Forecasting COVID-19 spreading through an ensemble of classical and machine learning models: Spain's case study

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• Predicting the COVID-19 pandemic's evolution is crucial for public health interventions.
• Traditional epidemiological models like SEIR require extensive data, including recovered patient counts, which are often unavailable.
• Machine learning (ML) models offer alternative predictive capabilities but may struggle with long-term trends.

Purpose of the Study:

• To evaluate the applicability of an ensemble of population and ML models for predicting COVID-19 pandemic evolution in Spain using only public data.
• To compare the performance of individual model types against their ensemble.
• To identify key features influencing ML model predictions.

Main Methods:

• Trained ML models and adjusted classical ODE-based population models using incidence data.
• Created an ensemble by combining population and ML models for improved prediction robustness.
• Enhanced ML models with additional features: vaccination, human mobility, and weather data.
• Utilized Shapley Additive Explanation (SHAP) values to determine feature importance.

Main Results:

• The ensemble of population and ML models provided robust predictions for COVID-19 pandemic evolution.
• Adding features like vaccination and mobility data did not consistently improve the ensemble's performance due to differing model family patterns.
• ML models showed performance degradation with new COVID variants post-training.
• SHAP analysis revealed the relative importance of input features for ML model predictions.

Conclusions:

• Ensemble models combining population dynamics and ML offer a promising alternative to traditional SEIR-like compartmental models.
• This approach is particularly advantageous as it does not require data on recovered patients.
• The study highlights the potential of data-driven and mechanistic model integration for pandemic forecasting.