Utilizing Google Search Data With Deep Learning, Machine Learning and Time Series Modeling to Forecast Influenza-Like Illnesses in South Africa

Abstract

Influenza-like illnesses (ILI) result in deaths and hospitalizations across the globe. Traditional surveillance systems rely on data from general medical practitioners. The process is resource-intensive and plagued with delay. Although recent studies have shown the potential utility of free and fast alternatives like web and social media data, the reliability cannot be generalized due to differences in technological culture. Meanwhile, there is a scarcity of studies exploring these free online data for (sub-Saharan) African countries. In this paper, we utilize Google trends (GT) data for ILI forecasting in South Africa. We study models based on deep learning (Long short-term memory (LSTM) and feedforward neural networks (FNN)), machine learning (Multiple linear regression (MLR), elastic net (EN), support vector machine (SVM)), and statistical time series (seasonal autoregressive integrated moving average (SARIMA)) algorithms. The FNN and SVM models using GT data alone, produce forecasts close in accuracy to those fitted to actual ILI data. The algorithms rank differently across various performance measures. Generally, the deep learning techniques perform better than the other algorithms in our study. However, tuning the former is quite intricate. Combining GT and historical ILI data enhances the models. The non-deep-learning algorithms benefit more from this enhancement. Furthermore, we observe that search volume increases proportional to and timeously with reported infection rates, suggesting that South Africans search Google in the week they feel flu symptoms. Thus, monitoring Google search trends is a reliable proxy for monitoring flu spread in South Africa.

Leveraging Google Search Data and Artificial Intelligence Methods for Provincial-level Influenza Forecasting: A South African Case Study

“…We used a range of statistical, machine learning, and deep learning algorithms that have been previously applied for influenza forecasting [10], [41], [43], [44], [46], [47] such as seasonal ARIMA (SARIMA), multiple linear regression (MLR), elastic net (EN), support vector machine regression (SVM), feedforward neural network (FNN), and long short-term memory (LSTM).…”

“…These models stem from the AI algorithms outlined in the previous section and the important input features for each province. For easy referencing and comparison, we maintain the same naming convention for the various provincial models as in the South African national ILI forecasting study [10]. For example, GT-MLR is a MLR model fitted to GT data only, while ILI-SARIMA is a SARIMA model based on past ILI data only.…”

“…We maintained the same number of training samples for the one and two weeks-ahead forecast models and reduced the test data by 1 and 2 instances respectively. Similar to Ref [10], the four sets of models under this category include GT-EN, GT-FNN, GT-MLR, and GT-SVM. The specific parameters tuning for the GT-FNN models for each of the two provinces concerned are outlined below: Gauteng: The GT-FNN models for Gauteng have 21 nodes in the input layer which depict the Google search volume of the 21 ILI-related terms.…”

“…Salathe [5] defined digital epidemiology as "epidemiology that uses data generated outside the public health system for disease surveillance". Some of the Internet-based data streams that have been explored in past studies include search engine [6]- [10] and Twitter data [11]- [14], news alerts [15]- [17], Wikipedia [18]- [20] and health-related blogs and websites [21]. Google trends (GT) is a commonly used search queries data source, which gives the relative amount of Google searches for particular queries in a given location.…”

“…One recent study investigated and established significant correlation between Google search volume for some ILI-related terms and real-world ILI incidence data [54]. A more recent study demonstrated the performance of Google search data for predicting ILI incidence in South Africa [10]. However, these studies have only focused on monitoring ILI at the 96 http://www.i-joe.org national level.…”