During the past few weeks we have been updating our Fever Monitor on a daily basis, based on the heart rate and activity data measured by the wearable devices and smart watches of more than 520.000 donors.
Some of you may have noticed that the fever curve changed significantly a few days ago. Noteably, the curves now appear smoother, without the charactistic up-and-down modulations that were previously seen on a weekly or bi-weekly basis. The reason behind this change is that we are now using an improved version of our fever detection algorithm for the Monitor.
During the past weeks, we have been working hard to refine the detection algorithm. One of our key goals was to account for the impact that weather and temperature have on heart rate and activity, and subsequently the performance of the algorithm, by removing the influence of seasonal variations.
Because the COVID-19 case counts in Germany follow the improved fever detections substantially better, we would like to take the opportunity to share in what ways we improved the algorithm.
Detektions and Case Counts
First, let’s have a look at the case counts compared to the detection curves for Germany and its federal states identified by the improved algorithm.
We found that the detections generated by the data donation project predict the temporal evolution of the case counts remarkably well. We can see that in the decline part of the first have during spring, the constant plateau during the summer, and the beginning to the second wave this fall.
We can also see this for the individual states, but not to the same extent because here the detection numbers are so low that statistical fluctuations dominate the signal.
What did we change?
During the past months, we have been working continuously on improving our methods. We always check the performance of these methods by comparing to different data sources and applying a variety of quality checks. For example, we use the RKI data on influenza like illnesses as a benchmark in addition to COVID-19 related data, such as the case counts, positive test rates and other measures.
We previously discussed the impact of weather on heart rate in the Blog Post Great weather quickens the heart. The goal is to account for this impact and not have the results skewed by it.
Our original method is derived from a number of publications that dealt with detecting fever during flu season. One of the fundamental differences between these studies is that we have a lot more data in terms of the total length of time of more than 5 months. Over this period, seasonal effect begin to play an increasingly large role. The magnitude of the effects were unavailable when we started in spring 2020.
In the improved algorithm, we investigate the vital data of each donor not as an absolute value, but rather relative to the data of all donors on that day. We then compare these values with the average values of the donor, the donor’s so called baseline. This way our algorithm automatically adapts to external factors like weather and climate and the impact of these on the detection curve is substantially reduced.
Calculating the Baseline
We also changed the method in which we calculate the donors’ baseline. In the previous iteration of the algorithm, we computed the baseline from 28 days prior to the reference day. Originally, with a much shorter timeframe and minimal seasonal effects, we had some success in eliminating minor external effects. But as our dataset expanded over several seasons, this approach proved unable to effectively account for substantial seasonal effects. Now, with the new method we use the data from the entire time window to compute the baseline, making it more stable as time goes on and less susceptible to external factors.