Boosting epidemic forecasting performance with enhanced RNN-type models

The emergence of COVID-19 significantly altered various aspects of everyday life. To capture the dynamics of the pandemic, numerous deterministic and stochastic compartmental models have been proposed. With neural networks (NNs) gaining popularity for their ability to handle complex data and deliver enhanced predictive performance, several deep learning (DL) architectures have been introduced to estimate the progression of the pandemic. These efforts often prioritize advanced network structures over preprocessing techniques or the integration of additional indicators that could improve model performance. This study aims to enhance the predictive capacity of widely used NN models by integrating epidemiological indicators obtained through particle filtering. We investigate how these indicators contribute to the predictive performance of five RNN-type architectures, namely RNN, GRU, LSTM, Bidirectional GRU, and Bidirectional LSTM, and demonstrate that the improvements are independent of the model choice. The inclusion of particle filtering as a preprocessing step bridges the field of DL-based forecasting with Bayesian estimation. The validation process focuses on predicting the number of new COVID-19 cases and deaths in Italy over a 500-day period. Experimental results demonstrate that the proposed methodology significantly enhances performance, particularly for both 1-day and 1-week ahead forecasts, with improvements reaching 95.39% and 53.61% for cases, and 62.15% and 25.30% for deaths, respectively. We emphasize that the enhanced predictive performance originating from the introduced method can support the implementation of more targeted isolation measures and the design of more effective vaccination strategies.