Impact of Demographic and Clinical Subgroups in Google Trends Data: Infodemiology Case Study on Asthma Hospitalizations.

Background: Google Trends (GT) data have shown promising results as a complementary tool to classical surveillance approaches. However, GT data are not necessarily provided by a representative sample of patients and may be skewed toward demographic and clinical groups that are more likely to use the internet to search for their health.

Objective: In this study, we aimed to assess whether GT-based models perform differently in distinct population subgroups. To assess that, we analyzed a case study on asthma hospitalizations.

Methods: We analyzed all hospitalizations with a main diagnosis of asthma occurring in 3 different countries (Portugal, Spain, and Brazil) for a period of approximately 5 years (January 1, 2012-December 17, 2016). Data on web-based searches on common cold for the same countries and time period were retrieved from GT. We estimated the correlation between GT data and the weekly occurrence of asthma hospitalizations (considering separate asthma admissions data according to patients' age, sex, ethnicity, and presence of comorbidities). In addition, we built autoregressive models to forecast the weekly number of asthma hospitalizations (for the different aforementioned subgroups) for a period of 1 year (June 2015-June 2016) based on admissions and GT data from the 3 previous years.

Results: Overall, correlation coefficients between GT on the pseudo-influenza syndrome topic and asthma hospitalizations ranged between 0.33 (in Portugal for admissions with at least one Charlson comorbidity group) and 0.86 (for admissions in women and in White people in Brazil). In the 3 assessed countries, forecasted hospitalizations for 2015-2016 correlated more strongly with observed admissions of older versus younger individuals (Portugal: Spearman ρ=0.70 vs ρ=0.56; Spain: ρ=0.88 vs ρ=0.76; Brazil: ρ=0.83 vs ρ=0.82). In Portugal and Spain, forecasted hospitalizations had a stronger correlation with admissions occurring for women than men (Portugal: ρ=0.75 vs ρ=0.52; Spain: ρ=0.83 vs ρ=0.51). In Brazil, stronger correlations were observed for admissions of White than of Black or Brown individuals (ρ=0.92 vs ρ=0.87). In Portugal, stronger correlations were observed for admissions of individuals without any comorbidity compared with admissions of individuals with comorbidities (ρ=0.68 vs ρ=0.66).

Conclusions: We observed that the models based on GT data may perform differently in demographic and clinical subgroups of participants, possibly reflecting differences in the composition of internet users' health-seeking behaviors.