Discovering Time-Varying Public Interest for COVID-19 Case Prediction in South Korea Using Search Engine Queries:...

Area of Science:

• Epidemiology
• Computational Linguistics
• Machine Learning

Background:

• Forecasting COVID-19 cases is vital for policy and lifestyle adjustments.
• Previous machine learning models for case prediction used static queries, limiting their ability to capture pandemic dynamics.
• Temporal variations in public search queries are crucial for understanding and predicting pandemic trends.

Purpose of the Study:

• To develop a novel framework for extracting COVID-19-related keywords that account for temporal variations.
• To investigate time-delayed web search behavior and its correlation with public interest in COVID-19.
• To improve COVID-19 case prediction by incorporating dynamically extracted, time-sensitive keywords.

Main Methods:

• Trained word embedding models on a news corpus to extract time-varying keywords related to "Corona" over 4-month intervals.
• Utilized time-lagged cross-correlation to identify optimal time lags between expanded queries and confirmed COVID-19 cases.
• Applied Principal Component Analysis (PCA) for feature reduction and ElasticNet regression for predicting daily case counts.

Main Results:

• Successfully extracted phase-specific keywords reflecting COVID-19 symptoms, societal impact, policy responses, and public sentiment (e.g., "economic crisis", "anxiety").
• Models trained with time-lagged, dynamically extracted queries significantly outperformed previous methods for 1-14 day ahead predictions.
• Demonstrated superior performance compared to models using only past case counts or static queries, particularly for 9-11 day ahead predictions (P<.01).

Conclusions:

• A novel COVID-19 case prediction model was developed using automated, time-aware keyword extraction via word embedding.
• The proposed model surpasses traditional methods relying on static or heuristic queries, requiring no prior expert knowledge.
• The approach effectively captures temporal shifts in public interest, offering a dynamic tool for pandemic monitoring and prediction.