ML-ABC: Machine-learning assisted Approximate Bayesian Computation for efficient calibration of agent-based models...

Area of Science:

• Epidemiology
• Computational Biology
• Mathematical Modeling

Background:

• Agent-based models (ABMs) are crucial for epidemic modeling but complex to calibrate.
• Bayesian methods for parameter uncertainty quantification in ABMs are computationally challenging.
• COVID-19 pandemic highlighted the need for efficient and robust epidemic modeling calibration.

Purpose of the Study:

• To introduce and evaluate a novel Machine-Learning Approximate Bayesian Computation (ML-ABC) method for calibrating ABMs.
• To improve the efficiency and robustness of ABM calibration for epidemic data.
• To quantify parameter uncertainty effectively in complex ABMs.

Main Methods:

• Developed ML-ABC by combining a Machine-Learning step with Approximate Bayesian Computation.
• Applied ML-ABC to calibrate the Covasim stochastic ABM using COVID-19 hospitalization and death data.
• Benchmarked ML-ABC against traditional Rejection-ABC (R-ABC) for efficiency and accuracy.

Main Results:

• ML-ABC achieved identical posterior distributions of calibrated parameters as R-ABC.
• ML-ABC demonstrated significant speed improvements: 52% faster for the first wave and 33% faster for the second wave.
• The method proved robust across different epidemic scenarios.

Conclusions:

• ML-ABC offers a more efficient and robust approach to calibrating ABMs compared to traditional methods.
• This novel method enhances the ability to quantify parameter uncertainty in ABMs.
• ML-ABC has the potential to make Approximate Bayesian Computation competitive with point-estimate calibration, crucial for real-time epidemic modeling.