Tue, Aug 16, 2022: 10:15 AM-10:30 AM
516A
Background/Question/Methods
Google Trends data have been used in predictive models for many infectious diseases including Zika, salmonellosis, encephalitis, and Lyme disease. Lyme disease incidence has increased in the US and there is a documented geographic expansion of cases. A previous study found that there were similar patterns between Google searches for Lyme disease and symptoms and Lyme disease incidence at the state-level. However, the authors noted that validation of the method is needed due to non-specific symptoms of Lyme that present with other conditions, such as arthritis and multiple sclerosis. Therefore, the objective of this study was to validate the use of Google Trends search data for predicting Lyme disease incidence at the state-level by comparing models with varying search terms. Monthly state-level Lyme disease incidence data from 2010-onward were requested from state public health departments. Preliminary results were generated from 3 states. Google Trend data was downloaded using the ‘gtrends’ package in R version 4.0.2. Time series models were built using random-effects generalized least squares (GLS) regression using the ‘xtreg’ command in Stata version 17.0. Monthly lags of searches were used as predictors (i.e., 1-month prior, 2-months prior) until statistical insignificance was achieved.
Results/Conclusions
In total, there were 360 observations from 3 states (CT, RI, and WA) used in preliminary analyses. Adjusting for season, the models indicated that the search terms most predictive of Lyme disease incidence included “Lyme,†“bone pain,†“tick bite,†and “tick rash.†The R2 for the most predictive models ranged from 19% to 69%. Interestingly, the predictive ability of terms for Lyme disease (“Lyme†versus “Lyme disease†versus “Lymesâ€) varied greatly. Other searches for conditions with symptoms similar to Lyme disease were not predictive of Lyme disease incidence, including “arthritis,†“bells palsy,†“chronic fatigue,†and “multiple sclerosis.†In addition, the search volume for a different tick-borne disease (“rocky mountain spotted feverâ€) was not predictive of Lyme disease. These results highlight that Google trends search data have potential as a predictive tool for Lyme disease incidence and show specificity for Lyme disease incidence. However, there is a need for additional testing and standardization of search terms to ensure the models optimize the use of search data, are generalizable to other states, and are reproducible.
Google Trends data have been used in predictive models for many infectious diseases including Zika, salmonellosis, encephalitis, and Lyme disease. Lyme disease incidence has increased in the US and there is a documented geographic expansion of cases. A previous study found that there were similar patterns between Google searches for Lyme disease and symptoms and Lyme disease incidence at the state-level. However, the authors noted that validation of the method is needed due to non-specific symptoms of Lyme that present with other conditions, such as arthritis and multiple sclerosis. Therefore, the objective of this study was to validate the use of Google Trends search data for predicting Lyme disease incidence at the state-level by comparing models with varying search terms. Monthly state-level Lyme disease incidence data from 2010-onward were requested from state public health departments. Preliminary results were generated from 3 states. Google Trend data was downloaded using the ‘gtrends’ package in R version 4.0.2. Time series models were built using random-effects generalized least squares (GLS) regression using the ‘xtreg’ command in Stata version 17.0. Monthly lags of searches were used as predictors (i.e., 1-month prior, 2-months prior) until statistical insignificance was achieved.
Results/Conclusions
In total, there were 360 observations from 3 states (CT, RI, and WA) used in preliminary analyses. Adjusting for season, the models indicated that the search terms most predictive of Lyme disease incidence included “Lyme,†“bone pain,†“tick bite,†and “tick rash.†The R2 for the most predictive models ranged from 19% to 69%. Interestingly, the predictive ability of terms for Lyme disease (“Lyme†versus “Lyme disease†versus “Lymesâ€) varied greatly. Other searches for conditions with symptoms similar to Lyme disease were not predictive of Lyme disease incidence, including “arthritis,†“bells palsy,†“chronic fatigue,†and “multiple sclerosis.†In addition, the search volume for a different tick-borne disease (“rocky mountain spotted feverâ€) was not predictive of Lyme disease. These results highlight that Google trends search data have potential as a predictive tool for Lyme disease incidence and show specificity for Lyme disease incidence. However, there is a need for additional testing and standardization of search terms to ensure the models optimize the use of search data, are generalizable to other states, and are reproducible.