Forecasting infectious disease prevalence with associated uncertainty using neural networks

Overview

• Infectious diseases pose significant human and economic burdens.
• Accurately forecasting disease incidence can enable public health agencies to respond effectively.
• Developing accurate forecasting models remains a significant challenge.
• This thesis proposes two methodological frameworks using neural networks (NNs) with associated uncertainty estimates.

Plain English Explanation

The paper focuses on developing better ways to forecast the spread of infectious diseases, like the flu, in the United States. Accurately predicting the number of people who will get sick can help public health agencies prepare and respond more effectively. However, creating accurate forecasting models has been very difficult.

The researchers propose two new methods that use neural networks, a type of machine learning model. These models not only make predictions, but also provide estimates of how certain they are about those predictions. This is an important feature that has been lacking in previous forecasting approaches.

The first method uses data from web searches related to the flu, along with historical flu case data, to train the neural network models. The best performing model reduced the average error in forecasts by 10.3% and improved the overall forecasting skill by 17.1% compared to other state-of-the-art methods.

The second method uses a type of neural network called a "neural ordinary differential equation" to bridge the gap between traditional mathematical models of disease spread and the more flexible neural network approach. This allows the model to benefit from the physical constraints provided by the mathematical models while still leveraging the power of neural networks.

Technical Explanation

The paper presents two novel methodological frameworks for forecasting infectious disease incidence using neural networks (NNs) with associated uncertainty estimates.

The first proposed method uses web search activity data in conjunction with historical influenza-like illness (ILI) rates as observations for training NN architectures. The models incorporate Bayesian layers to produce uncertainty intervals, positioning themselves as legitimate alternatives to more conventional approaches. The best performing architecture, the iterative recurrent neural network (IRNN), reduces mean absolute error by 10.3% and improves Skill by 17.1% on average in forecasting tasks across four flu seasons compared to the state-of-the-art.

The second framework uses neural ordinary differential equations to bridge the gap between mechanistic compartmental models and NNs, benefiting from the physical constraints that compartmental models provide. The researchers evaluate eight neural ODE models utilizing a mixture of ILI rates and web search activity data to provide forecasts, and compare them with the IRNN and a version of IRNN using only ILI rates (IRNN0). Models trained without web search activity data outperform the IRNN0 by 16% in terms of Skill.

Critical Analysis

The paper presents promising approaches for improving infectious disease forecasting by leveraging neural networks and incorporating uncertainty estimates. The use of web search data, in addition to traditional epidemiological data, is a valuable contribution, as it can capture early signals of disease activity.

However, the paper does not address the potential limitations of web search data, such as biases in who uses search engines or potential delays in the data reflecting actual disease activity. Additionally, the evaluation is limited to influenza-like illness forecasting in the United States, and further research is needed to assess the generalizability of the methods to other infectious diseases and geographic regions.

The neural ordinary differential equation approach is an interesting theoretical framework, but the paper does not provide a clear demonstration of its advantages over the more straightforward IRNN model. Further work is needed to fully capitalize on the benefits of incorporating physical constraints into the neural network architecture.

Conclusion

This thesis proposes two innovative frameworks for infectious disease forecasting using neural networks with uncertainty estimates. The first method leverages web search data to improve forecasting performance, while the second bridges the gap between mechanistic compartmental models and neural networks.

These approaches represent significant progress in the field of epidemic forecasting and have the potential to enable more effective public health decision-making. However, further research is needed to address the limitations and expand the applicability of these methods to a wider range of infectious diseases and settings.