Deep learning for disease outbreak prediction: a parallel LSTM-CNN model

Area of Science:

• Epidemiology and Public Health
• Computational Biology
• Dynamical Systems Theory

Background:

• Early warning signals (EWSs) are crucial for pandemic prevention, but emerging diseases present unique dynamics and noisy data challenges.
• Traditional time-series classification (TSC) methods struggle with the complexities of real-world outbreak data.
• Dynamical systems principles offer a framework for understanding disease spread, even for novel pathogens.

Purpose of the Study:

• To develop a robust deep learning model for reliable early warning signals (EWSs) in disease outbreak prediction.
• To address the challenges posed by noisy measurements and unique disease dynamics in outbreak surveillance.
• To improve the accuracy and applicability of EWSs in real-world public health crises.

Main Methods:

• Utilized a parallel long short-term memory-convolutional neural network deep learning architecture for TSC.
• Trained the model on two simulated datasets: one modeling novel disease behaviors via polynomial dynamical systems, and another simulating noise-induced dynamics.
• Evaluated model performance on diverse simulated data and real-world datasets including influenza, COVID-19, and monkeypox.

Main Results:

• The proposed deep learning model demonstrated superior performance compared to existing models and statistical indicators across most datasets.
• The model effectively provided early warning signals (EWSs) for impending outbreaks under various simulated and real-world conditions.
• The parallel LSTM-CNN model proved robust in handling noisy data inherent in disease outbreak measurements.

Conclusions:

• Advancements in deep learning, specifically the parallel LSTM-CNN model, significantly enhance the capability to provide improved EWSs.
• The model's effectiveness in noisy environments makes it highly applicable for real-world emerging disease outbreak surveillance and prediction.
• This research bridges sophisticated computational methods with practical public health needs for better pandemic preparedness.