Machine Learning Based Metamodels For Predicting Epidemic Trajectories From a SEIR Model
Epidemic transmission is one of the most common types of public health emergency that is difficult to forecast and often causes substantial harm to society. In epidemiology, mathematical models are used to analyze the spread and control of infectious diseases. Models use some basic assumptions and mathematics to find parameters for various infectious diseases and use these parameters to calculate the likely outcome of an epidemic. However, with more parameters introduced, models will become more complicated, they are made even more complex by the consideration of new requirements or interactions. Complex models take long time to run and it is very time consuming to rerun the models for sensitivity analysis. Therefore, we develop a simper model, known as metamodel, which is a model of a model, as the surrogate of the original model to speed up the process. In this study, we experiment the idea by using machine learning approach to predict the entire epidemic trajectory of a classic type of infectious disease model. Three different neural network models (Multilayer Perceptron [MLP], Long Short-Term Memory [LSTM], Gated Recurrent Units [GRU]) and a linear regression approach are used to construct the metamodels, and a SEIR (Susceptible, Exposed, Infectious, Recovered) model is developed for different configurations to train neural network and linear regression metamodels. We sample the different input parameters multiple times to generate the training and testing dataset for the model. In this study, 5000 samples are generated and are split into a training set of 4000 samples and a testing set of 1000 samples. We use Root Mean Error (RMSE), Mean Absolute Error (MAE) and coefficient of determination (R2) as performance indicators, the training time is also recorded. Among all the methods we apply, the LSTM network achieves the best performance of average R2 0.874 compare to MLP’s 0.563 and GRU’s 0.869, average RMSE 15.588 compare to GRU’s 16.263 and MLP’s 21.771, average MAE 11.664 compare to GRU’s 12.379 and MLP’s 16.239. Despite the MLP achieve the best average training time 30s compare to LSTM’s 61.2s and GRU’s 54.4s, it is considered not suited for time series prediction. The linear approach’s performance is not stable on the time span and the error is accumulative. Advisor: Qiushi Chen Assistant Professor Department of Industrial and Manufacturing Engineering
Penn State Only
Files are only accessible to users logged-in with a Penn State Access ID.
|Work Title||Machine Learning Based Metamodels For Predicting Epidemic Trajectories From a SEIR Model|
|License||All rights reserved|
|Work Type||Research Paper|
|Deposited||April 10, 2020|
This resource is currently not in any collection.