TL;DR
This study investigates how internet search query data (specifically Baidu Index) can be used to monitor pertussis (whooping cough) outbreaks in Shandong, China. It reveals that the effectiveness of this digital surveillance is significantly influenced by local sociodemographic factors, with higher education levels and urban populations showing stronger correlations with pertussis incidence. These findings are crucial for developing more accurate, location-specific early warning systems for infectious diseases.
What Problem Does This Paper Address?
Traditional infectious disease surveillance systems, while accurate, often suffer from a significant time lag (up to two weeks) between symptom onset and official notification. This delay hinders timely public health interventions. Internet search query data offers a promising alternative for near real-time monitoring, as individuals experiencing symptoms often search for information online. However, previous research has indicated substantial spatial variation in the accuracy of this digital surveillance. This paper aims to understand why this variation exists by examining the role of sociodemographic factors at a city level in Shandong province, China, and how these factors influence the correlation between internet search trends and actual pertussis cases.
Methods
The researchers employed a multi-faceted analytical approach:
- Data Collection: They gathered weekly pertussis notifications and corresponding Baidu Index (BI) data for cities in Shandong province from 2009 to 2017. They also collected yearly sociodemographic data for the same cities.
- Correlation Analysis: Spearman's correlation was used to assess the relationship between pertussis notifications and BI across different temporal aspects:
- Time-series correlation: Overall weekly correlation.
- Peak correlation: Correlation between the peak number of weekly cases and peak search interest.
- Increasing intensity correlation: Correlation reflecting the speed of epidemic spread.
- Composite Temporal Risk Metric (CTRM): To provide a single measure of the overall association between internet query data and pertussis surveillance, the Z-transformed correlation coefficients from the three temporal risk indices were averaged. A higher CTRM indicates a stronger correlation.
- Generalised Linear Models (GLMs): GLMs were used to model the relationship between the CTRM and various sociodemographic factors, calculating relative risks (RRs) to quantify the influence of each factor.
- Stratified Analysis: The dataset was stratified based on CTRM levels (low, medium, high) to examine how sociodemographic factors' influence changed across different levels of surveillance effectiveness.
- Regression Tree Analysis: This technique was used to identify hierarchical relationships and thresholds among sociodemographic factors that best predict the strength of the association between internet query data and pertussis surveillance.
Datasets
- Pertussis Notifications: Weekly clinical and laboratory-confirmed pertussis cases from the Chinese National Notifiable Disease Reporting System (CNNDRS) for cities in Shandong province (2009-2017).
- Baidu Index (BI) Data: Weekly search metrics for the top ten search terms related to "pertussis" in Shandong province, obtained from Baidu (2009-2017).
- Sociodemographic Data: City-specific yearly data including:
- Demographic categories (total population, urban population, age groups).
- Socioeconomic categories (highly educated population, education years, GDP per capita).
- Internet access categories (internet access rate by cell phone and PC). This data was sourced from the Shandong Provincial Bureau of Statistics.
Key Insights and Ideas
- Internet Search Data is a Viable Tool, But Not Uniformly: The study confirms that internet search query data (Baidu Index) can indeed correlate with pertussis infections, offering a potential for real-time surveillance. However, the strength of this correlation varies significantly across different cities.
- Sociodemographics Matter: The effectiveness of internet query surveillance is not solely a technical issue but is deeply intertwined with the social and economic characteristics of a population. This is a critical insight for tailoring surveillance strategies.
- Education is a Key Driver: The proportion of highly educated individuals in a city emerged as the most significant determinant influencing the accuracy of internet query-based pertussis surveillance. This suggests that more educated populations are more likely to use the internet for health-related queries, and perhaps more accurately report their symptoms online.
- Urbanization and Internet Access are Also Important: Higher percentages of urban populations and greater internet access rates (particularly via mobile phones) also showed a positive association with stronger correlations between search queries and pertussis cases.
- Hierarchical Influence: Regression tree analysis revealed a hierarchical influence, with education level acting as a primary split, followed by urban population and internet access, indicating a complex interplay of factors.
Key Results
- The study found significant correlations between pertussis infections and internet query data across Shandong province, with a Composite Temporal Risk Metric (CTRM) ranging from 0.59 to 1.24.
- Sociodemographic Factors and CTRM:
- Percentage of urban population: Relative Risk (RR) of 1.05 (95% CI 1.03–1.07).
- Proportion of highly educated population: RR of 1.27 (95% CI 1.16–1.39) - the strongest predictor.
- Internet access rate (cell phone): RR of 1.04 (95% CI 1.02–1.05).
- Stratified Analysis: Higher CTRM values (indicating better surveillance effectiveness) were associated with higher RRs for urban population, highly educated population, and mobile internet access. For instance, in areas with the highest CTRM (>70th percentile), the RR for highly educated population was 1.31 (95% CI 1.23–1.38).
- Regression Tree Findings: The percentage of highly educated population was identified as the most important factor, with thresholds like ≥4.80% significantly increasing the correlation coefficients.
Limitations and Future Work
- Reporting Bias: The study acknowledges that CNNDRS only captures reported cases, excluding individuals who don't seek medical care or are misdiagnosed. This underreporting might be more prevalent in less urbanized or less educated areas, potentially affecting the observed correlations.
- Internet Search Behavior: Differences in internet-seeking behaviors across communities could influence surveillance accuracy.
- Missing Factors: The authors suggest that future models could be improved by including additional predictors, such as climate factors, healthcare access, gender, and age, to provide a more comprehensive understanding.
- Spatially-Specific Criteria: The findings highlight the need for developing location-specific criteria for using internet query data, moving beyond a one-size-fits-all approach. This could lead to more robust early warning systems for pertussis and other infectious diseases.
Summary updated