Data-Driven Prediction of COVID-19 Daily New Cases through a Hybrid Approach of Machine Learning Unsupervised and Deep Learning

Abstract

Air pollution is associated with respiratory diseases and the transmission of infectious diseases. In this context, the association between meteorological factors and poor air quality possibly contributes to the transmission of COVID-19. Therefore, analyzing historical data of particulate matter (PM2.5 and PM10) and meteorological factors in indoor and outdoor environments to discover patterns that allow predicting future confirmed cases of COVID-19 is a challenge within a long pandemic. In this study, a hybrid approach based on machine learning and deep learning is proposed to predict confirmed cases of COVID-19. On the one hand, a clustering algorithm based on K-means allows the discovery of behavior patterns by forming groups with high cohesion. On the other hand, multivariate linear regression is implemented through a long short-term memory (LSTM) neural network, building a reliable predictive model in the training stage. The LSTM prediction model is evaluated through error metrics, achieving the highest performance and accuracy in predicting confirmed cases of COVID-19, using data of PM2.5 and PM10 concentrations and meteorological factors of the outdoor environment. The predictive model obtains a root-mean-square error (RMSE) of 0.0897, mean absolute error (MAE) of 0.0837, and mean absolute percentage error (MAPE) of 0.4229 in the testing stage. When using a dataset of PM2.5, PM10, and meteorological parameters collected inside 20 households from 27 May to 13 October 2021, the highest performance is obtained with an RMSE of 0.0892, MAE of 0.0592, and MAPE of 0.2061 in the testing stage. Moreover, in the validation stage, the predictive model obtains a very acceptable performance with values between 0.4152 and 3.9084 for RMSE, and a MAPE of less than 4.1%, using three different datasets with indoor environment values.