Enhancing the accuracy of COVID-19 incidence and mortality predictions using Google Trends data across the 50 US states and the District of Columbia

Abstract

Quick and accurate forecasts of incidence and mortality trends for the near future are particularly useful for the immediate allocation of available public health resources, as well as for understanding the long-term course of the pandemic. The surveillance data used for predictions, however, may come with some reporting delays. Consequently, auxiliary data sources that are available immediately can provide valuable additional information for recent time periods for which surveillance data have not yet become fully available. In this work, a set of Google search queries by individual users related to COVID-19 incidence and mortality is collected and analyzed. The information from these queries aims to improve quick forecasts. Initially, the identified search query keywords were ranked according to their predictive abilities with reported incidence and mortality. After that, the ARIMA, Prophet, and XGBoost models were fitted to generate forecasts using only the available reported incidence and mortality (baseline model) or together with combinations of searched keywords identified based on their predictive abilities (predictors model). In summary, the inclusion of top-ranked keywords as predictors significantly enhanced prediction accuracy for the majority of scenarios in the range from 50% to 90% across all considered models and is recommended for future use. The inclusion of low-ranked keywords did not provide such an improvement. In general, the ranking of predictors and the corresponding forecast improvements were more pronounced for incidence, while the results were less pronounced for mortality.