JMIR Publications

ABSTRACT

Background:

Megacities are at increased risk of respiratory infectious disease transmission due to high population density, frequent population mobility, and complex social contact patterns, which may rapidly increase pressure on healthcare services. Traditional surveillance based on reported cases is subject to delays in healthcare seeking, diagnosis, and reporting, and may therefore be insufficient for early warning. Regional pathogen surveillance and internet search behavior may provide potential upstream signals before increases in reported cases. However, how to integrate these two data sources and use spatial structure to improve city-level warning performance remains insufficiently studied.

Objective:

This study aimed to address the limited timeliness of case-based respiratory infectious disease surveillance in megacities by developing a multisource early-warning framework for Beijing. We evaluated whether spatially stratified regional pathogen indicators and Baidu Index search behavior could improve short-term prediction beyond historical case trends.

Methods:

This study developed a multisource early-warning framework for respiratory infectious diseases in Beijing. We integrated weekly reported cases of notifiable respiratory infectious diseases in Beijing from January 2023 to December 2025, weekly pathogen surveillance indicators from 31 provincial-level administrative regions in China, and Baidu Index search data for Beijing from January 2023 to June 2025. Regional pathogen indicators were innovatively constructed at multiple spatial scales, including Beijing local indicators, neighboring provinces, northern China, southern China, and national aggregation excluding Beijing. Baidu search terms were grouped into clinically meaningful composite indicators, including core respiratory symptoms and respiratory disease names. Lagged correlations from 0 to 4 weeks were evaluated. A series of LASSO-Poisson models were developed for one-week-ahead prediction of case counts, including a historical-case baseline model, regional pathogen models, a Baidu Index model, a regional–digital fusion model, a full multisource model, and a core multisource model. Model performance was evaluated using RMSE, MAE, SMAPE, and the percentage reduction in RMSE relative to the baseline model.

Results:

Regional pathogen activity and Baidu Index indicators showed clear temporal fluctuations and varying degrees of synchronous or lagged correlations with reported cases in Beijing. Overall pathogen positivity rates and coinfection-related indicators mainly showed synchronous correlations, whereas influenza A virus positivity rates in neighboring provinces and northern China showed certain leading correlations. For one-week-ahead prediction, the M10 core multisource model achieved the best performance, with an RMSE of 20,460.38, representing a 74.3% reduction compared with the M1 baseline model. The M7 Baidu Index model ranked second, with a 71.5% reduction in RMSE. Among regional pathogen models, the neighboring-province and northern-China models outperformed the southern-China control model, whereas the national aggregation model excluding Beijing performed worse than the baseline model. In the two-week-ahead sensitivity analysis, multisource models still maintained an approximately 50% reduction in RMSE.

Conclusions:

Integrating spatially structured regional pathogen surveillance with Baidu Index search behavior can substantially improve the short-term prediction of respiratory infectious disease activity in Beijing. Pathogen signals from neighboring provinces and northern China had greater predictive value than broad national aggregation or more distant regional indicators. This study provides a generalizable multisource data fusion framework for urban respiratory infectious disease early warning and supports the integration of laboratory-based pathogen surveillance, digital search data, and historical case trends to enhance public health emergency response capacity. Clinical Trial: Not applicable. This study was not a clinical trial and did not involve participant enrollment or assignment to interventions.