Methods: We used temperature readings collected by the Kinsa smart-thermometer and mobile device app to develop state-level forecasting models to predict real-time influenza activity (1-2 weeks in advance of surveillance reports). We used state-reported influenza-like illness (ILI) to represent state influenza activity for 48 US states with sufficient surveillance data. Counts of temperature readings, fever episodes and reported symptoms were computed by week. We developed autoregressive time-series models and evaluated model performance in an adaptive out-of-sample manner. We compared baseline time-series models containing lagged state-reported ILI activity to models incorporating exogenous thermometer readings.
Results: A total of 10,262,212 temperature readings were recorded from October 30, 2015 to March 29, 2018. In nearly all of the 48 states considered, weekly forecasts of ILI activity improved considerably when thermometer readings were incorporated. On average, state-level forecasting accuracy improved by 23.9% compared to baseline time-series models. In many states, such as PA, NM, MA, VA, NY and SC, out-of-sample forecast error was reduced by more than 50% when thermometer data was incorporated. In general, forecasts were most accurate in states with the greatest number of device readings. During the 2017-2018 influenza season, the average improvement in forecast accuracy was 24.4%, and thermometer readings improved forecasting accuracy in 41, out of 48, states.
Conclusion: Data from smart thermometers accurately track real-time influenza activity at a state level. Local surveillance efforts may be improved by incorporating such information. Such data may also be useful for longer-term local forecasts.
S. Pilewski, Kinsa Inc.: Employee and Shareholder , equity received and Salary .
V. Petrovic, Kinsa Inc.: Employee and Shareholder , equity received and Salary .
P. M. Polgreen, None