A Comparative Analysis of Traditional and Deep Learning Time Series Architectures for Influenza A Infectious Disease Forecasting

TL;DR

This study compares traditional and deep learning models for Influenza A forecasting, demonstrating that advanced deep learning architectures, especially Transformers, significantly outperform classical methods in predictive accuracy.

Contribution

It provides a comprehensive comparison showing deep learning models, particularly Transformers, offer superior performance over traditional models in infectious disease forecasting.

Findings

Deep learning models outperform traditional ARIMA and ETS models.

Transformers achieve the lowest MSE and MAE among tested architectures.

Results support using deep learning for improved public health forecasting.

Abstract

Influenza A is responsible for 290,000 to 650,000 respiratory deaths a year, though this estimate is an improvement from years past due to improvements in sanitation, healthcare practices, and vaccination programs. In this study, we perform a comparative analysis of traditional and deep learning models to predict Influenza A outbreaks. Using historical data from January 2009 to December 2023, we compared the performance of traditional ARIMA and Exponential Smoothing(ETS) models with six distinct deep learning architectures: Simple RNN, LSTM, GRU, BiLSTM, BiGRU, and Transformer. The results reveal a clear superiority of all the deep learning models, especially the state-of-the-art Transformer with respective average testing MSE and MAE of 0.0433 \pm 0.0020 and 0.1126 \pm 0.0016 for capturing the temporal complexities associated with Influenza A data, outperforming well known traditional baseline ARIMA and ETS models. These findings of this study provide evidence that state-of-the-art deep learning architectures can enhance predictive modeling for infectious diseases and indicate a more general trend toward using deep learning methods to enhance public health forecasting and intervention planning strategies. Future work should focus on how these models can be incorporated into real-time forecasting and preparedness systems at an epidemic level, and integrated into existing surveillance systems.