Summary: To better account for the effects of widespread social distancing, we updated our expected illness curve on Healthweather.us today. Our cumulative atypical illness map has also changed as a result. The new map correlates even better with COVID-19 cases and deaths than before. We’re also adding a new mode to the map, real-time atypical, which shows real-time levels of atypical illness.
Kinsa’s atypical illness early warning system has correctly identified the location and intensity of COVID-19 outbreak epicenters on average 14 days before the first COVID-19 death. However, now that social distancing directives have been implemented across the United States, we have observed a drop in overall illness levels nationwide. These drops coincide with the application of social distancing directives.
With COVID-19 case rates peaking or declining, many states are relaxing social distancing measures. There is an increasingly urgent need to detect possible COVID-19 resurgence as a result. Here, we present an updated approach to our atypical illness detection method where we account for social distancing on influenza transmission rates, such that we can identify new atypical illness levels when measures are relaxed.
The example of San Francisco, CA illustrates why we should incorporate social distancing into our forecasts, and exactly how this works. Figure 1 shows the history of atypical illness detections in San Francisco, where we observed a brief period of atypical illness levels during the onset of COVID-19 spread, followed by a drastic reduction of illness on March 16th, when schools were closed and shelter-in-place was mandated. What is most interesting is that following shelter-in-place, illness dropped well below the expected influenza levels for this time of year. In fact, shelter-in-place appears to have cut off most circulating illness transmission, and current influenza-like illness levels are near zero. This is important in that it shows community efforts are working.
It was clear that we needed to account for social distancing effects in our atypical illness detection. (You can read more about our atypical illness detection method here, and the approach accounting for social distancing is explained in detail within our preprint MedRxIV manuscript, under review at a leading scientific journal.) Briefly, we update our expected influenza forecast as follows:
- Reduce influenza transmission by a constant rate following social distancing policy interventions and behavioral changes (i.e. people choosing to avoid crowds)
- Flag anomalous illness levels in reference to the expected influenza levels with social distancing taken into account.
Let’s return to the San Francisco example to see how implementing these changes will work. Given that shelter-in-place and school closures occurred on March 16th, 2020, we can reduce influenza transmission rates for all days thereafter. In Figure 2, we observe a strong reduction in expected influenza levels following social distancing directives, which also leads to an increased period of atypical illness identifications. Additionally, our expected influenza levels are now quite similar to the illness levels we see today. This demonstrates that our actual data matches the drop in the influenza forecast by including social distancing. This suggests that social distancing does lead to a strong reduction in influenza transmission. Given that our expected influenza levels are now near-zero, if/when social distancing directives are lifted, Kinsa will be able to identify abnormally high levels as real-time illness exceeds this new range of expected influenza.
We can then take the approach outlined above and apply it to the entire country. Here, we make two simplifying assumptions: 1) Social distancing influences influenza transmission in all regions starting on March 17th, the most common date of school closures across the US, and 2) Distancing reduces influenza transmission by 25% for all days forward. These assumptions are supported by the observation of widespread declines in illness levels even with variable reductions in mobility patterns across the US; however, the magnitude of the reduction likely varies with policy implementation and adherence. Going forward, we will work to improve these estimates based on our own research and that published in academic journals.
Our county map of cumulative anomaly fevers is greatly improved by incorporating social distancing impacts on influenza transmission. We now capture atypical illness hotspots in Los Angeles (Figure 3), a region known to be a current COVID-19 hotspot. By accounting for social distancing Kinsa can more accurately track the spread of COVID-19 in communities, and these changes are live on healthweather.us. We have also added a real-time atypical mode to our national illness map, to capture new atypical illness where we observe it.
As shown in the San Francisco example, we are well-positioned to identify new outbreaks in real-time if illness levels begin to rise as social distancing directives are lifted. If real-time illness levels begin to exceed the expected range of social distancing-influenced influenza, real-time illness will be flagged as atypical. The Kinsa data team continues to tackle these problems and improve our approaches, in collaboration with academic researchers, to track and respond to the COVID-19 epidemic.