Neural networks for endemic measles dynamics: comparative analysis and integration with mechanistic models

Wyatt Madden, Wei Jin, Benjamin A. Lopman, Andreas Züfle, Benjamin D. Dalziel, C. Jessica E. Metcalf, Bryan T. Grenfell, Max S. Y. LaumedRxiv 2024 · DOI

TL;DR

Researchers have developed a hybrid modeling approach that combines the predictive power of neural networks with the biological rigor of mechanistic disease models to forecast measles outbreaks. Their "Physics-Informed" approach not only outperforms traditional models in long-term forecasting but also successfully uncovers the hidden spatial hierarchies of how disease spreads from major cities to smaller towns.

What Problem Does This Paper Address?

Predicting infectious disease outbreaks like measles is notoriously difficult because the dynamics are highly nonlinear and influenced by complex factors: seasonal changes (like school terms), birth rates, and the movement of people between cities.

Epidemiologists traditionally use mechanistic models (like the Time-Series SIR or TSIR), which are great for understanding why a disease spreads but often struggle with long-term accuracy or capturing complex spatial interactions. On the flip side, machine learning (ML) models are excellent at finding patterns and making predictions but are often "black boxes" that don't provide biological insight or help infer critical parameters like the basic reproduction number ($R_0$). This paper seeks to bridge that gap by integrating the two approaches.

Methods

The authors take a two-pronged approach:

  1. Spatial Feature Neural Network (SFNN): They built a high-dimensional feed-forward neural network that includes spatial features, such as the distance to the nearest "core cities" (the seven most populous cities in England and Wales). They used SHAP (SHapley Additive exPlanations) values—a technique from "Explainable AI"—to peek inside the black box and see which features the model prioritized.
  2. TSIR-PINN (Physics-Informed Neural Network): This is the paper's most novel contribution. PINNs are neural networks that include physical or biological equations (like differential equations for disease transmission) directly into their loss function. The authors improved this by using the TSIR model to reconstruct "latent" data—specifically, the number of susceptible individuals in a population—and feeding that into the PINN. This helps the neural network respect the underlying biology of the S-I-R (Susceptible-Infectious-Recovered) framework.

Datasets

The study utilizes a gold-standard dataset in epidemiology: measles incidence data from England and Wales (1944–1965).

  • Scope: 1,452 cities and towns.
  • Granularity: Biweekly reports over 21 years.
  • Significance: This pre-vaccination era data is famous for its "biennial" (every two years) cycles and provides a rigorous testbed for nonlinear dynamics.

Key Insights and Ideas

  • Geography Emerges Naturally: One of the most striking "aha moments" was that the SFNN, without being given any explicit rules about geography, "learned" a gravity-model-like hierarchy. It correctly identified that outbreaks in major cities drive the dynamics in smaller, surrounding towns.
  • The Power of Latent Variables: Pure neural networks often struggle to "see" the people who might get sick (the susceptibles). By using a mechanistic model (TSIR) to estimate this hidden population and then using those estimates to guide the PINN, the researchers significantly improved both the model's accuracy and its ability to estimate biological parameters.
  • Short-term vs. Long-term Trade-offs: The study found that while traditional mechanistic models are still slightly better at very short-term "nowcasting" (1–2 weeks ahead) in big cities, neural networks are superior for longer windows and for smaller, more volatile populations.

Key Results

  • Forecasting Accuracy: The SFNN outperformed the traditional TSIR model across most forecasting windows, particularly for long-term predictions (up to 52 biweeks ahead).
  • Parameter Inference: The hybrid TSIR-PINN model was remarkably effective at estimating $R_0$. In London, it estimated an $R_0$ of 26.9, which aligns with established epidemiological literature, whereas a "naive" PINN (without the TSIR integration) failed to converge on a realistic number, estimating an $R_0$ of only 3.0.
  • Interpretability: Using SHAP values, the authors proved that the model's predictions for small towns were heavily dependent on the infection status of nearby major cities, providing a post-hoc validation of the "hierarchical spread" theory.

Limitations and Future Work

The authors acknowledge several areas for growth:

  • Point Estimates: Currently, the models provide single-point forecasts. Future versions need to incorporate "rigorous statistical uncertainty" to provide a range of possible outcomes (probabilistic forecasting).
  • Architecture Complexity: The study used standard feed-forward networks. The authors suggest that more advanced architectures, like Convolutional Neural Networks (CNNs) or LSTMs, could better capture the "memory" of an epidemic.
  • Scalability of PINNs: While the SFNN was applied to over 1,400 locations, the more complex PINN was only tested on London data. Scaling the physics-informed approach to a full multi-city "metapopulation" remains a significant computational challenge.

Summary updated

Abstract

Abstract Measles is an important infectious disease system both for its burden on public health and as an opportunity for studying nonlinear spatio-temporal disease dynamics. Traditional mechanistic models often struggle to fully capture the complex nonlinear spatio-temporal dynamics inherent in measles outbreaks. In this paper, we first develop a high-dimensional feed-forward neural network model with spatial features (SFNN) to forecast endemic measles outbreaks and systematically compare its predictive power with that of a classical mechanistic model (TSIR). We illustrate the utility of our model using England and Wales measles data from 1944-1965. These data present multiple modeling challenges due to the interplay between metapopulations, seasonal trends, and nonlinear dynamics related to demographic changes. Our results show that, while the TSIR model yields more accurate very short-term (1 to 2 biweeks ahead) forecasts for highly populous cities, overall, our neural network model (SFNN) outperforms the TSIR in other forecasting windows. Furthermore, we show that our spatial-feature neural network model, without imposing mechanistic assumptions a priori , can uncover gravity-model-like spatial hierarchy of measles spread in which major cities play an important role in driving regional outbreaks. We then turn our attention to integrative approaches that combine mechanistic and machine learning models. Specifically, we investigate how the TSIR can be utilized to improve a state-of-the-art approach known as Physics-Informed-Neural-Networks (PINN) which explicitly combines compartmental models and neural networks. Our results show that the TSIR can facilitate the reconstruction of latent susceptible dynamics, improving both forecasts and parameter inference of measles dynamics within the PINN. In summary, our results show that appropriately designed neural network-based models can outperform traditional mechanistic models for short to long-term forecasts, while simultaneously providing mechanistic interpretability. Our work also provides valuable insights into more effectively integrating machine learning models with mechanistic models to enhance public health responses to measles and similar infectious disease systems. Author summary Mechanistic models have been foundational in developing an understanding of the transmission dynamics of infectious diseases including measles. In contrast to their mechanistic counterparts, machine learning techniques including neural networks have primarily focused on improving forecasting accuracy without explicitly inferring transmission dynamics. Effectively integrating these two modeling approaches remains a central challenge. In this paper, we first develop a high-dimensional neural network model to forecast spatiotemporal endemic measles outbreaks and systematically compare its predictive power with that of a classical mechanistic model (TSIR). We illustrate the utility of our model using a detailed dataset describing measles outbreaks in England and Wales from 1944-1965, one of the best-documented and most-studied nonlinear infectious disease systems. Our results show that, overall, our neural network model outperforms the TSIR in all forecasting windows. Furthermore, we show that our neural network model can uncover the mechanism of hierarchical spread of measles where major cities drive regional outbreaks. We then develop an integrative approach that explicitly and effectively combines mechanistic and machine learning models, improving simultaneously both forecasting and inference. In summary, our work offers valuable insights into the effective utilization of machine learning models, and integration with mechanistic models, for enhancing outbreak responses to measles and similar infectious disease systems.
BibTeX
@misc{madden2024neural,
  author = {Wyatt Madden and Wei Jin and Benjamin A. Lopman and Andreas Züfle and Benjamin D. Dalziel and C. Jessica E. Metcalf and Bryan T. Grenfell and Max S. Y. Lau and Max S. Y. Lau},
  title = {{Neural networks for endemic measles dynamics: comparative analysis and integration with mechanistic models}},
  year = {2024},
  doi = {10.1101/2024.05.28.24307979},
  howpublished = {medRxiv},
}
Figure