COVID-19 hospitalization forecasts using internet search data

[ad_1]

We focused on national hospital admission predictions and state-level predictions of 51 states in the United States, including Washington DC. The inputs consist of confirmed incremental cases, percentage of vaccinated population, confirmed new hospital admissions, and Google search query frequencies. Both state-level data and national data were directly obtained from respective data sources outlined in this section. Our prediction method is inspired by ARGO9, with details presented in this section as well.

Data/code availability

All data used in this study are publicly available from respective data sources outlined here. For completeness, the datasets analyzed during the current study and the code are also deposited in the Harvard dataverse repository, DOI: 10.7910/DVN/S7HOTD. All analyses, including the generation of all figures, were performed with the R statistical software20, version 4.1.1 (https://www.R-project.org/).

COVID-19 related data

We used reported COVID-19 confirmed incremental cases from JHU CSSE data21, percentage of fully vaccinated population from Centers for Disease Control and Prevention (CDC)22 which is averaged among percentages of fully vaccinated population in all states and COVID-19 confirmed new hospital admissions from HHS17. The data sets were collected from July 15, 2020 to January 15, 2021.

Figure 3

Left: National weekly COVID-19 new hospitalizations (black), national weekly COVID-19 confirmed Cases (blue) and national percentage of vaccinated population (red), scaled accordingly. Right: COVID-19 weekly new hospitalizations (red) and the top 3 Google queries (optimal lagged) that have largest correlation coefficients with hospitalizationsm: how long contagious (blue), loss of smell (red) and loss of taste (green).

Table 1 National level comparison error metrics. Table 2 State-level comparison error metrics. Google search data

Google Trends provides estimated Google search frequency for the specified query term23. We obtained online search data from Google Trends23 for the period from July 15, 2020 to January 15, 2021. To retrieve the time series search frequencies of a desired query, one needs to specify the querys geographical information and time frame on Google Trends. The returned frequency from Google Trends is obtained by sampling all raw Google search frequencies that contain this query23. The detailed data collection procedure and subsequent data pre-processing (introduced in sections below) are illustrated in the flowchart (Figure S1). In step 1 (green highlighted boxes in Figure S1), to curate the pool of potentially predictive queries, we first started with 129 influenza (flu) related queries based on prior studies9,24,25. Then, we changed Influnza and Flu keywords to Coronavirus and COVID-19, respectively. We also added in additional COVID-19 specific search terms from Google Trends Coronavirus Story page26. Lastly, for each query, we also included its top related queries and topics based on Google Trends website23. Finally, we end up with 256 COVID-19 related queries (Table S1). Next two sections will describe the subsequent data cleaning and pre-processing in detail, illustrated in step 2 and step 3 of Figure S1.

Inter-quantile range (IQR) filter and optimal lags for Google search data

The raw Google search frequencies obtained from Google Trends23 are observed to be unstable and sparse19. Such instability and sparsity can negatively affect prediction performance of linear regression models which are sensitive to outliers. To deal with such outliers in Google search data, we used an IQR filter19 to remove and correct outliers on a rolling window basis. The search data that is beyond 3 standard deviation from the past 7-day average are examined and removed19, which is also shown in step 2 (orange highlighted box) first sub-step in Figure S1.

The trends of Google frequencies search are often a few days ahead of hospitalizations, indicating that the search data might contain predictive information about hospitalizations. Figure 4 demonstrates the delay behavior between Google search query frequencies and national hospitalizations. To fully utilize the predictive information in national Google search terms, we found and applied optimal lags19 to filtered Google search frequencies to match the trends of national hospitalizations. For each query, a linear regression of COVID-19 new hospitalizations is fitted against lagged Google search frequency, considering a range of lags (from 4 to 35). The lag results in lowest mean square error is selected as the optimal lag for that query. The data used to find optimal lags are from August 1, 2020 to December 31, 2020, and this is shown in the second sub-step in step 2 (Figure S1).

Figure 4

Google search query how long contagious and COVID-19 weekly new hospitalizations Illustration of delay in peak between Google search query search frequencies (how long contagious in blue) and COVID-19 national level weekly new hospitalizations (red). Y-axis are adjusted accordingly.

Selection of google search terms

After applying optimal lags to the 256 COVID-19 related terms, we further selected the queries that have correlation coefficients with national COVID-19 hospitalizations larger than 0.5 for the period from August 1, 2020 to December 31, 2020. We applied 7-day moving average to further smooth out weekly fluctuations in the selected Google search queries. All three procedures above are shown in step 3 in Figure S1, serving as the final step of the entire data pre-processing procedure. Table 3 shows the 11 selected important terms plus their optimal lags. Table S2 displays the correlation coefficients of the optimally delayed 11 important Google search queries. Table 3 supports the intuition that when people get infected, they would search for general query like symptoms of the covid-19 first as this query has relatively large optimal lag. After the symptoms develop, people might begin to search for specific symptoms such as loss of smell which has relatively smaller optimal lag. Seemingly similar query terms can have slightly different search patterns, and thus different optimal lags, as discussed in the Descriptive analysis of the input data section.

Table 3 Optimal lags of selected important terms.ARGO inspired prediction

Let ({hat{y}}_{t,r}) be the daily hospital admissions of region r on day t; (X_{k,t}) be the Google search data of term k on day t; (c_{t,r}) be the JHU COVID-19 incremental confirmed cases on day t of region r; (v_{t,r}) be the cumulative percent of people who get vaccinated by day t of region r; ({mathbb {I}}_{{t, d}}) be the weekday indicator for day t (ie ({mathbb {I}}_{{t, 1}} ) indicates day t being Monday). Standing on day T, to predict l-day-ahead hospital admission of state r, ({hat{y}}_{T+l,r}), we used penalized linear estimator as following:

$$begin{aligned} begin{aligned} {hat{y}}_{T+l,r} = {hat{mu }}_{y,r}+sum ^{I}_ {i=0}{hat{alpha }}_{i,r}y_{Ti,r} + sum _{jin {J}}{hat{beta }}_{j,r }c_{T+lj,r}+ sum _{min {M}_{r}}{hat{gamma }}_{m,r}y_{T,m}+ sum _{ qin {Q}}{hat{phi }}_{q,r}v_{T+lq,r}+ sum ^{K}_{k=1}{hat{delta }} _{k,r}X_{k,T+l-{hat{O}}_k} + sum ^6_{d=1}{hat{tau }}_{d,r}{mathbb {I}}_{{T+l, d}} end{aligned} end{aligned}$$

(1)

where (I=6) considering consecutive one week lagged daily hospital admissions; (J=max left( {7,28},lright)), considering lagged confirmed cases; ({M}_{r}) is the set of geographical neighboring states of state r; (Q=max left( 7,lright)), considering vaccination data lagged by one week; ({hat{O}}_k=max left( O_k,lright)) is the adjusted optimal lag for term k; (K=11), considering 11 selected Google search terms. The coefficients for l-day-ahead predictions of region r, ({mu _{y,r},varvec{alpha }=(alpha _{1,r},ldots ,alpha _{ 6,r}), varvec{beta }=(beta _{1,r}, ldots , beta _{|J|,r}), varvec{gamma }=(gamma _{ 1,r},ldots ,gamma _{|{{M}_{r}}|,r}), varvec{phi }=phi _{max(7,l),r}, varvec{delta }=(delta _{1,r},ldots ,delta _{11,r}), varvec{tau }=(tau _{1,r}, ldots , tau _{6,r})}), were computed by

$$begin{aligned} begin{aligned} underset{mu _{y,r},varvec{alpha },varvec{beta },varvec{gamma },varvec{phi },varvec{delta },varvec{tau },varvec{lambda }}{mathrm {argmin}} sum _{t=TM-l+1}^{Tl}&omega ^ {Tl-t+1}Bigg ( y_{t+l,r}-mu _{y,r} – sum ^{6}_{i=0}{alpha }_{i,r} y_{ti,r}-sum _{jin {J}}{hat{beta }}_{j,r}c_{t+lj,r}-sum _{min {M }_{r}}{hat{gamma }}_{m,r}y_{t,m}\ ;;;&- sum _{qin {Q}}{hat {phi }}_{q,r}v_{t+lq,r} -sum ^{5}_{k=1}{hat{delta }}_{k,r}X_{k, t+l-{hat{O}}_k} – sum ^6_{d=1}{hat{tau }}_{d,r}{mathbb {I}}_{{t+ l, d}}Bigg )^2\ ;;;&+ lambda _alpha Vert varvec{alpha }Vert _1+lambda _beta Vert varvec{beta } Green _1+lambda _gamma Green varvec{gamma }Green _1+ lambda _phi Green varvec{phi }Green _1+lambda _delta Green varvec{delta }Green _1 +lambda _tau Vert varvec{tau }Vert _1 end{aligned} end{aligned}$$

(2)

M = 56 which is the length of our training period; (omega = 0.8) is the exponentially time-decaying weight which assigns higher weight on more recent observation. Region (varvec{r}) consists of US and its 51 states, including Washington DC. For US national level training, the hospitalizations of neighboring states, (y_{t,m}), and their coefficients, (varvec{gamma }), are excluded. To address the sparsity of Google search data, we used penalty of L1-norm. For simplicity, the hyperparameters (varvec{lambda }=(lambda _{alpha },lambda _{beta },lambda _{gamma },lambda _{phi },lambda _ {delta },lambda _{tau })) for L1-norm penalties were set to be equal and obtained via 10-folds cross-validation.

With the formulation above, on each Monday from January 4, 2021 to December 27, 2021, we iteratively trained our model and made national and state-level retrospective out-of-sample hospitalizations predictions up to 14 days into future. We then aggregated daily predictions into one-week-ahead and two-week-ahead predictions. For example, ({hat{y}}_{T+1:T+7,r}=sum ^7_{i=1}{hat{y}}_{T+i,r} ) and ({hat{y}}_{T+8:T+14,r}=sum ^{14}_{i=8}{hat{y}}_{T+i,r }) are the 1-week-ahead prediction and 2-week-ahead prediction on day T of region r, respectively.

Evaluation metrics

Root Mean Squared Error (RMSE) between a hospitalization estimate ({hat{y}}_t) and the true value (y_t) over period (t=1,ldots , T) is ( sqrt{frac{1}{T}sum _{t=1}^T left( {hat{y}}_t – y_tright) ^2}). Mean Absolute Error (MAE) between an estimate ({hat{y}}_t) and the true value (y_t) over period (t=1,ldots , T) is (frac{ 1}{T}sum _{t=1}^T left|{hat{y}}_t – y_tright|). Correlation is the Pearson correlation coefficient between (hat{varvec{y}}=({hat{y}}_1, dots , {hat{y}}_T)) and (varvec{y }=(y_1,dots , y_T)). All estimates ({hat{y}}_t) and the true value (y_t) were weekly aggregated before calculating RMSE, MAE and Cor.

Ethics approval and consent to participate

This study did not involve human participants, data, or tissue. It was conducted using only aggregated and anonymized data. Institutional review board approval was not required. All methods were carried out in accordance with relevant guidelines and regulations.

[ad_2]