The Corona Data Donation Project is designed to detect fever symptoms in the time-series of the donors' daily resting heart rate and daily step count signals. See blog post “How does it work?" for the mechanics of how the whole thing works and the basic idea behind it, or check out some of the preliminary results we obtained in the past.
In the context of the current covid-19-pandemic this is important because fever can be one of the symptoms of covid-19. So, if we detect an increase in the proportion of donors with potential fever symptoms this might be an indicator that covid-19 infections are on the rise as well.
But does it work?
One way to check this is comparing the time course of the new daily confirmed covid-19 cases in Germany with the estimated cases of fever in the community of donors that we detect and compute with our algorithms. If both curves have similar shapes as a function of time, if they follow similar “ups” and “downs”, it is an indication that it works. The figure below illustrates the comparison:
We see in the figure the complete time-series of confirmed cases of covid-19 in Germany from early March to late August. Superimposed is the detection rate of fever as computed from the Corona Data Donation Project. Note that the detection curve only covers a period starting in late May. Although the whole project started in April, in order for detections for work, a 4-5 week baseline has to be computed prior to the detection period. Because fever detections are only proportional to covid-19 cases, the vertical offset of the curves is not so important. Rather, the shapes of the curves matter, so how similarly they behave as time advances.
Bottomline: Both curves seem to follow very similar trends which is an indication that the fever detection algorithm seems to be “picking up the covid-19-signal”. This is good news. But keep in mind, this is a first shot using the approach and there’s still room for improvement. But as a preliminary results this is quite promising.
Let’s take a closer look and zoom into the most recent dozen weeks or so:
The figure above depicts the same comparison of covid-19 cases with fever detections during the last few weeks. When we compare the covid-19 case count (red) with the fever detections we see that important features, such as the continuous increase of cases throughout July until mid August 2020 are captured by the detections. Futhermore, the detection curve seems to precede the confirmed cases by a week. This could imply that we see a signal in the detections before covid-19 cases are reported in the system. If this is so, the whole Corona Data Donation Project could become a valuable tool in detecting trends in the pandemic in Germany early on. However, we really need to run more tests to see whether this result is robust and not mere coincidence.
Fever does not equal covid-19
This is all good news, but remember that the Corona Data Donation Project is designed to detect fever symptoms and cannot identify covid-19 disease directly. There are many other infectious diseases and ailments that have fever as a generic symptom. Our algorithms cannot (yet) diffenentiate between different infections that induce fever in a person, so it is important to check whether the fever curves that we measure are not dominated by other diseases like influenza like illnesses or acute respiratory diseases.
The figure below compares the confirmed covid-19 case count and the computed fever detections with surveillance data on influenza like illnesses and acute respiratory diseases monitored in by German GrippeWeb platform.
The data in this plot aggregates data by week. Comparing these time-series we see no indication that the detections computed in the Corona Data Donation Project are dominated by cases of influenza like illnesses or acute respiratory diseases. This, or course, does not rule out other sources that we have not considered yet. However, the evidence that we are actually measuring a trace of the actual covid-19 pandemic is growing.
Given the promising results presented in this post, we will now refine our methods. In addition to the statistical algorithms we are employing at the moment, we will try to implement other methods that fall into the category of pattern recognition. We hope to decrease the number of false positive detections and the impact of climate and temperature that induce modulations in our detection curves that are not systematic but rather induced by external factors. We will keep you posted about our progress.