Data collected from Fitbit® wearable devices significantly improved the real-time surveillance of influenza-like illness (known as “nowcasting”), according to study results published in The Lancet. Up to 650,000 influenza deaths occur worldwide each year, and influenza surveillance reports are often delayed by 1 to 3 weeks and then revised months later, allowing outbreaks to grow and spread unnoticed. Because the physiological response to the inflammation from acute infections can cause an elevated resting heart rate and changes in daily activity, investigators aimed to assess if seasonal respiratory infection trends, such as influenza, could be tracked through wearable devices that collect sleep data and resting heart rate.
Deidentified sensor data from 200,000 US Fitbit users who wore a device between March 1, 2016, and March 1, 2018, for ≥60 days were taken from users in the 5 states with the most Fitbit users: California, Illinois, New York, Pennsylvania, and Texas. Inclusion criteria included a self-reported birth year between 1930 and 2004, a weight >20 kg, and a height >1 m. Sensor data were compared with weekly state-level estimates of influenza-like illness rates as reported by the US Centers for Disease Control and Prevention (CDC). For each state, influenza-like illness case counts were modeled with a negative binomial model, including 3-week lagged CDC influenza-like illness rate data (null model) and the proportion of Fitbit users with increased sleep duration above a specified threshold and elevated resting heart rate each week (full model). Predicted influenza-like illness rates vs CDC reported rates were compared using Pearson correlation.
After assessing the 200,000 Fitbit users with available data, 47,249 met the inclusion criteria. The mean age was 42.7 years and 60.2% (n=28,465) were women. Two different combinations of resting heart rate and sleep measurements were tested as data abnormality thresholds for the final models. Using model 1 classified 24.3% of 2,186,559 weekly Fitbit measurements as abnormal, and using model 2 classified 11.2% of the measurements as abnormal. The highest correlation with influenza-like illness rates reported by the CDC were found using the model 1 thresholds — that is, defining abnormal data as 0.5 standard deviation above the individual’s average resting heart rate combined with sleep measurements more than 0.5 SD below their average.
The wearable device data significantly improved influenza-like illness predictions across all 5 states, with an average Pearson correlation increase of 0.12 over baseline. Improvements ranged from 6.3% in New York to 32.9% in California. CDC influenza-like illness rate correlations with the final model ranged from 0.84 to 0.97, with week-to-week changes in the proportion of device users with abnormal data in most cases associated with week-to-week changes in influenza-like illness rates. Prediction levels from model 1 were high in all states, with the highest correlation seen in California (r=0.97; P <.0001), and the lowest seen in New York (r=0.89; P <.0001).
Study limitations included no activity data, including resting heart rate averages from sick and well days, and the low accuracy of sleep devices. However, the investigators concluded that their findings showed that using metrics from wearables, particularly those devices with continuous sensors, could improve real-time influenza surveillance.
“In the future, with access to real-time data from these devices, it might be possible to identify [influenza-like illness] rates on a daily, instead of weekly basis, providing even more timely surveillance,” the researchers wrote. “As these devices become more ubiquitous, this sensor-based surveillance technique could even be applied at a more global level where surveillance sites and laboratories are not always available.”
Radin JM, Wineinger NE, Topol EJ, Steinhubl SR. Harnessing wearable device data to improve state-level real-time surveillance of influenza-like illness in the USA: a population-based study [published online January 16, 2020]. Lancet Digital Health. doi:10.1016/S2589-7500(19)30222-5