Deep neural network-based emulators for COVID-19 hospitalization simulations
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
In March of 2020, Delaware Emergency Management Agency, DEMA created the COVID-19 Predictive Modeling group with epidemiologists and medical professionals from Delaware hospitals and Professors from the University of Delaware. University of Delaware Professors Gregory Dobler and Federica Bianco who are members of this group developed a hybrid compartmental and agent-based model using Markov Chain Monte Carlo (MCMC) framework to predict hospitalization for Delaware hospitals, DHARMA – Delaware Hospitalization and Recovery Metrics Analysis. This thesis work presents the contributions by the author in modifying the DHARMA model for factors that impacted hospitalization during the spring and summer of 2021. Several transitions happened during this period, such as relaxation of social policies, administration of vaccinations, and the emergence of the dominant COVID-19 delta variant, all of which required modifications to the DHAMA code to predict hospitalization during this period. ☐ DHARMA program uses MCMC methods, specifically the affine invariant MCMC model, to sample the posterior distribution of parameters of interest by random sampling in a probabilistic space, running parallel chains for thrice the number of the parameters (i.e., 45 independent parallel chains, thrice the number of parameters). The MCMC method is guaranteed to sample the full posterior in the limit of infinite samples. We typically configure the runs for 10000 steps to stabilize and use the optimized parameters to predict hospitalization for the next 100 days. As the number of steps increases the execution time increases and for a 10000 steps configuration, it takes 4-6 hours to complete on an 8CPU computer (Intel(R) 1.65 GHz). ☐ We considered building an emulator: a supervised model based on deep neural network architecture to reduce the computational costs. This thesis presents the work by the author in building emulators for DHARMA. Emulators are built based on supervised learning deep neural network architecture. DHARMA emulators are trained using DHARMA-generated hospitalization simulations, using infection rate, β time series as features, and corresponding hospitalization time series as the target. Neural network architecture for emulators included Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and two variations of CNN with additional hidden layers of Long Short-Term Memory (LSTM). The results of these exploratory studies show the models can learn the temporal trends of hospitalization and significantly reduce the computational costs of predictions. However, the prediction of hospitalization had significant inaccuracies. A plot of target vs. predicted values showed a linear relationship, and the slope from the linear regression was used to revise the prediction. This revision significantly improved prediction accuracy. We also provide a few suggestions for further studies and to improve prediction accuracies.
Description
Keywords
COVID-19, Deep neural network, Emulator, MCMC, SIR model