Study population and setting
Geo-located and self-reported temperature data from more than 1 million smartphone-connected personal thermometers managed by Kinsa, Inc. were anonymized and aggregated at the US county-level. Temperature readings were used to construct and track influenza-like illness (ILI) signals and generate county-specific ILI-incidence forecasts from the beginning of March (before widespread outbreaks were underway) to mid-May 2020. Forecasts were based on day-of-year reproduction numbers (R) estimated back to August 1, 2016, and forecasts were made from March 1, 2020 using R estimates for that day-of-year. However, it was unclear what time frame of ILI data from thermometer readings were used for this back calculation. Influenza forecasts with lower levels of transmission due to the assumed impact of “shelter-in-place” social distancing measures were also produced to assess levels of potential bias. Authors compared this ILI forecast to the real-time thermometer-based ILI incidence at the city-level to identify “anomalous” ILI incidence which was defined as the difference between the real-time thermometer ILI and the upper uncertainty bound of the ILI forecast. Authors estimated the probability that this anomalous incidence was driven by normal seasonal influenza and assessed whether these detected anomalous incidence were indicative of COVID-19 cases.
Summary of Main Findings
Results show that self-reported temperature or ILI data can be used to produce forecasts of ILI incidence comparable to other influenza models for the first few weeks. ILI anomalies correlated strongly to county- and state-level positive COVID-19 case counts. County-level ILI anomalies also corresponded spatially to areas where major outbreaks have occurred including the Seattle Area, San Francisco Bay Area, New York Metro Area, and Florida. However, anomalous ILI incidence was not observed in Minnesota, Wisconsin, and South Dakota where COVID-19 cases have been confirmed. Authors found that ILI forecasts were likely sensitive to any social distancing measures implemented which could impact influenza transmission and affect whether ILI anomalies would occur.
Authors re-ran their influenza forecasts assuming that “shelter-in-place” orders and social distancing reduced influenza transmission by different amounts which could bias their estimates of anomaly incidence. The study uses county-level real-time syndromic data which has the potential to reflect changes in case numbers in a more timely manner.
As the study relies on self-reported data from smartphone-connected thermometers, it is unclear how representative the real-time ILI incidence used in this study is. For example, users of the Kinsa Inc product may only represent a certain age-group or there may not be any or a very limited number of users in certain counties or states. Only the total COVID-19 case counts have been used to validate the ILI anomalies making it difficult to assess whether these anomalies are indicative of COVID-19 case counts over time. It is not clear from the methods or the supplementary material exactly how, or on what time frame of Kinsa data the ILI forecasts were derived.
By using temperature data from a network of over 1 million users across the US, the study demonstrates the potential use of self-reported syndromic data to identify potential hotspots of COVID-19 transmission. This may provide an early-warning system to locate areas where an outbreak response is needed. Such syndromic surveillance for emerging pathogens can be valuable in settings with strong surveillance in place for other diseases with similar symptoms (in this case influenza and respiratory symptoms) and settings where a lack of diagnostic capacity may lead to under-reporting.
This review was posted on: 18 July 2020