Can Machines Learn Respiratory Virus Epidemiology?: A Comparative Study of Likelihood-Free Methods for the Estimation of Epidemiological Dynamics.

Heidi L Tessmer,Kimihito Ito,Ryosuke Omori

doi:10.3389/fmicb.2018.00343

Abstract

To estimate and predict the transmission dynamics of respiratory viruses, the estimation of the basic reproduction number, R0, is essential. Recently, approximate Bayesian computation methods have been used as likelihood free methods to estimate epidemiological model parameters, particularly R0. In this paper, we explore various machine learning approaches, the multi-layer perceptron, convolutional neural network, and long-short term memory, to learn and estimate the parameters. Further, we compare the accuracy of the estimates and time requirements for machine learning and the approximate Bayesian computation methods on both simulated and real-world epidemiological data from outbreaks of influenza A(H1N1)pdm09, mumps, and measles. We find that the machine learning approaches can be verified and tested faster than the approximate Bayesian computation method, but that the approximate Bayesian computation method is more robust across different datasets.

Highlights

Prediction of infectious disease epidemics is essential to their control, and a difficult process
The first is the development of an individual-based (IBM) Susceptible Exposed Infectious Removed (SEIR) epidemiological model for generating data; the second is the approximate Bayesian computation (ABC) method used for estimating parameters; the third is the learning by multi-layer perceptrons (MLP), convolutional neural networks (CNN), and long-short term memory models (LSTM) machine learning models, again to estimate parameters; the fourth is the dataset creation and bootstrapping of the real-world and test data to create confidence intervals on the machine learning solutions; and calculation of the time it took for each method to obtain its estimates
The average errors of estimates made with ABC and ML are most similar for R0

Summary

Introduction

Prediction of infectious disease epidemics is essential to their control, and a difficult process. This is because the epidemiological dynamics, i.e., the time evolution of the number of infected individuals, are non-linear, with the probability of a susceptible individual acquiring infection depending on the number of infected individuals. Previous studies have constructed mathematical models describing the transmission dynamics of infectious disease, known as the SusceptibleInfectious-Removed (SIR) model, and fit the model to the time series data of the number of infected individuals (Bjrnstad et al, 2002). Conventional statistical methods, e.g., maximum likelihood estimation, require explicit solution of the time series data of the number of infected individuals from the SIR model. A mathematical model taking into account stochasticity is required to estimate parameters

Methods

Results

Conclusion