With the increasing prevalence of smartwatches and fitness trackers, what’s the best way to harness the potential of these devices? A team of NIH-funded researchers has an idea: use these wearable sensors as a way to predict clinical trial results, which could potentially serve as an early warning sign for underlying health problems.
“Consumer wearable devices have enormous untapped potential to facilitate monitoring (and potentially predicting) human health and disease,” said Grace Peng, Ph.D., NIBIB program director in Mathematical Modeling, Simulation and Analysis. “This study, which investigates how smartwatch data is associated with clinical laboratory testing, is an important step forward in this burgeoning field.”
To better understand how smartwatches could be incorporated into routine healthcare, the study authors first evaluated how data is captured using a wearable device compared to measurements taken in a clinical setting. To do this, they followed 54 participants for approximately three years. During this time, each participant made around 40 clinic visits and wore a smartwatch for around 340 days. The smartwatch measured four vital signs: heart rate, skin temperature, step count, and electrodermal activity (a measure of skin conductance). In the clinic, heart rate and oral temperature were measured, and clinical laboratory panels were performed, including a complete blood count, a complete metabolic panel, and a cholesterol panel.
The researchers evaluated the difference in vital signs measured during clinic visits versus continuous measurements from a smartwatch. They found that temperature measurements were more consistent when assessed in a clinical setting, as oral temperature generally had less variability than skin temperature measured with wearable devices. However, the researchers found that the smartwatch provided more accurate heart rate readings, as the measurements taken in the clinic had significantly greater variability between them. “When your vital signs are measured in a doctor’s office, there are a wide variety of variables that can affect the heart rate measurement, such as the time of day, what types of activities you were doing before your appointment, or even whether you were nervous during the test,” explained the study’s lead author, Michael Snyder, Ph.D., chair of the Department of Genetics at Stanford University School of Medicine. “On the other hand, because a smartwatch is worn continuously, the user’s heart rate can be measured throughout the day, resulting in a much more consistent measurement with significantly less variability.”
Next, the researchers wanted to determine if they could predict clinical laboratory test results using information collected from smartwatches. Generating such prediction models requires an immense amount of information, which was captured during the extensive monitoring period. Therefore, the researchers analyzed the longitudinal data collected from the devices and converted the measurements into more than 150 different characteristics, such as average heart rate during exercise, nocturnal variability in skin temperature, and overall electrodermal activity. Then, using machine learning models, they combined these features to predict clinical laboratory results.
The researchers then compared the predicted results generated from their models with the results observed in laboratory tests performed in the clinic. They found that their predictions aligned with the results of several clinical tests, with four blood tests having the most predictable results. These tests included red blood cell (RBC) count, absolute monocyte count, hemoglobin (HBG) levels, and hematocrit (HCT) levels. Interestingly, the researchers found that measurements related to electrodermal activity were an important factor in predicting RBC, HBG, and HCT test results.
“Electrodermal activity is not typically measured in a clinical setting, but it is one of the technologies used in a lie detector test,” explained first author Jessilyn Dunn, Ph.D., assistant professor of biomedical engineering at Duke University. “Basically, we’re looking at the opening of sweat glands, which could be in response to stress, temperature, emotional state or even a measure of hydration,” he said. “Electrodermal activity likely has great clinical potential, such as detecting dehydration, especially in older people, but has not yet been exploited as a resource in this setting.”
Other smartwatch measurements were key in predicting specific blood test results. For example, the most important measurements for predicting absolute monocyte count were based on step count and skin temperature. On the other hand, platelet count prediction was based on heart rate-related measurements, while fasting plasma glucose prediction used a combination of skin temperature, heart rate, and step count measurements. “Our results suggest that different physiological characteristics are associated with the prediction of different clinical measurements,” Dunn said.
However, the researchers emphasized that smartwatch data is not a substitute for clinical testing, but rather could serve as an early warning sign, potentially prompting the user to consult their doctor. “The power of wearable devices is their ability to detect changes from initial readings,” Snyder said. Even if some specific measurements aren’t very accurate, the watch’s ability to detect changes in the wearer’s vital signs could be immensely useful, he added. “The current medical paradigm focuses on treating patients when they are already sick, not on monitoring healthy people for early detection of disease,” Snyder said. “We believe that data from smart watches could help intercept emerging diseases, which could ultimately prevent more serious diseases.”
This research was reported in Nature medicine.
This study was funded by NIH Common Fund Human Microbiome Project grant DE023789-01 and an NIH National Center for Advancing Translational Sciences (NCATS) Clinical and Translational Sciences Award (TR001085). The authors of this study were supported by the Mobilize Center grant (EB020405) and the NIH Career Development Award (ES028825).
Study reference: Dunn, J., Kidzinski, L., Runge, R. et al. Wearable sensors enable personalized predictions of clinical laboratory measurements. Nat Med 27, 1105–1112 (2021). https://doi.org/10.1038/s41591-021-01339-0
