Evaluation of the predictive capabilities of the SuEIR model

Abstract

COVID-19 was an outburst that led to a global pandemic across the nations which started in late 2019 in Wuhan, China. The virus had several adverse implications on the health, economy, education, and daily activities of a human being. Due to these implications, the development of epidemic models became important for the well-being of the people. Epidemic models are developed on mathematical, computational, and statistical grounds which could assist in preventing the widespread of the disease. These models could be an important aspect for guiding the decision-makers, governments, and the public for themselves. UCLA has developed an epidemic model known as the UCLA-ML model that has the potential to forecast the mortality rates, confirmed cases, and ICU hospitalization cases. The differentiating factor of the UCLA-ML model is that it takes into consideration unreported cases and hospitalizations as well. Moreover, it inculcates machine learning algorithms for selecting the parameters. The model is developed using the data for 27 countries and 53 US states from authentic data sources such as the New York Times and Johns Hopkins University (JHU). However, there is a need to evaluate the predictive capabilities of the model under different circumstances. As a result, the thesis focuses on evaluating the prediction performance of the model by proposing a sliding window approach. This approach trains and tests the model with different time frames of data which will eventually generate statistical performance metrics to evaluate the prediction framework of the UCLA-ML model. The evaluation framework can also be used for comparing predictions of other epidemic models that are capable of forecasting mortality rates.