Smartphones, activity trackers and other mobile devices come equipped with sensors that enable the unobtrusive collection of behavioral and physiological data as people go about daily activities. Smartwatches can track heart rate, heart rhythms and blood pressure. Apps on many devices enable people to log streams of information about mood, stress levels or use of medications. The combination of sensors and self-reporting can provide insights into activity levels, weight and disease symptoms. Such patient-generated health data (PGHD) has enormous potential to provide real-world insights into disease and treatment.

Luca Foschini, PhD, is the co-founder and chief data scientist at Evidation Health and an expert in machine learning, behavioral economics and medical informatics. He answers our questions about how artificial intelligence (AI) and PGHD can advance clinical studies for biotech and pharmaceutical companies.

Edited and condensed for clarity

Much of AI in healthcare is focused on the use of electronic healthcare data, which provides a very limited and biased view of a patient's health. When thinking about opportunities in healthcare for AI, what do you think wearables and other forms of PGHD can bring that traditional medical data cannot?

Most people, if they are lucky, spend less than 1% of their time visiting a doctor. PGHD enables us to collect data about the remaining 99% of their time. When someone has their blood pressure measured, for example, we obtain one measure taken at a particular point in time. Yet that measurement is sometimes completely biased by the “white coat effect,” when a patient’s blood pressure spikes in the clinic. By measuring blood pressure continually in daily life, we can establish a patient’s baseline and then better understand when it deviates from the norm.

PGHD also provides insights we can’t obtain at the clinic. For example, in a recent feasibility study we used smartphones or smartwatches to monitor real-world behaviors such as typing speed, pace of speech and text messaging in a group of 31 people with mild cognitive impairment and 82 healthy controls. We also administered a battery of validated cognitive tests. The study provided insights into how technology can help us better understand and identify the earliest signs of cognitive decline.

Wearable data can be very noisy, which can present problems for developing an AI system. What are approaches you have used to mitigate this?

From a statistical perspective, wearables and other forms of PGHD are indeed noisy — data that may be meaningless or is hard to interpret. There are three ways to mitigate noise.

First, PGHD enables us to collect data continually over time, providing a longitudinal picture. AI then becomes helpful for anomaly detection. A machine model learns what a particular patient’s normal data stream looks like, day after day. Whenever a measure doesn’t fit with the learned model of normal — that is, it is a few standard deviations away from the mean — AI can detect it and provide an alert.

Second, the data collected is multimodal. A sensor may collect data about a patient’s heart rate but will also provide data about their activity level. By triangulating different data streams, researchers can understand if the data points are noise or are true signals.

Third, and most important, PGHD is not just a measurement engine; it provides a direct connection to an individual. You can ask questions to provide context for the data. Let’s say your Apple Watch measures an equivalent of a six-minute walk test, which is a measure of overall health extensively used in clinical trials, and that data is shared with your doctor. If you are fit, yet perform poorly on the test, your doctor can ask questions. Maybe you are sleep-deprived or ran a marathon recently. Or let’s say a smartwatch detects what appears to be tachycardia in a person who normally has a steady heart rate. The doctor can ask if they wore the watch in the shower, where the water can interfere with the sensors in a smartwatch. This two-way communication ensures that participants are not just passive data collectors, but are partners in the process. Not only is this the most effective way to combat statistical noise, but it also helps build trust with participants about how their data is being used.

Evidation is not a device manufacturer but instead is a platform which connects individuals who own wearables with organizations conducting clinical research. How did you arrive at this strategy, and to what extent have AI and machine learning been an enabling feature?

Evidation is a two-sided platform that connects individuals and organizations. On the consumer side are the individuals who use our Achievement app, where people earn points for walking more, getting enough sleep and engaging in other healthy behaviors. The enterprise side offers Achievers opportunities to join research studies sponsored by biopharma, leading tech companies, academic institutions and government partners. Our platform allows complete transparency about where someone’s data goes and how it is used: participants sign a consent for each study, even when reusing the same data.

We started by building the individual side of the platform, to give people a way to improve their health. Then we realized that people who are motivated to get healthier might also want to participate in research studies. For example, someone who has migraines might want to participate in a study about triggers of migraines in the real world and why people switch from one medication to another, so that a research group can improve treatment.

In recent federally-funded preliminary research, we developed a system using AI to detect data collected from commercial wearable sensors. The research analyzed data on sleep, heart rate and activity levels collected from smartwatches and other wearable sensors, as well as self-reported symptoms by participants. This was just one of several investigations we’ve conducted to better understand how PGHD can enhance the understanding of COVID-19 symptoms in the real world (note: automatic download).

We recently developed a similar model to alert participants who had developed flu symptoms in the past 48 hours about how to enroll in a clinical trial of a new flu treatment. This is the type of research that would be difficult to conduct in the traditional way. You can’t go to a hospital and post flyers and expect people to remember who to contact when they get the flu. There were many clinical trial sites in the United States. Whenever someone experienced symptoms and lived near a site, our model sent them an alert and gave them an opportunity to go to a local site and join the study. That activation at scale would not have been possible without AI.

— Ann MacDonald

Continue the conversation by joining Luca Foschini at the upcoming Designing and Implementing AI Solutions for Health Care program, or connect with us on Twitter @HMS_ExecEd or @evidation(Evidation) and @calimagna (Luca Foschini).