Lifestyle Disease Surveillance Using Population Search Behavior: Feasibility Study

Area of Science:

• Public Health
• Digital Epidemiology
• Health Informatics

Background:

• Official health statistics for lifestyle diseases are slow to produce.
• Web search data offers a potential proxy for lifestyle disease surveillance.
• Previous studies faced limitations like ad-hoc keyword selection and lack of generalization.

Purpose of the Study:

• To enhance methods for lifestyle disease surveillance using web search data.
• To identify limitations of Google Trends for disease surveillance.
• To test the generalizability of the methodology to other countries.

Main Methods:

• Collected lifestyle disease prevalence data (diabetes, obesity, exercise).
• Utilized Google Trends data with a rigorous keyword selection and denormalization process.
• Applied L1-regularized regression, multivariate spatio-temporal models, and transfer learning to predict prevalence.

Main Results:

• Proposed models outperformed prior work and baselines in the US, showing significant Mean Absolute Error (MAE) improvements.
• Achieved 24% improvement for diabetes, 18% for obesity, and 34% for exercise.
• Transfer learning to Canada demonstrated promising correlations (Spearman 0.70 for diabetes, Pearson 0.90-0.91 for obesity).

Conclusions:

• Modeling lifestyle diseases remains challenging, requiring substantial data and innovative strategies.
• Google Trends shows low-to-moderate quantitative validity for lifestyle disease surveillance, even with corrective approaches.
• Qualitative analysis and transfer learning (especially for low-resource countries) represent the most practical applications of Google Trends in this domain.