Abstract
Reliable forecasts of influenza can aid in the control of both seasonal and pandemic outbreaks. We introduce a simulation optimization (SIMOP) approach for forecasting the influenza epidemic curve. This study represents the final step of a project aimed at using a combination of simulation, classification, statistical and optimization techniques to forecast the epidemic curve and infer underlying model parameters during an influenza outbreak. The SIMOP procedure combines an individual-based model and the Nelder-Mead simplex optimization method. The method is used to forecast epidemics simulated over synthetic social networks representing Montgomery County in Virginia, Miami, Seattle and surrounding metropolitan regions. The results are presented for the first four weeks. Depending on the synthetic network, the peak time could be predicted within a 95% CI as early as seven weeks before the actual peak. The peak infected and total infected were also accurately forecasted for Montgomery County in Virginia within the forecasting period. Forecasting of the epidemic curve for both seasonal and pandemic influenza outbreaks is a complex problem, however this is a preliminary step and the results suggest that more can be achieved in this area.
Introduction
Forecasting the "epidemic curve"—the daily or weekly count of new infections—is a critical objective for public health officials. Accurate forecasts allow for better resource allocation, such as vaccine distribution and hospital staffing. However, traditional forecasting often relies on deterministic compartmental models (like basic SIR models) that assume a "well-mixed" population, ignoring the complex reality of human movement and interaction.
While individual-based models (IBMs) capture this complexity, they are computationally expensive and difficult to "tune" in real-time when a novel strain emerges. This study addresses the challenge of rapidly estimating unknown disease parameters (like transmissibility and incubation periods) from noisy, partial surveillance data during the early stages of an outbreak.
Methodology: The SIMOP Framework
The researchers developed a two-stage approach. Initially, a Random Forest classifier determines if a current outbreak matches any previously simulated scenarios in a pre-computed library. If no match is found, indicating a novel outbreak, the Simulation Optimization (SIMOP) procedure is triggered.
The SIMOP framework consists of three core components:
- Individual-Based Model (IBM): A stochastic simulation over a synthetic social contact network that models individuals moving between home, work, and school using an SEIR (Susceptible, Exposed, Infectious, Recovered) framework.
- Nelder-Mead Simplex Optimization: Because the IBM is stochastic and non-differentiable, the authors utilize the Nelder-Mead method—a derivative-free "direct search" algorithm—to explore the parameter space (transmissibility, incubation, and infectious period distributions).
- Optimization Loop: The algorithm proposes parameters, runs the simulation, compares the resulting curve to observed data using a Sum of Squared Errors (SSQ) objective function, and adjusts parameters to minimize error.
Experimental Setup and Datasets
The study utilizes synthetic social networks to ensure a known "ground truth" for validation. These networks were built using US Census data and travel surveys to mimic realistic contact patterns in three distinct regions:
- Seattle: 3.2 million people.
- Miami: 2 million people.
- Montgomery County, VA: 0.16 million people.
To simulate real-world conditions, the authors added 15% and 25% noise to the "true" epidemic curves, representing the reporting errors and lags common in public health surveillance.
Key Results and Regional Insights
The model demonstrated that regional specificity is vital. Demographics, such as the proportion of children who drive flu spread, vary significantly between regions, necessitating localized forecasting.
Key performance metrics include:
- Peak Timing: Depending on the city, the peak time could be predicted within a 95% confidence interval as early as seven weeks before the actual peak.
- Magnitude Accuracy: For Montgomery County, the peak number of infected individuals and total cumulative infections were accurately forecasted within the first four weeks.
- Noise Tolerance: Even with 25% noise, the model predicted peak timing within a one-week margin of error by day 28 of the outbreak.
- Convergence: The SIMOP approach showed a clear reduction in Root Mean Squared Error (RMSE) as more temporal data was provided, proving the optimization loop successfully "learned" the outbreak's trajectory.
Limitations and Future Work
While promising, the model currently assumes static social networks, whereas real-world behavior (e.g., social distancing) changes during a pandemic. Future iterations aim to account for biological differences between age groups and adapt the SIMOP approach for multi-wave outbreaks. Additionally, the computational cost of running thousands of IBM simulations requires significant high-performance computing resources, which remains a hurdle for widespread real-time deployment.
Summary updated