COVID-19 hospitalizations forecasts using internet search data

Abstract

As the COVID-19 spread over the globe and new variants of COVID-19 keep occurring, reliable real-time forecasts of COVID-19 hospitalizations are critical for public health decisions on medical resources allocations. This paper aims to forecast future 2 weeks national and state-level COVID-19 new hospital admissions in the United States. Our method is inspired by the strong association between public search behavior and hospitalization admissions and is extended from a previously-proposed influenza tracking model, AutoRegression with GOogle search data (ARGO). Our LASSO-penalized linear regression method efficiently combines Google search information and COVID-19 related time series information with dynamic training and rolling window prediction. Compared to other publicly available models collected from COVID-19 forecast hub, our method achieves substantial error reduction in a retrospective out-of-sample evaluation from Jan 4, 2021, to Dec 27, 2021. Overall, we showed that our method is flexible, self-correcting, robust, accurate, and interpretable, making it a potentially powerful tool to assist healthcare officials and decision making for the current and future infectious disease outbreaks.

Using Internet Search Data to Forecast COVID-19 Trends: A Systematic Review

“…None of them examined in detail the importance of the search queries with a carefully designed large-scale data-driven query identification process or attempted to forecast COVID-19 incidences and future outbreaks by incorporating the aforementioned prediction techniques. While the majority of the existing studies proclaim the potential predictive power of internet search data in COVID-19 forecasting through the corresponding correlation analyses, only a few have built prediction models [45][46][47][48][49][50][51][52][53] to fully utilize and demonstrate the predictive power of internet search data. This review study examines each of those prediction models, by elaborating the relevant internet search data as well as other alternative data, and analyzing the prediction models.…”

“…Lampos et al [49] determined the list of search terms by COVID-19-related symptoms and keywords, and computed the time lags between the search frequencies and COVID-19 confirmed cases and deaths, as inputs to their forecasting model. Ma and Yang [51], Wang et al [52], and Ma et al [53] developed an end-to-end data-driven selection mechanism that selected 23 important search queries from the 256 top searched COVID-19-related Google search terms, by ranking the Pearson correlation coefficient between optimal lagged search term and COVID-19 trends (forecasting target), with a cutoff threshold of 0.5, using summer 2020 as the training period. The selected 23 important search terms contained more specific COVID-19-related symptoms, such as "loss of taste" and "loss of smell", compared to other studies.…”

“…Liu et al [45] employed a 2-day moving average in all the selected queries to enhance signal and reduce noise, while Rabiolo et al [48] conducted principle component analysis (PCA) and used two principle components from all the selected queries for subsequent forecasts. On the other hand, Ma and Yang [51], Wang et al [52], and Ma et al [53] first addressed the high sparsity in the state-level data through a weighted average of state-level and regional or national-level search queries' frequencies, and incorporated an interquantile range (IQR)-based data filtering mechanism to further smooth out extreme outliers due to sudden spikes/drops. Wang et al [52] further applied a 7-day moving average to remove the weekly seasonality in the selected Google search queries, before employing them in the hospitalizations forecasts.…”

“…On the other hand, Ma and Yang [51], Wang et al [52], and Ma et al [53] first addressed the high sparsity in the state-level data through a weighted average of state-level and regional or national-level search queries' frequencies, and incorporated an interquantile range (IQR)-based data filtering mechanism to further smooth out extreme outliers due to sudden spikes/drops. Wang et al [52] further applied a 7-day moving average to remove the weekly seasonality in the selected Google search queries, before employing them in the hospitalizations forecasts. Lampos et al [49] took a step further by not only filtering out the noise in the search queries' frequencies but also the news media effects led by public fear.…”

“…In addition, one can incorporate L 1 -norm penalty into an ARMA structure to achieve model selection while enjoying the interpretability and simplicity from the linear structure. Ma and Yang [51], Wang et al [52], and Ma et al [53] proposed an autoregression-like model that combines COVID-19 time series information and Google search data with L 1 -norm penalty for U.S. national COVID-19 forecasts. They demonstrated that past COVID-19 time series and Google search information efficiently complement each other in a rolling-window forecasting manner.…”