Physics-Informed Neural Networks Integrating Compartmental Model for Analyzing COVID-19 Transmission Dynamics

Xiao Ning, Jinxing Guan, Xian Li, Yongyue Wei, Feng ChenViruses 2023 · DOI

TL;DR

This paper introduces a novel Physics-Informed Neural Network (PINN) approach to model COVID-19 transmission dynamics by integrating a modified SEIRD compartmental model with deep neural networks. This method effectively estimates time-varying parameters, offering more accurate predictions and insights into the impact of interventions compared to traditional methods.

What Problem Does This Paper Address?

Predicting and understanding the spread of infectious diseases is crucial for public health. Compartmental models, like the SEIRD model, are powerful tools for this, but they rely on parameters that are often unknown and can change over time due to factors like interventions or viral mutations. Estimating these time-varying parameters from real-world data, which is often noisy and incomplete, presents a significant challenge. Traditional parameter estimation methods struggle with computational costs and the inability to capture dynamic parameter changes, limiting their practical application for real-time epidemic analysis and forecasting.

Methods

The core of this work is the application of Physics-Informed Neural Networks (PINNs) to address the limitations of traditional compartmental modeling. The authors propose a PINN architecture that consists of two main components:

  1. State Network: A deep neural network (DNN) that learns to represent the state variables of the SEIRD model (Susceptible, Exposed, Infectious, Recovered, Dead) over time.
  2. Parameter Network: Another DNN designed to estimate the time-varying parameters of the SEIRD model, specifically the transmission rate ($\beta(t)$), recovery rate ($\gamma(t)$), and death rate ($\mu(t)$).

The PINN framework integrates the known physics (the SEIRD differential equations) with the observed data. The loss function is designed to minimize two key components:

  • Data Mismatch: The difference between the network's predicted states and the actual observed data.
  • ODE Residual: How well the network's predicted states satisfy the underlying SEIRD differential equations.

This dual-network approach, combined with a modified tanh activation function for constant parameters ($\alpha$, $\epsilon$), allows the model to learn complex dynamics even with limited data. The use of ResNet-like skip connections in the neural network architecture further aids in gradient flow and model performance.

Datasets

The study utilizes real-world COVID-19 data from Italy. Specifically, the data includes:

  • Cumulative infected, recovered, and deceased cases.
  • The data spans from February 20, 2020, to June 30, 2020.
  • A 7-day moving average was applied to smooth the data and mitigate weekly fluctuations.

No new benchmark datasets were introduced; the focus was on applying the method to existing, publicly available data.

Key Insights and Ideas

  • PINNs for Dynamic Parameters: The paper demonstrates that PINNs can effectively learn and estimate time-varying parameters in epidemiological models, a significant advancement over methods that assume constant parameters.
  • Dual-Network Architecture: Separating the state dynamics from parameter estimation into distinct neural networks allows for a more robust and efficient learning process, especially when dealing with limited data.
  • Physics-Informed Learning: By embedding the differential equations of the SEIRD model directly into the loss function, the PINN approach ensures that the learned dynamics are physically plausible, even when data is sparse.
  • Intervention Impact Analysis: The time-varying transmission rate ($\beta(t)$) estimated by the model clearly reflects the impact of interventions like lockdowns in Italy, showing a significant reduction in transmission.
  • Forecasting Power: The model shows strong performance in predicting future trends (up to 7 days ahead), indicating its potential for real-time epidemic monitoring and decision-making.

Key Results

  • The proposed PINNs method achieved accurate fitting of observed COVID-19 data in Italy for current infections, cumulative recoveries, and cumulative deaths.
  • The model demonstrated robust forecasting capabilities, with MAE, RMSE, and MAPE values indicating high accuracy for 3-day, 5-day, and 7-day predictions. For instance, for 7-day predictions after May 19, 2020, the MAE for infections was 3046, RMSE was 3514, and MAPE was 5.41%.
  • The estimated time-varying transmission rate ($\beta(t)$) showed a significant decrease following the implementation of lockdown measures in Italy.
  • The estimated recovery rate ($\gamma(t)$) showed an increasing trend, and the death rate ($\mu(t)$) showed a decreasing trend over time, which the authors attribute to improved understanding and treatment of the disease.
  • The effective reproduction number ($R_t$) estimated by the PINNs method closely matched results from a serial Bayesian model, validating the approach.

Limitations and Future Work

  • Prior Knowledge Requirement: A key limitation of PINNs, as acknowledged by the authors, is the requirement for prior knowledge of the physical laws and constraints governing the system. If these underlying epidemiological laws are not well understood or if the available data contradicts them, the model's accuracy may be compromised.
  • Model Structure Dependence: The accuracy of the compartmental model itself can impact the PINN's performance. Changes in the underlying epidemiological dynamics might necessitate adjustments to the model structure.
  • Generalizability: While the authors suggest the method can be extended to other regions and diseases, the current study focuses on COVID-19 in Italy. Further validation on diverse datasets and disease contexts would be beneficial.
  • Future Work: The authors suggest extending the method to model and analyze infectious disease transmission dynamics in various regions and for different infectious diseases, especially as more data becomes available. They also highlight the potential for this method to aid in designing appropriate control measures.

Summary updated

Abstract

Modelling and predicting the behaviour of infectious diseases is essential for early warning and evaluating the most effective interventions to prevent significant harm. Compartmental models produce a system of ordinary differential equations (ODEs) that are renowned for simulating the transmission dynamics of infectious diseases. However, the parameters in compartmental models are often unknown, and they can even change over time in the real world, making them difficult to determine. This study proposes an advanced artificial intelligence approach based on physics-informed neural networks (PINNs) to estimate time-varying parameters from given data for the compartmental model. Our proposed PINNs method captures the complex dynamics of COVID-19 by integrating a modified Susceptible-Exposed-Infectious-Recovered-Death (SEIRD) compartmental model with deep neural networks. Specifically, we modelled the system of ODEs as one network and the time-varying parameters as another network to address significant unknown parameters and limited data. Such structure of the PINNs method is in line with the prior epidemiological correlations and comprises the mismatch between available data and network output and the residual of ODEs. The experimental findings on real-world reported data data have demonstrated that our method robustly and accurately learns the dynamics and forecasts future states. Moreover, as more data becomes available, our proposed PINNs method can be successfully extended to other regions and infectious diseases.
BibTeX
@inproceedings{ning2023physics,
  author = {Xiao Ning and Jinxing Guan and Xian Li and Yongyue Wei and Feng Chen},
  title = {{Physics-Informed Neural Networks Integrating Compartmental Model for Analyzing COVID-19 Transmission Dynamics}},
  booktitle = {Viruses},
  year = {2023},
  doi = {10.3390/v15081749},
}
Figure