Comparison of an artificial neural network and Gompertz model for predicting the dynamics of deaths from COVID-19 in Mexico

The present work is focused on modeling and predicting the cumulative number of deaths from COVID-19 in Mexico by comparing an artificial neural network (ANN) with a Gompertz model applying multiple optimization algorithms for the estimation of coefficients and parameters, respectively. For the modeling process, the data published by the daily technical report COVID-19 in Mexico from March 19th to September 30th were used. The data published in the month of October were included to carry out the prediction. The results show a satisfactory comparison between the real data and those obtained by both models with a R-2 > 0.999. The Levenberg-Marquardt and BFGS quasi-Newton optimization algorithm were favorable for fitting the coefficients during learning in the ANN model due to their fast and precision, respectively. On the other hand, the Nelder-Mead simplex algorithm fitted the parameters of the Gompertz model faster by minimizing the sum of squares. Therefore, the ANN model better fits the real data using ten coefficients. However, the Gompertz model using three parameters converges in less computational time. In the prediction, the inverse ANN model was solved by a genetic algorithm obtaining the best precision with a maximum error of 2.22% per day, as opposed to the 5.48% of the Gompertz model with respect to the real data reported from November 1st to 15th. Finally, according to the coefficients and parameters obtained from both models with recent data, a total of 109,724 cumulative deaths for the inverse ANN model and 100,482 cumulative deaths for the Gompertz model were predicted for the end of 2020.

Automated summary

This study compares the performance of an artificial neural network (ANN) with a Gompertz model using different optimization algorithms for modeling and predicting the cumulative deaths from COVID-19 in Mexico. The results show that the ANN model better fits the real data, and different optimization algorithms have different impacts on coefficients and parameters.