This study adhered to the methodological framework proposed by Mavragani and Ochoa [21]. In a previous study by our team, we had (1) established a correlation between GT data related to common cold–related search terms and asthma hospitalizations and (2) evaluated whether GT data on the common cold, combined with data on admissions, could help forecast asthma hospitalizations. In this study, we applied the same methodology (correlations and forecast models) and used the same GT data but specifically considered those hospital admissions occurring in patients of each sex, age group (18-64 years old versus ≥65 years old), ethnicity (White versus Black or Brown [“pardo”]), and the presence or absence of at least one Charlson comorbidity [22]. We assessed a period of approximately 5 years (2012-2016), assessing data from Portugal, Spain, and Brazil.

Data analysis encompassed 2 different steps: (1) assessing the correlations between GT data and asthma hospitalizations in each country after applying time series analysis methods and (2) building models forecasting asthma hospitalizations for a period of 1 year based on GT and hospitalization data from the previous 4 years. To evaluate the predictive capability of the models, we compared the forecasted asthma admissions with the observed hospitalization data. For both GT and the frequency of hospitalizations, we worked with data displayed on a weekly basis (as it allowed detection of short-term variations while mitigating the impact of large random fluctuations that can occur when data are examined on a daily basis).

First, we performed a cross-correlation analysis to examine the correlation between GT data and asthma hospitalizations (cross-correlation can be understood as a statistical method used to analyze the relationship [correlate] between 2 continuous variables that can be measured or sampled at different points in time) [29]. Given that (1) for GT data, a relevant secular trend was expected (reflecting an increase in Google searches over time) and (2) GT and hospitalization data are expressed on different scales (GT results are expressed as relative search values [ie, percentages in relation to the maximum observed value of the whole period], whereas hospitalizations are expressed as absolute values), we removed the secular trend component for both GT and hospitalization data then assessed the correlation between search volumes and asthma hospitalizations. We analyzed Spearman correlations between GT and hospitalization data registered in the same week, as well as cross-correlation coefficients for a lag of 1 and 2 weeks to determine if search volumes demonstrated a stronger correlation with asthma hospitalizations occurring afterward rather than those happening concurrently. Correlation coefficients were presented alongside 95% CIs, which were computed using bootstrap methods for Spearman correlation coefficients.

Second, we built seasonal autoregressive integrated moving average (ARIMA) models to forecast variations in asthma hospitalizations over a period of 1 year [19,30]. Seasonal ARIMA models are used to forecast time series data that exhibit repeating patterns over fixed intervals (in this case, yearly cycles). These models take into account both the nonseasonal patterns and the seasonal variations in the data to make accurate predictions for future time points [31]. For each analysis, seasonal ARIMA models parameters (p, d, q)(P', D, Q)s were defined, where p denotes the order of autoregression, d denotes the degree of difference, q denotes the order of the moving average part, P' denotes the seasonal order of autoregression, D denotes the degree of difference following seasonal integration, Q denotes the seasonal moving average, and s denotes the length of the seasonal period. We chose d and D so that the 2012-2016 time series appeared stationary (ie, with constant variance and no extreme fluctuations or overall increasing or decreasing behavior), including by testing a measure of seasonal strength [32]; we chose s=52 weeks (since there are roughly 52 weeks in a year); we chose p and P' based on spikes in partial autocorrelation function plots; and we chose q and Q based on spikes in autocorrelation function plots. Identification of these parameters using autocorrelation and partial autocorrelation plots allowed us to define candidate models. Final seasonal ARIMA models were then selected based on the results of the Ljung-Box test (which was applied to assess whether residuals look like white noise) and on the minimization of the Akaike information criteria and Bayesian information criteria (see Table S1 in Multimedia Appendix 1 for the parameters defined for each model). In this study, we used asthma hospitalization data alongside GT data to forecast future asthma hospitalizations. The data set was split into a training set and a testing set. Specifically, the training set comprised asthma hospitalizations and GT data collected from July 1, 2012, to June 20, 2015. We then used this trained model to forecast asthma hospitalizations for the testing set, which included hospitalizations between the weeks of June 21, 2015, and June 19, 2016. This split allowed for the evaluation of model performance on previously unseen data.

To evaluate the predictive performance of the models, several measures were used: (1) the Spearman correlation coefficients between the predicted variation in hospitalizations and the actual number of asthma hospitalizations (ie, without time series decomposition), (2) the average weekly difference between the numbers of predicted and observed hospitalizations, and (3) the number of weeks whose number of observed asthma hospitalizations fell outside the 95% CI for predicted admissions.

All analyses were performed using R software, version 4.3.0 (R Foundation for Statistical Computing) [33], using the forecast and urca packages.

In this study, we assessed the correlations of GT data and the performance of GT-based models in different subgroups of patients (as defined by clinical and demographic characteristics: age, sex, ethnicity, and presence of comorbidities) using the case study of asthma. Overall, our results point out that GT-based models may not necessarily have the same performance in all subgroups of patients, highlighting that GT data may vary across different segments of users. In fact, we observed stronger correlations between GT data and asthma hospitalization data or between forecasted and observed hospitalizations when assessing admissions of women or patients without comorbidities. Less consistent results were observed, in particular in Brazil, according to age group.

Overall, studies using GT data for surveillance purposes have obtained mixed results. On the one hand, GT has been shown to have an effective potential to monitor the spread of infectious diseases, track public interest in health-related topics, and identify emerging trends in public health [37,38]. However, there have also been instances where GT has shown inconsistencies or failed to provide accurate predictions, emphasizing the need to carefully interpret the data [23]. In part, these failures may be related to differences in the composition of internet users compared with that of patients with a particular disease. Although we cannot necessarily generalize the results observed in the use case of asthma hospitalizations to other conditions or countries, this paper is relevant from a methodological point of view, as it demonstrates, through a case study, how the association between GT and disease data is not always the same for all groups of individuals, pointing to the need to study these associations according to the characteristics of the patients.

This study is also relevant for asthma care, as this was the condition we particularly assessed. Regarding our findings of the performance of GT-based models in the distinct subgroups of asthma hospitalizations, we observed relevant gender-related differences. In fact, women have higher asthma prevalence, severity, and health care utilization than men [39,40]. The better correlations and model performance observed in female admissions may be related to the higher prevalence of asthma in this population. On the other hand, women often exhibit more proactive health information–seeking behaviors, with a particular emphasis on their own health and well-being as well as that of their families—which may partly explain the higher internet use by women than men [41]. In addition, in some cultures, women may have primary caregiving responsibilities for family members’ health, including managing asthma [42]. This can also contribute to increasing interest and information-seeking behavior and enhance engagement with online platforms, possibly explaining the higher correlations and better performance of models in admissions of women.

Younger adults, especially those who are generally healthy, may exhibit different health-seeking behaviors than older adults or individuals with chronic illnesses [43]. Younger adults may tend to be more proactive in seeking health care information online and to be more likely to use search engines [44], in part given their historically higher access and rates of internet use [44] (which can be attributed to factors such as greater digital literacy, increased reliance on technology for information and communication, and higher rates of smartphone ownership [44,45]). However, that access has also been proliferating very quickly among older adults, who may possibly be more concerned about their health [46]. These changing patterns may partly explain the heterogeneity of our results obtained on age groups, with higher correlations found for younger adults in contrast with better performance of forecasting models in admissions of older adults. Such a pattern was also observed when performing separate analyses for age groups of 15 to 44-year-olds and 45 to 64-year-olds in Spain and Brazil. All things considered, our findings may offer insights into digital divides, hinting at disparities in internet access, digital literacy, and health information–seeking behaviors across demographic groups. The use of GT-based tools for complementing surveillance systems may have important implications in terms of health equity, considering the discrepancies in internet access across clinical and demographic subgroups.

Information on the presence of comorbidities or ethnicity of patients was only available for one country each (Portugal and Brazil, respectively). The presence of comorbidities has been associated with worse health outcomes for asthma admissions [47]. In addition, we observed a less pronounced seasonal pattern in Portuguese hospitalizations of individuals with comorbidities than in those without, potentially explaining the worse performance of GT-based models in forecasting those admissions. Small differences were observed regarding ethnicity, with a slightly better performance of models for Whites in Brazil, possibly reflecting different regional demographics, internet use patterns, or health care–seeking patterns.

Several limitations should be discussed. First, the differences observed in the performance of GT across different subgroups are not necessarily generalizable to other countries and conditions (eg, we cannot state that GT-based models always display better performance when considering data from female participants). Second, data availability on hospitalizations was limited to 3 countries and, regarding the presence of comorbidities or ethnicity, we only had that information for Portugal and for Brazil, respectively. In addition, the small frequency of weekly admissions precluded the comparison of the performance of the models’ unspecific sets of Charlson comorbidity groups. Third, GT provides data on search term popularity and relative interest (ie, GT presents searches as a relative volume instead of as an absolute number of searches), which makes comparisons between queries difficult and reveals less information about the absolute search interest in each aspect being assessed. In addition, it does not provide detailed information about the context or intent behind the searches, thus making it prone to bias due to possible media coverage [48]. In the particular context of this study, we were not able to quantify the number of searches on the “common cold” resulting from individuals experiencing cold symptoms versus reflecting other intentions (eg, search for news on the common cold). This lack of specificity can make it challenging to establish a causal link between search behaviors and the studied outcomes. However, this is an inherent limitation of GT, and our goal was not so much to establish an association between searches on “common cold” and asthma hospitalizations but rather to assess how that association varies considering different subgroups. Finally, during the assessed period, there was an increase in the use of the internet. However, we applied time series analysis methods, removing the estimated trend components for both GT and hospitalization data.

This study also had several strengths. In particular, this study has an important novelty component—to the best of our knowledge, this is the first time that the performance of GT-based models has been investigated across diverse demographic and clinical subgroups, with relevant potential implications for considering digital divides and health equity–related aspects in interpreting results of GT-based tools. In addition, we assessed 3 different countries (1 in Europe and 1 in South America) using nationwide data for a period of 5 years. We applied 2 different strategies to retrieve common cold–related GT data—GT data on the pseudo-influenza syndrome topic and search terms regarding the common cold—which obtained comparable results. We examined asthma as a case study since (1) asthma, in comparison with other diseases (such as COVID-19), is less subject to a high or variable media coverage, thus not particularly biased for GT data [30,48]; (2) the relationship between asthma admissions and GT data on the common cold has been already established [19]; and (3) the influence of patients’ characteristics on asthma outcomes has been assessed [20]. Although this study relied on a case study on asthma admissions, there is potential application of this methodology to other diseases and segments of the population to better understand the context in which GT-based models can be better applied.

In this study, we observed better performance of models forecasting asthma hospitalizations in women, White individuals (Brazil), and patients without comorbidities (Portugal), suggesting that the models based on GT may perform differently in subgroups of participants, which may indicate variations in the patterns of health-related information seeking among different segments of internet users. Although GT data have increasingly been assessed as a potential complementary tool to more classical surveillance approaches, determining the best practices for using GT data and understanding its limitations requires exploring in which segments of users it performs better. Although this study assessed the use case of asthma in 3 countries and shows differences in different segments of the population, future studies should explore how GT-related models may differentially perform in accordance with other variables, such as sociodemographic variables (like age, gender, education, income, urban/rural context, underserved populations), as well as to test differences observed in other diseases, countries, and clinical data sources. This study contributes to advancing our understanding of the complexities inherent in the infodemiology field and hints at the need to consider population subgroups and health contexts for the applicability of GT-based surveillance systems.