The predictive skill of convolutional neural networks models for disease forecasting

Area of Science:

• Epidemiology
• Machine Learning
• Computational Biology

Background:

• Deep learning, particularly Recurrent Neural Networks (RNNs), has shown success in forecasting diseases like Influenza-Like Illness (ILI).
• Conventional models like ARIMA have limitations in capturing complex disease dynamics.

Purpose of the Study:

• To investigate the effectiveness of one-dimensional Convolutional Neural Networks (CNNs) for epidemiological forecasting.
• To compare the performance of CNN-based models against traditional RNNs in disease forecasting.

Main Methods:

• Adaptation of two neural networks utilizing one-dimensional temporal convolutional layers: Temporal Convolutional Networks and Simple Neural Attentive Meta-Learners.
• Testing these CNN models using US influenza data from 2010-2019.

Main Results:

• Epidemiological forecasting using CNNs is demonstrated to be feasible.
• CNN models achieved forecasting skill comparable to, and in some instances superior to, standard RNNs.
• CNNs and RNNs introduce nonlinear transformations, enhancing purely data-driven epidemiological models.

Conclusions:

• One-dimensional CNNs are a viable and effective tool for epidemiological forecasting.
• CNNs offer a powerful alternative or complement to RNNs in disease prediction models.
• The integration of nonlinear transformations via CNNs and RNNs advances data-driven epidemiological modeling.