A Recurrent Neural Network and Differential Equation based Spatiotemporal Infectious Disease Model with Application to COVID-19

Zhijian Li, Yunling Zheng, Jack Xin, Guofa Zhou

TL;DR

This paper introduces IeRNN, a novel spatiotemporal model that combines a simplified SIR epidemic model with a Recurrent Neural Network (RNN) to forecast infectious disease spread. It excels at predicting COVID-19 trends, especially with limited data, by effectively capturing both local temporal dynamics and latent spatial interactions between regions.

What Problem Does This Paper Address?

Accurately forecasting infectious disease outbreaks, like COVID-19, is crucial for effective public health interventions. While data-driven methods like Recurrent Neural Networks (RNNs) are powerful, they often struggle with limited daily data, leading to poor performance. Traditional epidemic models, such as the SIR model, are good at capturing temporal trends within a single region but fail to account for spatial interactions (e.g., how infections spread between neighboring areas). Existing spatiotemporal models often require substantial amounts of data, which is a significant limitation for emerging diseases like COVID-19 where data is initially scarce. This paper aims to bridge this gap by developing a model that is both accurate with limited data and captures crucial spatiotemporal dynamics.

Methods

The core of the proposed IeRNN model is a hybrid approach that integrates two key components:

  • Discrete-Time SIR Model with Latent Spatial Forcing (Ie-Equation): The authors start with the classic SIR differential equations and adapt them into a discrete-time form. Crucially, they introduce a term, $I_e$, representing "external infection influence" from neighboring regions. This $I_e$ term is treated as latent information because it's not directly observed. The simplified SIR model with this latent forcing captures the temporal dynamics within a region.
  • Recurrent Neural Network (RNN) for Latent Spatial Information: To model the unobserved $I_e$ term, the researchers employ an RNN, specifically using Long Short-Term Memory (LSTM) cells. This RNN takes the infection data from neighboring regions as input and learns to predict the latent spatial influence ($I_e$) for the target region. This allows the model to capture how infections in one area affect others, even if those connections aren't explicitly defined by simple adjacency.

The model is trained end-to-end, optimizing both the parameters of the Ie-equation and the weights of the RNN simultaneously.

Datasets

The model was trained and tested using COVID-19 data from Italy. Specifically, they used daily new (current) case data for 20 Italian regions from February 24, 2020, to June 18, 2020. This dataset is noted for its relatively accurate reporting and complete duration of the epidemic's rise and fall compared to other available datasets.

Key Insights and Ideas

  • Latent Spatial Information is Key: The paper highlights that explicitly modeling the unobserved "latent spatial information" ($I_e$) is crucial for improving infectious disease forecasting, especially when data is limited. Simply relying on temporal data or basic spatial adjacency is insufficient.
  • Hybrid Models Offer the Best of Both Worlds: Combining a simplified mechanistic model (SIR) with a data-driven approach (RNN) allows the model to leverage the interpretability and temporal modeling strengths of the SIR framework while using the RNN to capture complex, unobserved spatial interactions.
  • Robustness with Limited Data: The IeRNN model demonstrates remarkable performance even when trained on a reduced dataset (40% of the available data), outperforming other models that struggle significantly under such conditions. This is a critical advantage for forecasting emerging diseases.
  • Compactness and Efficiency: The discrete-time SIR component makes the temporal part of the model more compact than a pure LSTM approach, contributing to its efficiency and ability to perform well with less data.

Key Results

  • Superior Forecasting Accuracy: IeRNN consistently outperforms baseline models (I-model, fully-connected NN, and ARIMA) in 1-day, 3-day, and 7-day ahead forecasting across multiple Italian regions. This superiority is particularly pronounced when using limited training data.
  • Reduced RMSE: In 1-day ahead forecasts, IeRNN achieved the smallest Root Mean Square Errors (RMSE) compared to other models, especially when trained on 70% of the data. Even with only 40% of the data, IeRNN maintained competitive performance.
  • Effective with Limited Data: The model's ability to track the general trend of infected populations even with only 40% of the training data is a significant finding.
  • Computational Efficiency: The model is computationally efficient, with training times under 1 second and inference times in fractions of a second, making it suitable for real-time applications.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the $I_e$ term is a latent variable and its direct mathematical modeling is difficult.
  • The current model focuses on interactions with nearest geographic neighbors.
  • Future Work:
    • The authors suggest strengthening the I-equation by incorporating social and control mechanisms.
  • They propose expanding the spatial interaction modeling to include data beyond nearest neighbors, potentially using traffic data to better understand inter-regional movement and its impact on disease spread.

Summary updated

Abstract

The outbreaks of Coronavirus Disease 2019 (COVID-19) have impacted the world significantly. Modeling the trend of infection and real-time forecasting of cases can help decision making and control of the disease spread. However, data-driven methods such as recurrent neural networks (RNN) can perform poorly due to limited daily samples in time. In this work, we develop an integrated spatiotemporal model based on the epidemic differential equations (SIR) and RNN. The former after simplification and discretization is a compact model of temporal infection trend of a region while the latter models the effect of nearest neighboring regions. The latter captures latent spatial information. We trained and tested our model on COVID-19 data in Italy, and show that it out-performs existing temporal models (fully connected NN, SIR, ARIMA) in 1-day, 3-day, and 1-week ahead forecasting especially in the regime of limited training data.
BibTeX
@inproceedings{li2020recurrent,
  author = {Zhijian Li and Yunling Zheng and Jack Xin and Guofa Zhou},
  title = {{A Recurrent Neural Network and Differential Equation based Spatiotemporal Infectious Disease Model with Application to COVID-19}},
  booktitle = {KDIR},
  year = {2020},
  url = {https://openreview.net/forum?id=cxjPlMGdQI},
}
Figure