Data Analysis and Forecasting of the COVID-19 Spread: A Comparison of Recurrent Neural Networks and Time Series Models

Abstract

To understand and approach the spread of the SARS-CoV-2 epidemic, machine learning offers fundamental tools. This study presents the use of machine learning techniques for projecting COVID-19 infections and deaths in Mexico. The research has three main objectives: first, to identify which function adjusts the best to the infected population growth in Mexico; second, to determine the feature importance of climate and mobility; third, to compare the results of a traditional time series statistical model with a modern approach in machine learning. The motivation for this work is to support health care providers in their preparation and planning. The methods compared are linear, polynomial, and generalized logistic regression models to describe the growth of COVID-19 incidents in Mexico. Additionally, machine learning and time series techniques are used to identify feature importance and perform forecasting for daily cases and fatalities. The study uses the publicly available data sets from the John Hopkins University of Medicine in conjunction with the mobility rates obtained from Google’s Mobility Reports and climate variables acquired from the Weather Online API. The results suggest that the logistic growth model fits best the pandemic’s behavior, that there is enough correlation of climate and mobility variables with the disease numbers, and that the Long short-term memory network can be exploited for predicting daily cases. Given this, we propose a model to predict daily cases and fatalities for SARS-CoV-2 using time series data, mobility, and weather variables.

Forecasting the spread of COVID-19 based on policy, vaccination, and Omicron data

“…For instance, some studies forecasted cumulative confirmed cases 29 or cumulative vaccination rates 30 using past observations. Other studies used predictor variables including temperature, mobility, and social distancing 26 , 27 . However, these studies only compare the prediction models, not predictor variables 26 – 28 .…”

Estudios sobre el desarrollo económico en México. Un enfoque multifactorial

“…A su vez, la combinación de métodos de aprendizaje automatizado con modelos tradicionales arima y de crecimiento ha sido empleada para modelar la expansión de contagios en México y otros países, como es el análisis de Naemi et al (2022), en donde se analizan los brotes de covid-19 utilizando modelos híbridos. Asimismo, el análisis de Gómez-Cravioto et al (2022) para predecir picos de contagios se basa en modelos de series de tiempo, específicamente en Vectores Autorregresivos (var) combinados con una especificación de red neuronal autorregresiva (Long Short-Term Memory, lstm, por sus siglas en inglés), propuesto por Hochreiter y Schmidhuber (1997), mismos a que vez pertenecen a los modelos de aprendizaje profundo.…”

Deep learning for Covid-19 forecasting: State-of-the-art review.

“… Brazil, France, India, Mexico, Russia, Saudi Arabia, the US; Jan - Sep, 2020 0.63 - 6.02. Gomez et al, 2021 [28] Compared univariate population growth models, VAR, and M-LSTM and found that the M-LSTM model achieved the lowest errors. Mexico; Jan - Mar, 2020 0.47 Kafieh et al, 2021 [41] Compared RF, MLP, LSTM-R, LSTM-E, M-LSTM models and found that M-LSTM achieved the smallest error.…”