Using Baidu Search Engine to Monitor AIDS Epidemics Inform for Targeted intervention of HIV/AIDS in China

Kang Li, Meiliang Liu, Yi Feng, Chuanyi Ning, Weidong Ou, Jia Sun, Wudi Wei, Hao Liang, Yiming ShaoScientific 2019 · DOI

Abstract

China's reported cases of Human Immunodeficiency Virus (HIV) and AIDS increased from over 50,000 in 2011 to more than 130,000 in 2017, while AIDS-related search indices on Baidu rose from 2.1 million to 3.7 million in the same period. In China, individuals frequently seek health information on Baidu, one of the world's largest search engines. This study examines the relationship between national HIV surveillance data and the Baidu Index (BDI) to monitor the AIDS epidemic and inform targeted interventions. After screening keywords and constructing a composite index, we utilized seasonal autoregressive integrated moving average (ARIMA) modeling. The most accurate epidemic predictions were obtained using an ARIMA model with external variables (ARIMAX). A significant correlation between monthly HIV/AIDS reported cases and the Baidu Composite Index (r = 0.845, P < 0.001) was observed. Compared to the standard ARIMA model, the ARIMAX model incorporating the Baidu Composite Index achieved a lower Akaike information criterion (AIC, 839.42) and superior predictive accuracy (MAPE of 6.11%). These findings suggest that Baidu search query data is a reliable indicator for monitoring and predicting the HIV/AIDS epidemic in China.

Introduction

Traditional infectious disease surveillance often functions as a "rear-view mirror" exercise. In China, official HIV/AIDS data is typically released with a two-week lag and relies on a resource-intensive network of sentinel hospitals. Furthermore, the heavy social stigma surrounding HIV/AIDS may discourage individuals from seeking early clinical consultation. However, these same individuals often turn to the internet for discreet self-diagnosis and information. This research explores whether this "digital exhaust" from Baidu can fill the information gap, providing health authorities with a faster way to track and predict the spread of the virus.

Methodology

Data Sources

Researchers utilized two primary datasets covering January 2011 to June 2017:

  • Official Surveillance Data: Monthly aggregated new HIV/AIDS case counts from the China Centre for Disease Control and Prevention (China CDC).
  • Baidu Index (BDI): Daily and monthly search volume data for AIDS-related keywords.

Statistical Pipeline

The study employed a multi-stage approach to refine search data into a predictive tool:

  1. Keyword Selection: Starting with 100 keywords across categories like Etiology and Symptoms, researchers used Spearman’s correlation and cross-correlation analysis to identify 17 high-relevance terms.
  2. Composite Index Construction: A "Baidu Composite Index" (BDI) was created by weighting these 17 keywords based on their correlation strength with actual reported cases.
  3. Time-Series Modeling: The study compared a standard ARIMA model (predicting based on past case data) against an ARIMAX model, which integrated the BDI as an exogenous predictor.

Key Results and Insights

Model Performance and Correlation

The Baidu Composite Index showed a strong, statistically significant correlation with monthly reported HIV/AIDS cases ($r = 0.845$). The integration of search data markedly improved model accuracy. The ARIMAX model's AIC dropped to 839.42 (from 1184.78 in the standard model), and the Mean Absolute Percentage Error (MAPE) was reduced to 6.11%.

Figure 1: Time series plots comparing specific Baidu search indices (AIDS spread and Thrush) with monthly HIV/AIDS report cases from 2011 to 2017.

Behavioral Observations

  • The "Digital Confessional": Search data proved powerful for tracking stigmatized diseases. Keywords like "HIV" and "AIDS symptoms" were frequent, suggesting the internet is a primary point of contact for at-risk individuals.
  • Publicity Effects: A consistent surge in searches occurs every December, coinciding with World AIDS Day. While this reflects increased public awareness rather than a spike in infections, it highlights the impact of media reporting on search behavior.
  • Nowcasting Capabilities: The best correlation occurred at "lag 0" (the same month), indicating that search data is most effective for "nowcasting"—providing a real-time snapshot of the current month before official reports are finalized.

Discussion and Limitations

While the model is robust at a national level, its effectiveness is currently tied to internet penetration. In 2017, rural internet penetration in China was only 35.4%, potentially leading to under-representation of rural epidemics. Additionally, search behavior evolves; keyword libraries must be constantly updated to account for changing terminology and slang.

Future work should focus on developing province-specific ARIMAX models to account for geographical disparities in health literacy and epidemic severity, allowing local health authorities to target interventions more precisely.

Summary updated

Abstract

China's reported cases of Human Immunodeficiency Virus (HIV) and AIDS increased from over 50000 in 2011 to more than 130000 in 2017, while AIDS related search indices on Baidu from 2.1 million to 3.7 million in the same time periods. In China, people seek AIDS related knowledge from Baidu which one of the world's largest search engine. We study the relationship of national HIV surveillance data with the Baidu index (BDI) and use it to monitor AIDS epidemic and inform targeted intervention. After screening keywords and making index composition, we used seasonal autoregressive integrated moving average (ARIMA) modeling. The most correlated search engine query data was obtained by using ARIMA with external variables (ARIMAX) model for epidemic prediction. A significant correlation between monthly HIV/AIDS report cases and Baidu Composite Index (r = 0.845, P < 0.001) was observed using time series plot. Compared with the ARIMA model based on AIDS surveillance data, the ARIMAX model with Baidu Composite Index had the minimal an Akaike information criterion (AIC, 839.42) and the most exact prediction (MAPE of 6.11%). We showed that there are close correlations of the same trends between BDI and HIV/AIDS reports cases for both increasing and decreasing AIDS epidemic. Therefore, the Baidu search query data may be a good useful indicator for reliably monitoring and predicting HIV/AIDS epidemic in China.
BibTeX
@inproceedings{li2019using,
  author = {Kang Li and Meiliang Liu and Yi Feng and Chuanyi Ning and Weidong Ou and Jia Sun and Wudi Wei and Hao Liang and Yiming Shao},
  title = {{Using Baidu Search Engine to Monitor AIDS Epidemics Inform for Targeted intervention of HIV/AIDS in China}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-018-35685-w},
}
Figure