An Agent-Based Model for COVID-19 in Bangladesh

BackgroundThe COVID-19 pandemic, that has resulted in millions of deaths and hundreds of millions of cases worldwide, continues to affect the lives, health and economy of various countries including Bangladesh. Despite the high proportion of asymptomatic cases and relatively low mortality, the viruss spread had been a significant public health problem for densely populated Bangladesh. With the healthcare system at stress, understanding the disease dynamics in the unique Bangladesh context became essential to guide policy decisions. MethodsWith a goal to capture the COVID-19 disease dynamics, we developed two stochastic Agent-Based Models (ABMs) considering the key characteristics of COVID-19 in Bangladesh, which vastly differ from the developed countries. We have implemented our ABMs extending the popular (but often inadequate) SIR model, where the infected population is sub-divided into Asymptomatic, Mild Symptomatic and Severe Symptomatic populations. One crucial issue in Bangladesh is the lack of enough COVID-19 tests as well as unwillingness of people to do the tests resulting in much less number of official positive cases than the actual reality. Although not directly relevant to the epidemiological process, our model attempts to capture this crucial aspect while calibrating against official daily test-positive cases. Our first model, ABM-BD, divides the population into age-groups that interact among themselves based on an aggregated Contact Matrix. Thus ABM-BD considers aggregate agents and avoids direct agent level interactions as the number of agents are prohibitively large in our context. We also implement a scaled down model, ABM-SD, that is capable of simulating agent level interactions. ResultsABM-BD was quite well-calibrated for Dhaka: the Mean Absolute Percentage Error (MAPE) between official and forecasted cases was 1.845 approximately during the period between April 4, 2020 and March 31, 2021. After an initial model validation, we conducted a number of experiments - including retrospective scenario analysis, and hypothetical future scenario analysis. For example, ABM-BD has demonstrated the trade off between a strict lockdown with low infections and a relaxed lockdown with reduced burden on the economy. Leveraging the true agent level interaction capability of ABD-SD, we have also successfully analyzed the relative severity of different strains thereby (confidently) capturing the effect of different virus mutations. ConclusionsOur models have adequately captured the COVID-19 disease transmission dynamics in Bangladesh. This is a useful tool to forecast the impact of interventions to assist policymakers in planning appropriate COVID response. Our models will be particularly useful in a resource constrained setting in countries like Bangladesh where the population size is huge.