A novel ensemble ARIMA-LSTM approach for evaluating COVID-19 cases and future outbreak preparedness.

Abstract

Background: The global impact of the highly contagious COVID-19 virus has created unprecedented challenges, significantly impacting public health and economies worldwide. This research article conducts a time series analysis of COVID-19 data across various countries, including India, Brazil, Russia, and the United States, with a particular emphasis on total confirmed cases.

Methods: The proposed approach combines auto-regressive integrated moving average (ARIMA)'s ability to capture linear trends and seasonality with long short-term memory (LSTM) networks, which are designed to learn complex nonlinear dependencies in the data. This hybrid approach surpasses both individual models and existing ARIMA-artificial neural network (ANN) hybrids, which often struggle with highly nonlinear time series like COVID-19 data. By integrating ARIMA and LSTM, the model aims to achieve superior forecasting accuracy compared to baseline models, including ARIMA, Gated Recurrent Unit (GRU), LSTM, and Prophet.

Results: The hybrid ARIMA-LSTM model outperformed the benchmark models, achieving a mean absolute percentage error (MAPE) score of 2.4%. Among the benchmark models, GRU performed the best with a MAPE score of 2.9%, followed by LSTM with a score of 3.6%.

Conclusions: The proposed ARIMA-LSTM hybrid model outperforms ARIMA, GRU, LSTM, Prophet, and the ARIMA-ANN hybrid model when evaluating using metrics like MAPE, symmetric mean absolute percentage error, and median absolute percentage error across all countries analyzed. These findings have the potential to significantly improve preparedness and response efforts by public health authorities, allowing for more efficient resource allocation and targeted interventions.