Publicatie
Publication date
Using web search queries to monitor influenza-like illness: an exploratory retrospective analysis, Netherlands, 2017/18 influenza season.
Schneider, P.P., Gool, C.J.A.W. van, Spreeuwenberg, P., Hooiveld, M., Donker, G.A., Barnett, D.J., Paget, J. Using web search queries to monitor influenza-like illness: an exploratory retrospective analysis, Netherlands, 2017/18 influenza season. Eurosurveillance: 2020, 25(21), p. 10 p..
Read online
Background
Despite the early development of Google Flu Trends in 2009, standards for digital epidemiology methods have not been established and research from European countries is scarce.
Aim
In this article, we study the use of web search queries to monitor influenza-like illness (ILI) rates in the Netherlands in real time.
Methods
In this retrospective analysis, we simulated the weekly use of a prediction model for estimating the then-current ILI incidence across the 2017/18 influenza season solely based on Google search query data. We used weekly ILI data as reported to The European Surveillance System (TESSY) each week, and we removed the then-last 4 weeks from our dataset. We then fitted a prediction model based on the then-most-recent search query data from Google Trends to fill the 4-week gap (‘Nowcasting’). Lasso regression, in combination with cross-validation, was applied to select predictors and to fit the 52 models, one for each week of the season.
Results
The models provided accurate predictions with a mean and maximum absolute error of 1.40 (95% confidence interval: 1.09–1.75) and 6.36 per 10,000 population. The onset, peak and end of the epidemic were predicted with an error of 1, 3 and 2 weeks, respectively. The number of search terms retained as predictors ranged from three to five, with one keyword, ‘griep’ (‘flu’), having the most weight in all models.
Discussion
This study demonstrates the feasibility of accurate, real-time ILI incidence predictions in the Netherlands using Google search query data.
Despite the early development of Google Flu Trends in 2009, standards for digital epidemiology methods have not been established and research from European countries is scarce.
Aim
In this article, we study the use of web search queries to monitor influenza-like illness (ILI) rates in the Netherlands in real time.
Methods
In this retrospective analysis, we simulated the weekly use of a prediction model for estimating the then-current ILI incidence across the 2017/18 influenza season solely based on Google search query data. We used weekly ILI data as reported to The European Surveillance System (TESSY) each week, and we removed the then-last 4 weeks from our dataset. We then fitted a prediction model based on the then-most-recent search query data from Google Trends to fill the 4-week gap (‘Nowcasting’). Lasso regression, in combination with cross-validation, was applied to select predictors and to fit the 52 models, one for each week of the season.
Results
The models provided accurate predictions with a mean and maximum absolute error of 1.40 (95% confidence interval: 1.09–1.75) and 6.36 per 10,000 population. The onset, peak and end of the epidemic were predicted with an error of 1, 3 and 2 weeks, respectively. The number of search terms retained as predictors ranged from three to five, with one keyword, ‘griep’ (‘flu’), having the most weight in all models.
Discussion
This study demonstrates the feasibility of accurate, real-time ILI incidence predictions in the Netherlands using Google search query data.
Background
Despite the early development of Google Flu Trends in 2009, standards for digital epidemiology methods have not been established and research from European countries is scarce.
Aim
In this article, we study the use of web search queries to monitor influenza-like illness (ILI) rates in the Netherlands in real time.
Methods
In this retrospective analysis, we simulated the weekly use of a prediction model for estimating the then-current ILI incidence across the 2017/18 influenza season solely based on Google search query data. We used weekly ILI data as reported to The European Surveillance System (TESSY) each week, and we removed the then-last 4 weeks from our dataset. We then fitted a prediction model based on the then-most-recent search query data from Google Trends to fill the 4-week gap (‘Nowcasting’). Lasso regression, in combination with cross-validation, was applied to select predictors and to fit the 52 models, one for each week of the season.
Results
The models provided accurate predictions with a mean and maximum absolute error of 1.40 (95% confidence interval: 1.09–1.75) and 6.36 per 10,000 population. The onset, peak and end of the epidemic were predicted with an error of 1, 3 and 2 weeks, respectively. The number of search terms retained as predictors ranged from three to five, with one keyword, ‘griep’ (‘flu’), having the most weight in all models.
Discussion
This study demonstrates the feasibility of accurate, real-time ILI incidence predictions in the Netherlands using Google search query data.
Despite the early development of Google Flu Trends in 2009, standards for digital epidemiology methods have not been established and research from European countries is scarce.
Aim
In this article, we study the use of web search queries to monitor influenza-like illness (ILI) rates in the Netherlands in real time.
Methods
In this retrospective analysis, we simulated the weekly use of a prediction model for estimating the then-current ILI incidence across the 2017/18 influenza season solely based on Google search query data. We used weekly ILI data as reported to The European Surveillance System (TESSY) each week, and we removed the then-last 4 weeks from our dataset. We then fitted a prediction model based on the then-most-recent search query data from Google Trends to fill the 4-week gap (‘Nowcasting’). Lasso regression, in combination with cross-validation, was applied to select predictors and to fit the 52 models, one for each week of the season.
Results
The models provided accurate predictions with a mean and maximum absolute error of 1.40 (95% confidence interval: 1.09–1.75) and 6.36 per 10,000 population. The onset, peak and end of the epidemic were predicted with an error of 1, 3 and 2 weeks, respectively. The number of search terms retained as predictors ranged from three to five, with one keyword, ‘griep’ (‘flu’), having the most weight in all models.
Discussion
This study demonstrates the feasibility of accurate, real-time ILI incidence predictions in the Netherlands using Google search query data.