You are here
The Blind Spots in Our Biomedical Data
Between measuring our activity levels, heart rate and sleeping schedules, today’s smart watches seem to give us a better picture of our overall health. Duke BME’s Jessilyn Dunn explores the endless potential—and hidden limitations—of this data.
This is Rate of Change, a podcast from Duke Engineering dedicated to the ingenious ways engineers are solving some of society’s toughest problems. I’m Michaela Kane.
It’s hard to imagine a time when smartwatches weren’t a common accessory decorating people’s wrists. What was once a tool used almost exclusively by serious athletes and self-proclaimed weekend warriors has transformed into an everyday tool that can provide information about heart rates, daily step counts, and workout statistics to the user. They can also provide helpful––if sometimes annoying––nudges telling you to get moving if you’ve been sedentary for most of the day.
Most people are content with the basic knowledge these devices can provide about their health, but Jessilyn Dunn, an assistant professor of biomedical engineering and biostatistics and bioinformatics at Duke University, is aiming to explore how the data from these tools can help develop new methods of disease prevention, detection and treatment. Dunn accomplishes this work with her team, which makes up the Big Ideas Lab at Duke.
Dunn: The name actually stands for something. It’s the Biomedical Informatics Group, and the idea is that we’re integrating engineering, data and analytics, and really what we’re doing is working across multiple different disease areas with the goal of developing what we call digital biomarkers, so methods of monitoring health and disease using less traditional data sources than someone going into the clinic at a single visit per year and getting a physical and workup done at that one time. What we’re hoping to do is use more continuous methods of monitoring whether that’s through wearable devices or mobile devices. We also do a bit of work with biomolecular data and other types of biomolecular sensors. So our goal is to really bring all this data and technology together to better be able to predict illness states.
Kane: Can you give me an example of what that would look like, whether it’s one disease or for an average patient, what would that involve?
Dunn: Yeah, absolutely. So one of the projects we’ve been working on for a little over a year has been the prediabetes detection project. We were funded through Duke MEDx and the CTSA to essentially try to develop methods using more commonly available wearables like an Apple Watch or Fitbit that would actually be able to pick up prediabetes. So that’s kind of the big idea, if you will. And the way that we’re doing that is actually by using devices that are a little more complex than the actual common consumer devices. But the goal is that we’d have a non-invasive method of being able to detect who is likely to have prediabetes or be at risk of developing either pre or type-2 diabetes.
And the reason that this is such an important area for us is that actually 1/3 of the US population is prediabetic and 90 percent of those people who are prediabetic don’t actually know that they’re prediabetic. So that’s a really scary stat. Without people having an appropriate diagnosis, they will never know that they should be making an appropriate lifestyle change to prevent the onset of prediabetes. So the earlier that we can catch this actually prediabetes is reversible, but people need to know that they have it.
Kane: According to Dunn, the team uses both data driven methods and hypothesis driven methods to develop these digital biomarkers. On the data driven side of things, the team is focused on developing machine learning algorithms that can parse through huge amounts of data from different sensors and create predictions about who may be sick with certain illnesses.
Dunn: On the other side of the coin are hypothesis-driven methods, where we know that there are physiologic changes when someone has pre or type-2 diabetes. For example there is some damage to the autonomic nervous system that can take place, and that represents by various different physiologic changes, one of those is changes in heart rate variability, and so we’re looking into how that may relate to diabetic status.
Kane: The ability to create these biomarkers that can be tracked with common wearable devices also addresses an important gap in health care knowledge, which is long-term health monitoring. This is especially helpful for people who are unable to see a doctor on a regular basis.
Dunn: So, I think one of the major challenges with our current healthcare system is that people very sporadically go into the clinic, and the often go into the clinic when something is wrong, so we don’t actually get a good picture of what health and illness look like outside of the time when someone is physically present in the clinic. And so these wearable sensors and mobile sensors enable continuous monitoring, so we’re actually getting a picture of somebody’s physiology 24 hours a day, seven days a week. And another key point here is that there is actually circadian differences in our physiology, so if you go to a doctor’s appointment first thing in the morning one year and then at the end of the day another year those are actually not directly comparable because your body changes over the course of the day and over time. So it’s very difficult to understand somebody’s baseline health with only these sporadic pieces of information.
Kane: Having a comprehensive picture of someone’s baseline helps doctors identify when something is wrong, and Dunn and her team are optimistic that wearable technology can act as a solution to this issue and provide valuable medical insight to both doctors and patients. But before this is possible, Dunn and her collaborators across the digital medicine world are also working to ensure that both common smartwatches and clinical sensors are as accurate and useful as possible.
Dunn: We’ve actually been heavily involved with some of the initiatives going on in the digital health and the digital medicine space to ensure that tools are used in a way that we call ‘fit for purpose’. The challenge here is that the tools are incredibly useful so your standard consumer smartwatch can tell you a lot about your heart measurements, your physical activity and your sleep, and lots of researchers have realized this and have started to integrate them into clinical studies and clinical trials. The challenge though is that there’s no set standards on information that needs to be measured or released about these devices to ensure that the measurements that are coming off of them are accurate enough to be used for clinical purposes.
Kane: To help standardize the data collected from smartwatches and other devices, Dunn a team of researchers across digital medicine proposed a V3 framework that appeared in Nature Digital Medicine. In their three step process, engineers would verify, analytically validate and clinically validate the sensors, algorithms and data produced by wearables and other devices to identify any limitations in their accuracy.
Dunn: In our lab we’ve applied that framework, specifically to look at how skin tone and physical activity affect the accuracy of heart rate monitors that are optically based.
Kane: Fitness trackers currently measure heart rate using a process called photoplethysmography, or PPG. This involves shining a specific wavelength of light, which usually appears green, from a pulse oximeter sensor on the underside of the device where it touches the skin on the wrist. As the light illuminates the tissue, the pulse oximeter measures changes in light absorption and the device then uses this data to generate a heart rate measurement.
But the lab had seen anecdotal evidence that suggested that skin tone could affect the accuracy of the PPG sensors, as melanin, which tones our skin, could also absorb the green light.
Dunn: We really wanted to see systematically ‘is this true, is there evidence for this,’ and we had even seen that some of the wearable device companies that we were evaluating had in their user materials ‘these devices will not work as well in darker skin tones,’ and for us that was shocking. If we’re developing these technologies to work for people who may not be able to come into the clinic, a large portion of those people may have darker skin and we need this technology to work for everyone. So we designed this study with that in mind and we looked across the Fitzpatrick skin tone scale, so essentially we had several different skin tone categories to lightest skin which is very low melanin content to darkest skin with the highest melanin content.
Kane: The team explored the accuracy of several different commercial devices, including the FitBit, the Apple Watch, a Garmin watch and the Show Me Me-Band, as well as research grade devices like the Empatica E4 and the Biovotion.
And what we saw actually we were glad to see that there were not significant differences in the accuracy of the heart rate across different skin tones, but throughout the process we had people go through different types of activities because there was a potential that there was interactions between the skin tone and types of activity, and we did see then in fact in higher levels of physical activity the heart rate was less accurate. In fact on average it was up to 30% less accurate than heart rate at rest. So that’s definitely a big problem when we think about how we’re using the data from these devices. So overall it seems like there’s been a lot of algorithmic methods to address the potential skin tone issue, but we still don’t have algorithms to address the movement issue.
Kane: In addition to studying the physical and algorithmic limitations of these devices, Dunn is also exploring how potential biases in data collection may affect how researchers can identify biomarkers for a variety of diseases.
Dunn: There are a lot of communities that are not typically represented in a lot of these larger scale wearable device studies because a lot of these studies are designed as a bring your own device study, and we know that the people who would typically buy a device are not necessarily representative of the overall US population and are particularly not representative of underserved communities. So the problem with that is that when we’re developing these digital biomarkers, they are based on predictive algorithms. So we have these data driven approaches and we have these hypothesis driven approaches, and when we’re developing data driven approaches to making predictions, the data is critical. If we don’t have data that represents all people then our algorithms are only as good as the data that we put into them.
And that becomes really dangerous when we think about the populations that are represented in the studies that are developing these new technologies, because all of a sudden it’s very easy to develop technologies that work for some but not for others. So it is really critical that we have this appropriate representation of all groups in our data sets.
Kane: All of this work has helped to inform and set up the lab’s newest research project, CovIdentify. Originally launched in early April, CovIdentify was designed to explore how data collected by smartphones and smartwatches could help determine whether or not device users have COVID-19, which is the disease caused by the novel coronavirus.
Dunn: So what we want to do is we want to be able to detect, even before somebody might know they have a COVID-19 infection, that they’re sick. The idea there, one, is obviously to prevent the spread, right, so if somebody knows that they have the infection they won’t be spreading it to their family or community members. They would know to quarantine, they would know they should probably get in touch with their doctor and let them know what’s going on, and the other piece is that we’re hoping to learn more about what the trajectory of infection looks like.
Different people have very different outcomes from COVID infection. We’ve seen that some people are completely asymptomatic. They have no indication at all that they’ve been infected, and what’s even scarier about that is that some of those people do have physiologic changes that they themselves just weren’t able to detect. So they may have some lung damage or other damage that they themselves don’t even feel. So it would be important to know that they’ve had an infection.
And then there’s the other dramatic end of the spectrum where this disease is really lethal for a lot of underserved communities, people from communities of color, where they have significantly worse outcomes. They have higher rates of infection due to higher risk living or working environments. A lot of times these communities have higher comorbidities. So they have other diseases like diabetes and cardiovascular disease that can compound the danger of COVID. So there is a lot there that needs to be examined and uncovered, and the sooner that we can understand what mobile and wearable sensors can tell us about the disease the sooner we can develop better interventions.
Kane: In the first phase of CovIdentfy, the team launched the covidentify.org website, where people can sign up and input their relevant demographics and medical information. The study then involves a daily survey, which can be delivered by text, email or through the CovIdentify app, which asks two simple questions. The first question asks if the participant has been in contact with anyone outside of their household and if they’ve been maintaining social distancing, and the second question asks if they feel sick. If they say that they do feel sick, the survey then expands to ask about specific symptoms.
Participants are also asked to share biometric data from their smartphones and smartwatches. The team then matches the biometric data with answers from the survey questions, and they’ll use the relationship between reported symptoms and any changes in the biometric data to begin to develop biomarkers for COVID-19.
Dunn: But because we have this problem with this ‘bring your own device’ research model, we’ve actually been working on getting device donations and devices purchased through some grant funding to be able to hand out devices to people from underserved communities. So we’re currently in the process of that and we’ve been working with folks from the CTSA and the mobile app Gateway and also the VA to better engage with the community and get devices to people who otherwise wouldn’t have them.
Kane: The CovIdentify team also developed a plan to target recruitment of people that have a higher risk of contracting the disease, including delivery drivers, grocery store workers, hospital cleaning and cafeteria staff and nurse aids. They are also exploring how they can deploy watches in high-density housing like nursing homes, college dorms, military barracks and homeless shelters.
If CovIdentify is successful, it would be a non-invasive and accessible tool that can help control the spread of the coronavirus. It also speaks to the lab’s larger goals of arming health care professionals with accessible tools and information to detect illness early, which helps doctors save more lives.
Dunn: One of the ideas here is also to provide some patient empowerment or even just person empowerment. I think part of the idea here is that if we can provide health insights, those are useful to clinicians to help care for their patients, but those are useful for us as individuals, to know things about our health and to be able to understand what to do with health recommendations. So what we’re trying to do here is really close the loop on this healthcare system to make sure that people have access to information about their health and also have access to what comes next. You know, what kind of actions can they take to potentially reverse their prediabetes, for example.
Kane: Thanks for tuning into this week’s episode of Rate of Change. Remember to follow us on social media for updates, and be sure to subscribe. Thanks for listening, and stay safe!