Interpretable surrogate modelling for agent-based simulations to better inform COVID-19 decision making

Laura Rosella, Dalla Lana School of Public Health; Dionne Aleman, Faculty of Applied Science & Engineering

SARS-CoV-2 (COVID-19) was officially declared a pandemic in March 2020 and infections spread rapidly around the world. Researchers, healthcare practitioners, and government officials recognized the need for effective mitigation strategies to reduce the incidence of infection. The Medical Operations Research Lab's Pandemic Outbreak Planner (morPOP) is an agent-based simulation (ABS) model utilized to illustrate the potential outcomes of COVID-19 mitigation strategies at various compliance rates. Despite numerous applications of ABSs in the peer-reviewed literature, important gaps remain, with the high dimensional nature of these models resulting in high computational complexity and prohibitively long run times. Applications of surrogate machine learning (ML) have been implemented to overcome these challenges and demonstrated the ability to accurately predict ABS outputs while reducing overall run time.

This study aims to develop an interpretable ML surrogate model for the morPOP agent-based simulation model. Four model architectures are proposed – logistic regression, eXtreme Gradient Boosted trees, Support Vector Machines, and Artificial Neural Networks. Model performance will be evaluated using F1 Score and sensitivity and specificity metrics. The goal of the ML model is to accurately predict the number of cases resulting from implementing a specific mitigation strategy at a range of compliance rates. This information can be used by decision-makers about the potential efficacy of these strategies and therefore inform the selection of the optimal mitigation methods. Due to its intended use in public health decision-making, the focus is on maintaining interpretability and efficiency in the ML model while achieving high accuracy.