wearables and other digital health data methods.
How do you apply biometric data to solve public health problems?
My research group’s expertise is in digital biomarker development. We take continuous, longitudinal data coming from wearable health monitors such as smartwatches and determine how the data relate to disease outcomes. We use a lot of statistical learning and machine learning techniques. We define an outcome of interest, infection status for example, and then we collect a bunch of variables in the raw data from a wearable health monitor. We bring the data together using predictive modeling methods like linear regression to develop some cool, complex models that detect patterns in the data to predict the outcome of interest.
What did you learn from your recent study that can be applied to the COVIDENTIFY study?
A lot of the physiological changes that happen with flu or rhinovirus infections are like those that happen with COVID-19. We can identify digital biomarkers and apply them to see if we can detect who does and does not have COVID-19 from those biomarkers alone. That is what we have been working on non-stop for the past two years.
Can you use these digital biomarkers to distinguish between different viral infections?
We see changes associated with heart rate, physical activity, and sleep, so we combine these different metrics together. But these are pretty common among different types of illnesses, which makes it difficult to distinguish between infections. We’re not quite there yet with identifying specific digital biomarkers for SARS-CoV-2 infection, but we hope that we will identify a unique signature.
We don’t have as much data as we expected. We have a lot of people enrolled, but not everyone wears their watch regularly. These are just people living their daily lives, and when they get sick, they tend to take their watches off. The people who wear their watches regularly seem to have lower infection rates. That’s one of the reasons it’s been so hard to develop these digital biomarkers for COVID-19; the data we collect are tricky to analyze. If there are people out there who had COVID-19 and wore their smartwatch during the course of their infection, those are really the people we want to donate their data so we can analyze it retrospectively.
Have many of the participants in COVIDENTIFY have developed COVID-19?
We imagined that if we recruited people who didn’t have COVID-19 that a number of people enrolled in the study would come down with it, but that didn’t happen. It turns out that people who seek out research are also more careful about taking preventative measures against infection. The prevalence of COVID-19 in our study population was much lower than the broader population, which is to be expected, but it makes our job more challenging.
How do you know if someone in the cohort had COVID-19?
We recruited many people from Duke University because there is continuous surveillance and a lot of regular COVID-19 testing on campus. We have a nice, clear set of COVID-19 negative people who are regularly tested. But we cannot assume that someone who has not had a COVID-19 test is not infected because patients are often asymptomatic. We have to rely on testing or self-reported symptoms, which also limits our data.
Does it matter how people collect their personal biometric data?
That’s still a work in progress. We’ve done head-to-head comparisons of different devices for accuracy. We anticipate that there will be some differences, but whether the variability matters depends on what the most predictive factor for SARS-CoV-2 infection is. For example, a lot of commercial devices don’t do a great job of measuring heart rate variability. If heart rate variability is the best digital biomarker for SARS-CoV-2 infection, then what device the participant uses really matters. Research grade devices like the E4 wearable used in our recent JAMA Network Open study are the best for data collection, but they aren’t as stylish. We hope that devices that collect high resolution data become stylish and fun to use at some point, but they haven’t yet.
What is the most predictive digital biomarker for respiratory infection based on your studies so far?
Biometric data collected during sleep is the cleanest because people aren’t moving around and there are expected and predictable changes in heart rate and temperature that happen during sleep. If we see changes that don’t happen as expected, that gives us a clue that something is up.
How can predictive digital biomarkers influence public health?
If we know that people have an infection before they develop symptoms, we can prevent them from going out and spreading it. This will give us the ability to do population level surveillance to determine where there are outbreaks. I hope that we can eventually use these devices not just for measurements, but for communication. Imagine if we could get information that someone is likely to be sick, and we could send them a notification that says they are likely to be infected and should go get tested. We could even send them precautions they should take. In the longer term, I see us combining digital health monitoring with intervention.
This interview has been edited and condensed for clarity.
Reference
Grzesiak, E. et al. Assessment of the Feasibility of Using Noninvasive Wearable Biometric Monitoring Sensors to Detect Influenza and the Common Cold Before Symptom Onset. JAMA Netw Open 4, e2128534 (2021).
