Abstract: Infectious disease modeling and forecasting have played a key role in helping assess and respond to epidemics and pandemics. Recent work has leveraged data on disease peak infection and peak hospital incidence to fit compartmental models for the purpose of forecasting and describing the dynamics of a disease outbreak. Incorporating these data can greatly stabilize a compartmental model fit on early observations, where slight perturbations in the data may lead to model fits that project wildly unrealistic peak infection. We introduce a new method for incorporating historic data on the value and time of peak incidence of hospitalization into the fit for a Susceptible-Infectious-Recovered (SIR) model by formulating the relationship between an SIR model's starting parameters and peak incidence as a system of two equations that can be solved computationally. This approach is assessed for practicality in terms of accuracy and speed of computation via simulation. To exhibit the modeling potential, we update the Dirichlet-Beta State Space modeling framework to use hospital incidence data, as this framework was previously formulated to incorporate only data on total infections.
The Challenge: Distinguishing Incidence from Prevalence
When a new disease outbreak begins, public health officials rely on compartmental models like the SIR (Susceptible-Infectious-Recovered) model to predict when the peak will hit and how many hospital beds will be needed. However, fitting these models to sparse, noisy data at the start of an epidemic is notoriously difficult—small fluctuations in early case counts can lead to wildly unrealistic predictions of a massive future peak.
To stabilize these models, researchers often use "priors" based on historical data from similar past outbreaks (like previous flu seasons). Specifically, they look at the value and timing of the peak. The catch is that most existing mathematical shortcuts for this stabilization rely on prevalence (the total number of people sick at a given time). In reality, almost all public health data is reported as incidence (the number of new cases or hospitalizations per day).
Until now, researchers often naively treated incidence data as if it were prevalence. This paper addresses this gap by providing a rigorous way to map peak incidence data directly to the internal parameters of a forecasting model.
Key Insights: The Timing Gap
A key "aha moment" in this research is the visualization that peak incidence and peak prevalence do not occur at the same time. As shown in the figure below, incidence usually peaks earlier than prevalence. Failing to account for this lag leads to systematic timing errors in forecasts and biased biological parameters.
If researchers naively use incidence data in a prevalence-based model, the underlying parameters—transmission rate ($\beta$) and recovery rate ($\gamma$)—will be biased. This is dangerous if health officials use those parameters to estimate the "basic reproduction number" ($R_0$).
Methods and Computational Strategies
The authors formulate the relationship between an SIR model’s starting parameters and the Peak Incidence Value (PIV) and Peak Incidence Time (PIT) as a system of two equations. Because these equations involve complex integrals without simple "closed-form" solutions, the authors evaluated four different computational strategies:
- Compute Integral: A brute-force numerical integration.
- Taylor Approximation: A simplified algebraic shortcut.
- Single ODE Approximation: Solving a single Ordinary Differential Equation.
- Full ODE Approximation: Numerically solving the entire SIR system for every iteration.
Key Results
The researchers validated their approach using 1,000 simulations and historical CDC seasonal influenza data. The "Compute Integral" method emerged as the clear winner for practical application.
As shown in the performance table, the "Compute Integral" method achieved near-zero error for Peak Incidence Time (PIT) and very low error for Peak Incidence Value (PIV), while remaining fast enough for real-time use (approximately 0.5 seconds per calculation).
When applied to the 2010 US flu season, the incidence-aware model produced significantly tighter prediction intervals early in the outbreak compared to the old prevalence-based method. Furthermore, the new method yielded median estimates for the reproduction number ($S_0\beta/\gamma$) that were more consistent with biological reality (1.09 for the incidence model vs. 1.21 for the misspecified prevalence model).
Limitations and Future Work
The primary limitation is that the current mapping is specific to the standard SIR model. Real-world diseases often require more complex structures, such as the SEIR model (including an "Exposed" stage) or models that explicitly track hospitalizations (SEIRH). Future work will focus on developing similar "incidence-to-parameter" maps for these more complex models and finding even faster, high-accuracy approximations for large-scale simulations.
Summary updated