Benchmarking Disease Modeling Techniques on the Philippines' COVID-19 Dataset

Abstract

: The COVID-19 pandemic has emphasized the importance of timely and accurate prediction of disease outbreaks. Mathematical disease models can help simulate the trajectory of diseases and guide policymakers in identifying priorities and gaps in current policies. This study evaluates the performance, on various metrics, of three different parameter estimation algorithms in compartmental models, i.e., Nelder-Mead, Simulated Annealing, and L-BFGS-B, together with the ARIMA time-series modeling, in modeling COVID-19 cases. Using the daily number of confirmed cases of COVID-19 in the Philippines as the dataset, the models were trained on 90 different periods, with each period having 30 days of case data. After training, the models were used to predict the cases up to 30 days later. The Negative Log Likelihood (NLL), time spent, iterations per second, and memory allocation were all measured. The results show that ARIMA performed better in terms of accuracy, time, and space efficiency than each of the other algorithms. This suggests that ARIMA should be preferred for predicting the number of cases. However, policymaking sometimes requires scenario-based modeling, which ARIMA is unable to provide. For such requirements, any of the three compartmental models may be preferred, as each performed generally very well, too.