DSA-BEATS: Dual Self-Attention N-BEATS Model for Forecasting COVID-19 Hospitalization

Abstract

The high number of hospitalization cases of COVID-19 made public health providers overloaded. Forecasting the number of hospitalized patients related to COVID-19 can help public health providers make informed decisions for controlling the spread. In this study, we present the Dual Self-Attention NBEATS (DSA-BEATS) model, a novel approach that effectively combines the self-attention mechanism of transformers with the proficiency of the N-BEATS model in dealing with multivariate forecasting problems. We expanded the dataset to a multivariate one by including data from Canadian transportation hub cities and SARS-CoV-2 RNA load in wastewater, which allowed for a more comprehensive modeling of the complex relationships impacting COVID-19 hospitalizations. These transportation hub cities were the major ports of entry for international travelers coming to the country. The DSA-BEATS model was tested on a 55-day test set with a 12-day horizon, resulting in a Mean Absolute Percentage Error (MAPE) of 14.23%, which implies an accuracy of 85.77%. These results demonstrate substantial improvements over state-of-the-art models such as N-BEATS and Informer, validating the efficacy of the DSA-BEATS model in accurately predicting COVID-19 hospitalizations. The study provides a significant contribution to the ongoing development of enhanced timeseries forecasting methods, particularly in the context of public health crises. The DSA-BEATS model's ability to capture complex temporal relationships and effectively handle multivariate data inputs underscores its potential in a wide range of forecasting tasks beyond the COVID-19 pandemic.