Unifying Epidemic Models With Mixtures

Arnab Sarker, Ali Jadbabaie, Devavrat Shah

Introduction

Epidemiological modeling typically falls into two distinct camps: mechanistic and non-mechanistic. Mechanistic models, such as the classic Susceptible-Infected-Recovered (SIR) model, utilize differential equations to describe disease transmission based on physical principles. While highly interpretable, these models often struggle with the noise of real-world data and the volatility of human behavioral changes. Conversely, non-mechanistic models rely on statistical curve-fitting. These are flexible and accurate for short-term forecasting but often lack a physical explanation for the observed trends.

This research introduces a novel framework that bridges this gap. By representing infection waves as a mixture of Gaussian curves, the authors provide a model that achieves high accuracy in short-term forecasting while remaining physically interpretable through a concept termed "degree pruning" in social networks.

Theoretical Framework and Methodology

The core of the proposed approach is a Mixture of Gaussians model for time-series data. The number of cases $N(t)$ is modeled as a sum of $r$ components, where each component follows a Gaussian-like trajectory: $e^{-at^2 + bt + c}$.

Bridging SIR and Statistical Curves

To ground this statistical approach in epidemiology, the authors provide a mechanistic justification. They demonstrate that a stochastic SIR process (illustrated in Figure 1) on a social network, where individuals reduce their contact rate ("prune" their social ties) at a constant rate as the epidemic progresses, naturally results in an infection curve that takes a Gaussian form. This links modern networked SIR models with Farr’s Law—the 1840s observation that epidemics follow symmetric bell curves.

The classic SIR model representing the flow of individuals between Susceptible, Infected, and Recovered states.

The Learning Algorithm

Fitting these mixtures is a non-convex optimization challenge. The authors developed a two-stage approach:

  1. Initialization: A peak-finding algorithm identifies mid-points between waves to provide intelligent starting parameters. This method outperformed random initialization significantly, achieving an $R^2$ of 0.984 on synthetic benchmarks.
  2. Alternating Minimization: Gradient descent is used to refine these parameters by minimizing the squared error against observed data.

The Bayesian Information Criterion (BIC) is employed to automatically determine the optimal number of waves ($r$).

Empirical Evaluation

The model was validated using US COVID-19 case data from Johns Hopkins University and the Reich Lab, alongside mobility data from Google and SafeGraph.

Performance and Interpretability

The mixture model demonstrated competitive accuracy, ranking 14th overall out of 34 models in the COVID-19 Forecasting Hub and 4th among models using only case counts. It achieved a median absolute percent error (MAPE) of 15.9% for one-week-ahead forecasts.

The Mixture of Gaussians model applied to US COVID-19 data, showing the decomposition of the total case count into distinct detected outbreaks.

A key insight from the model is the interpretation of the Gaussian "curvature" parameter ($a$). The authors found that higher $a$ values correspond to faster population reactions to the virus (e.g., mask adoption or lockdowns). This was confirmed by a statistically significant correlation ($p < 0.01$) between the learned "degree pruning" parameters and actual SafeGraph mobility data, proving that the model captures real-world behavioral shifts.

Conclusion and Future Directions

While the model excels at short-term forecasting and provides a robust theoretical bridge, it faces challenges at the onset of entirely new waves before sufficient data is available to define a trajectory. Additionally, accuracy degrades for long-term projections (4+ weeks) as it cannot predict future policy changes or viral mutations. Future work aims to integrate real-time mobility data directly into the learning process to anticipate new waves earlier and to develop data-driven control strategies for future public health crises.

Summary updated

Abstract

The COVID-19 pandemic has emphasized the need for a robust understanding of epidemic models. Current models of epidemics are classified as either mechanistic or non-mechanistic: mechanistic models make explicit assumptions on the dynamics of disease, whereas non-mechanistic models make assumptions on the form of observed time series. Here, we introduce a simple mixture-based model which bridges the two approaches while retaining benefits of both. The model represents time series of cases and fatalities as a mixture of Gaussian curves, providing a flexible function class to learn from data compared to traditional mechanistic models. Although the model is non-mechanistic, we show that it arises as the natural outcome of a stochastic process based on a networked SIR framework. This allows learned parameters to take on a more meaningful interpretation compared to similar non-mechanistic models, and we validate the interpretations using auxiliary mobility data collected during the COVID-19 pandemic. We provide a simple learning algorithm to identify model parameters and establish theoretical results which show the model can be efficiently learned from data. Empirically, we find the model to have low prediction error. The model is available live at covidpredictions.mit.edu. Ultimately, this allows us to systematically understand the impacts of interventions on COVID-19, which is critical in developing data-driven solutions to controlling epidemics.
BibTeX
@article{sarker2022unifying,
  author = {Arnab Sarker and Ali Jadbabaie and Devavrat Shah},
  title = {{Unifying Epidemic Models With Mixtures}},
  journal = {IEEE Trans. Signal Inf. Process. Over Networks},
  year = {2022},
  doi = {10.1109/tsipn.2024.3375600},
}
Figure