Leveraging advances in data-driven deep learning methods for hybrid epidemic modeling.

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• Mathematical modeling is vital for understanding epidemic dynamics and informing public health decisions.
• Existing models include mechanistic (SEIR-type), data-driven (DD), and hybrid approaches.
• The COVID-19 Scenario Modeling Hub (SMH) has utilized these models for over 12 rounds since early 2021.

Purpose of the Study:

• To summarize work within the COVID-19 SMH, emphasizing deep learning for epidemic modeling.
• To present a flexible data-driven framework that complements mechanistic models for evaluating future epidemic scenarios.
• To introduce novel Long Short-Term Memory (LSTM) network models for improved epidemic forecasting.

Main Methods:

• A traditional curve-fitting approach based on SEIR mechanisms was used initially.
• Two multivariate Long Short-Term Memory (LSTM) models were developed: a dependent input LSTM and an independent multivariate LSTM.
• The independent multivariate LSTM was designed to incorporate diverse data sources beyond traditional surveillance data.

Main Results:

• LSTM models effectively capture long-term and short-term epidemic behaviors by learning appropriate functions from data.
• The independent multivariate LSTM demonstrated the capability to integrate heterogeneous data sources like syndromic, environmental, and mobility data.
• The data-driven framework, especially LSTM, offers a feasible alternative and complement to mechanistic models for complex socio-epidemiological systems.

Conclusions:

• Deep learning techniques, particularly LSTM, offer a powerful enhancement to epidemic modeling.
• Data-driven approaches, leveraging big data, significantly expand the scope and accuracy of epidemic scenario evaluation.
• These advanced modeling strategies are crucial for informed decision-making during health emergencies.