In the last post Feverish we presented the first national “fever curve” based on and computed from the donors' data. In this post, we not only want to give you an update on the national fever curve, but also want to show the results we obtained for the individual federal states in Germany.
Recall: In this project, we evaluate the time series of daily averages of resting heart rate and daily step counts of individual donors. When, if compared to individual baseline values in both quantities, the resting heartrate increases and remains elevated for a few days while simultaneously the step count drops and remains low for the same number of days, that anomaly is an indication that a person is sick and has developed a fever. Here’s a figure from the last post that illustrates a typical candidate for a positive detection:
The deviations from individual baselines are computed for all donors with sufficiently long data streams and are automatically identified with the detection algorithms we developed.
Update: Fever curve for Germany
The following figure depicts the fever curve for Germany as measured by our detection algorithms. The way detection works is explained in detail in the post Feverish. Currently we are using two different detection algorithms (labeled A and B) that are based on two different measures of statistical deviations from the baseline central tendancy. Both algorithms produce functionally similar curves.
In the time period from May 26th - Aug. 3rd, 2020 we measured the detection rate, so the fraction of detections with respect to the analyzed donor cohort. The detection rate is typically in the range of 0.01-0.06 so between 1% - 6% are “diagnosed” with fever. We see two different important qualitative features in the curves. First, we see clear temporal modulations in the curves with a typical period of around 2 weeks. We were able to find out that these modulations correlate with variations in outside temperature that exhibit approximately the same periodicity. We will discuss this topic in one of the next blog posts.
More important, however, are the trend lines. In the trends that ignore the short time variations, we see a systematic increase in fever detections in July. Or course, we cannot conclude from this that this increase is caused by COVID-19 cases. Because we “only” measure fever symptoms and not COVID-19 symptoms directly, the observed trend can also be caused by other infections of diseases that yield fever symptoms. This is why it is important to compare the measured curves to results obtained by other surveillance systems. This is explained in detail in the previous blog post Feverish.
Fever curves of federal states in Germany
By now we can compute the fever curves for the 16 federal states in Germany individually. Our goal is to fine-tune the geographic resolution and compute fever curves on a county and city level. At the moment, however, we do not see a sufficiently large number of detections on a county level.
The next figure depicts the fever curves for individual federal states.
Interestingly, the fever curves of individual states do not vary much apart from slight variations. All curves follow the overall national trend.
This effect can be seen more distuinctly the fever curves of all 16 states are depicted in one figure. This is shown in the figure below, which also depicts the national aggregated fever curve for comparison.
Currently we are working on an automated pipeline for producing an online fever curve monitor that is automaticcally updated every day. Although this sounds simple, the process actually requires quite a number of computational steps that must be aligned and implemented. So bear with us. We need a few more days.