Emergence of the reproduction matrix in epidemic forecasting

Hossein Gorji, Noé Stauffer, Ivan LunatiJournal of The Royal Society Interface 2024 · DOI

TL;DR

This paper introduces a "reproduction matrix" to go beyond the limitations of the scalar reproduction number ($R(t)$) for epidemic forecasting. By accounting for population stratification (like age or vaccination status) and non-linear dynamics, their method provides more accurate predictions of epidemic trajectories, especially for the timing and magnitude of infection peaks, as demonstrated on synthetic and real-world COVID-19 data.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical role of the reproduction number, $R(t)$, in guiding public health interventions. Conventionally, $R(t)$ is treated as a scalar value that indicates whether an epidemic will grow or decay. However, this simplification fails to capture the complexities of real-world epidemics, which are influenced by factors like finite population sizes, age stratification, vaccination status, and heterogeneous mixing patterns. These factors lead to non-linear dynamics and can cause the epidemic trajectory to deviate significantly from simple exponential growth. Existing methods often oversimplify these complexities, leading to inaccurate forecasts, particularly in predicting the peak of an epidemic wave. This paper aims to address this gap by developing a more sophisticated forecasting framework that explicitly incorporates population heterogeneity and non-linear effects.

Methods

The core of the paper's contribution lies in the generalization of the scalar reproduction number to a reproduction matrix, denoted as $R(t)$. This matrix, $R_{ij}(t)$, quantifies the expected number of secondary infections in class $i$ generated by an individual infected in class $j$ at time $t$. This is achieved by extending the standard renewal equation to a multidimensional renewal equation that accounts for the susceptibility of different population classes.

The forecasting scheme involves two main steps:

  1. Computing the Reproduction Matrix: The reproduction matrix $R(t)$ is estimated from observed incidence data using a derived formula (Equation 2.6) that incorporates contact matrices, infectiousness kernels, and generation time distributions. This step requires more detailed input than traditional methods but allows for a richer representation of epidemic dynamics.
  2. Forecasting with a Transfer Functional: Once the initial reproduction matrix $R(t_0)$ and susceptibility are known, a transfer functional ($G$) is used to project the epidemic course into the future. This functional, derived from the multidimensional renewal equation, allows for iterative updates of the reproduction matrix and incidence over time. The paper also outlines a time-marching Monte Carlo scheme to quantify the uncertainty in these forecasts.

Datasets

The paper evaluates its method using two types of datasets:

  • Synthetic Test Cases: These are generated using an SIR (susceptible-infectious-recovered) compartmental model, designed to mimic heterogeneous populations (e.g., age-stratified, vaccinated vs. unvaccinated). This allows for controlled testing of the model's performance against known dynamics.
  • Real-World COVID-19 Data: The authors analyze data from the Omicron waves in Switzerland and Scotland during Spring 2022. These datasets were chosen for their relatively low initial immunity and limited spatial heterogeneity, making them suitable for testing the model's ability to capture real epidemic dynamics without incorporating complex spatial or mobility factors.

Key Insights and Ideas

  • The scalar $R(t)$ is insufficient for complex epidemics: The paper strongly argues that relying on a single scalar value for $R(t)$ leads to significant overestimation of epidemic peaks and an inaccurate understanding of epidemic trajectories, especially in heterogeneous populations.
  • Heterogeneity matters: The stratification of a population (by age, vaccination status, etc.) is not a minor detail but a crucial factor that fundamentally alters epidemic dynamics. The reproduction matrix captures these inter-class dependencies.
  • Non-linearity is key: Finite population sizes and the resulting depletion of susceptibles introduce non-linear effects that cannot be ignored. The multidimensional renewal equation explicitly accounts for this.
  • A matrix approach offers a more holistic view: The reproduction matrix provides a more complete picture of epidemic spread by detailing how infections propagate between different population groups, rather than just an average rate.
  • Forecasting requires more than just current trends: Simple exponential extrapolation or even homogeneous models fail because they don't account for how the epidemic's characteristics (like susceptibility) change over time within different groups.

Key Results

  • Improved Peak Prediction: The proposed heterogeneous forecasting scheme, using the reproduction matrix, significantly improves the accuracy of predicting the timing and magnitude of epidemic peaks compared to scalar $R(t)$ or homogeneous models. This is demonstrated on both synthetic and real-world COVID-19 data.
  • Realistic Trajectories: The model generates more realistic epidemic trajectories that better reflect the complex interplay of different population groups.
  • Robustness to Uncertainty: The Monte Carlo approach for uncertainty quantification shows that the model remains relatively robust even with uncertainties in input parameters, such as the contact matrix.
  • Overestimation by Scalar $R(t)$: In synthetic scenarios, the scalar $R(t)$ consistently overestimates the peak incidence, leading to an "overshoot" in predictions. This is because it assumes the initial growth rate can be sustained indefinitely, ignoring the impact of waning susceptibility in specific cohorts.

Limitations and Future Work

  • Parameter Estimation: While the paper outlines how to estimate the reproduction matrix, it acknowledges that obtaining accurate input parameters (contact matrices, infectiousness, generation times) for real-world scenarios can be challenging and requires careful consideration of data quality and potential biases (e.g., under-reporting).
  • Time-Invariant Assumptions: The current model assumes that certain parameters (like mitigation measures, contact rates, infectivity) remain constant during the forecasting period. The authors note that this can be a limitation for rapidly changing epidemic situations.
  • Spatial and Mobility Factors: The current framework does not explicitly incorporate spatial heterogeneity or population mobility, which can be significant drivers of epidemic spread.
  • Future Work: The authors suggest several promising avenues for future research:
  • Time-Dependent Parameters: Developing methods to incorporate time-varying mitigation measures, contact rates, infectivity, and vaccine protection.
  • Spatial Integration: Combining their approach with models that account for spatial heterogeneity and epidemicity indices.
  • Data-Driven Granularity: Exploring methods to learn the necessary model granularity directly from data, rather than relying solely on predefined stratifications.
  • Generalization: Applying the framework to other types of heterogeneous systems described by renewal processes, beyond epidemiology.

Summary updated

Abstract

During the recent COVID-19 pandemic, the instantaneous reproduction number, <i>R</i>(<i>t</i>), has surged as a widely used measure to target public health interventions aiming at curbing the infection rate. In analogy with the basic reproduction number that arises from the linear stability analysis, <i>R</i>(<i>t</i>) is typically interpreted as a threshold parameter that separates exponential growth (<i>R</i>(<i>t</i>) > 1) from exponential decay (<i>R</i>(<i>t</i>) < 1). In real epidemics, however, the finite number of susceptibles, the stratification of the population (e.g. by age or vaccination state), and heterogeneous mixing lead to more complex epidemic courses. In the context of the multidimensional renewal equation, we generalize the scalar <i>R</i>(<i>t</i>) to a reproduction matrix, [Formula: see text], which details the epidemic state of the stratified population, and offers a concise epidemic forecasting scheme. First, the reproduction matrix is computed from the available incidence data (subject to some <i>a priori</i> assumptions), then it is projected into the future by a transfer functional to predict the epidemic course. We demonstrate that this simple scheme allows realistic and accurate epidemic trajectories both in synthetic test cases and with reported incidence data from the COVID-19 pandemic. Accounting for the full heterogeneity and nonlinearity of the infection process, the reproduction matrix improves the prediction of the infection peak. In contrast, the scalar reproduction number overestimates the possibility of sustaining the initial infection rate and leads to an overshoot in the incidence peak. Besides its simplicity, the devised forecasting scheme offers rich flexibility to be generalized to time-dependent mitigation measures, contact rate, infectivity and vaccine protection.
BibTeX
@article{gorji2024emergence,
  author = {Hossein Gorji and Noé Stauffer and Ivan Lunati},
  title = {{Emergence of the reproduction matrix in epidemic forecasting}},
  journal = {Journal of The Royal Society Interface},
  year = {2024},
  doi = {10.1098/rsif.2024.0124},
}
Figure