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

Mathematical modelling with agent-based models (ABMs) has gained popularity during the COVID-19 pandemic, but their complexity makes efficient and robust calibration to data challenging, particularly when applying Bayesian methods to quantify parameter uncertainty. We propose a method for calibrating ABMs that combines a Machine-Learning step with Approximate Bayesian Computation (ML-ABC). We showcase ML-ABC application with a proof-of-principle case study, in which we calibrate the Covasim -a stochastic ABM that has been used to model the English COVID-19 epidemic and inform policy at important junctions. Benchmarking against traditional Rejection-ABC (R-ABC), we illustrate the advantage of ML-ABC application in calibrating Covasim to data on hospitalisations and deaths from COVID-19 during the first and the second COVID-19 epidemic waves of 2020 and early 2021. Across scenarios, we demonstrate that using an ML screening step allows us to derive identical posterior distributions of the calibrated Covasim parameters as with the traditional R-ABC method, but faster. Specifically, we derive posterior distributions for input parameters around 52% faster when calibrating to the first epidemic wave and around 33% faster when calibrating parameters for the second epidemic wave, compared to the traditional R-ABC. Policy modelling requires calibration which is both efficient to adapt to fast-changing pandemic environments and robust to ensure confidence in policy decisions. However, existing ABM calibration often relies on myopic non-exhaustive searches in order to remain tractible, resulting in point parameter estimates. In this preliminary study, ML-ABC strictly improves upon existing ABC calibration approaches in all tested scenarios, indicating its potential to make ABC competitive with point-estimate calibration approaches. This novel approach offers a pathway to effectively calibrate ABMs in a way which is both efficient and quantifies parameter uncertainty, crucial for realising the potential of ABMs for timely and responsively modelling during an emerging epidemic.