Fitbit has published data on its effort to develop an algorithm that detects COVID-19 using wearables before symptoms start.
Writing in the Nature journal npj digital medicine, researchers showed variables such as respiration rate and heart rate change in the days preceding the emergence of symptoms of respiratory disease. With a reported 90% sensitivity, the algorithm may detect 21% of cases the day before symptoms show.
The algorithm is yet to be validated in a prospective study. Fitbit is working with Northwell Health to validate the algorithm's early detection of COVID-19.
Efforts to use data from wearables such as Fitbit devices in the monitoring of infectious diseases predates the coronavirus pandemic, with groups at Scripps Research Translational Institute and other centers publishing the results of their previous explorations of the concept. However, COVID-19 intensified interest as public health experts scrambled for ways to stop chains of viral transmission.
In theory, changes in vital signs that occur between viral infection and the onset of symptoms could facilitate early detection of COVID-19, enabling the person to act to stop them infecting others and, if suitable treatments are available, get care before their condition worsens.
As wearables provide data to support such predictions, Fitbit began enrolling users of its devices into a study earlier this year, before going on to share early results in August. The company published its complete findings on Monday.
The paper describes the development of an algorithm based on data from 2,745 North American Fitbit users diagnosed with COVID-19. The heart rates and, in particular, respiratory rates of the subjects rose in the few days before the onset of symptoms. Fitbit used the data to develop an algorithm to predict whether an individual has COVID-19.
At 90% sensitivity, the algorithm detected 43% of cases the day after symptoms showed and 21% cases the day before symptoms appeared. Increasing the sensitivity lowered the detection rate. At 99% sensitivity, the algorithm detected 7% of cases the day before the onset of symptoms.
Fitbit used data during the peak of the COVID-19 outbreak in New York state to show hypothetically how its algorithm may perform in a major outbreak. The researchers estimate the daily disease prevalence in the state in mid-April was 0.65%. In that scenario, at 99% specificity around one in seven of the algorithm’s positive predictions would be a true positive. However, the true positive rate will be lower in areas with fewer cases.
While the algorithm would generate far more false than true positives even when prevalence is high, the researchers argue it could encourage people to take precautions or get tested. The spread of the virus could slow if presymptomatic people self-isolated upon being alerted by the algorithm. At 99% specificity, the algorithm is almost always right when it predicts a person is not infected.
Fitbit’s findings come with several caveats, including the fact that users self-recalled when symptoms started. A prospective study is needed to show whether the algorithm can actually predict if a person has COVID-19. Fitbit, which has U.S. Department of Defense funding, plans to run such a trial. Other groups including Sonica Health have also received government funding for coronavirus wearables.