📊 epidemiology

Adaptive multi-model ensembles for improved epidemic projections and decision support

This paper introduces an adaptive multi-model ensemble approach that dynamically selects individual model trajectories based on observed data to improve the accuracy of long-term epidemic projections and short-term forecasting while offering a resource-efficient alternative to traditional coordinated modeling efforts.

Original authors: Fiandrino, S., Paolotti, D., Bay, C., Chinazzi, M., Davis, J. T., Bents, S. J., Perofsky, A. C., Turtle, J. A., Riley, P., Ben-Nun, M., Moore, S. M., Perkins, A., Camargo Espana, G. F., Srivastava, A.

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fiandrino, S., Paolotti, D., Bay, C., Chinazzi, M., Davis, J. T., Bents, S. J., Perofsky, A. C., Turtle, J. A., Riley, P., Ben-Nun, M., Moore, S. M., Perkins, A., Camargo Espana, G. F., Srivastava, A., Aawar, M. A., Bandekar, S. R., Bi, K., Bouchnita, A., Fox, S. J., Meyers, L. A., Venkatramanan, S., Porebski, P., Adiga, A., Lewis, B., Marathe, M., Haghpanah, F., Klein, E., Loo, S. L., Jung, S.-m., Smith, C. P., Contamin, L., Hochheiser, H., Carcelen, E. C., Howerton, E., Shea, K., Yan, K., Runge, M. C., Viboud, C., Pearson, C. A. B., Truelove, S. A., Lessler, J., Borchering, R., Biggerstaff,

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to predict the weather for the next three months. Instead of relying on just one weather forecaster, you ask a whole team of experts. Each expert runs their own computer model, creating hundreds of different "what-if" scenarios: some say it will be a dry summer, others a wet one, some predict a heatwave, and others a cold snap.

In the world of disease modeling, this is exactly what happens. A group of research teams (the "Flu Scenario Modeling Hub") creates a massive collection of computer simulations to predict how many people might get sick with the flu and end up in the hospital. They do this before the season even starts, creating a "static" bundle of all possible futures.

The Problem: The Bundle Gets Stale
The problem with this "static bundle" is that it's like a weather forecast made in September for the whole winter. As the weeks pass and real data comes in (like actual flu cases reported in the news), some of those original predictions start to look silly. Maybe the experts predicted a huge wave of illness, but so far, things are quiet. Or maybe they predicted a mild season, but cases are spiking.

Usually, to update these predictions, you'd have to ask all the experts to re-run their complex computer models with the new data. This takes a lot of time, money, and computing power. It's like asking ten chefs to re-cook a massive banquet every time a guest mentions they are allergic to nuts.

The Solution: The "Adaptive Ensemble" (The Smart Filter)
This paper introduces a clever, low-cost trick called an Adaptive Multi-Model Ensemble.

Think of the original bundle of predictions as a giant bag of marbles. Each marble represents a possible future path of the flu season. Some marbles are red (predicting a mild season), some are blue (predicting a severe one), and some are green (predicting a specific virus type).

The new method acts like a smart filter. Every week, as new real-world data arrives, the researchers look at the bag of marbles. They ask: "Which of these marbles (predictions) look most like what is actually happening right now?"

They then discard the marbles that look wrong (the ones predicting a mild season when people are actually getting very sick) and keep only the top-performing ones. They don't ask the chefs to cook again; they just pick the best dishes that were already prepared.

How It Works in Practice
The researchers tested this idea using flu data from the 2023-2024 and 2024-2025 seasons in the US.

The Setup: They took the original "static" bag of all predictions from all the experts.
The Filter: Every week, they selected the top 5% to 75% of predictions that matched the real hospitalization numbers best.
The Result: They created a new, "adaptive" forecast using only those selected marbles.

What They Found

Better Accuracy: The "filtered" forecast was usually more accurate than the original "static" bag of all predictions. It was better at guessing how many people would end up in the hospital.
Spotting the Truth: The method was surprisingly good at figuring out which "story" was actually playing out. For example, the experts had different scenarios: "Will the H1N1 virus dominate?" or "Will the H3N2 virus dominate?" or "Will vaccination rates be high or low?" The adaptive filter could look at the real data and say, "Ah, it looks like the H1N1 virus is winning, and vaccination rates are a bit lower than usual," even early in the season.
Short-Term Forecasting: They also used this method to predict the next few weeks (short-term). It beat the "baseline" model (a simple guess that assumes next week will be exactly like this week) but didn't quite beat the full "FluSight" ensemble (which is a massive, constantly updated collaboration of dozens of teams). However, the authors note that their method did this with a fraction of the resources, using only pre-made scenarios rather than constantly re-running complex models.

The Catch
The paper notes that this "filter" isn't magic. It works best when there is enough real data to make a good choice. Right at the very start of the season (when data is scarce) or right at the peak of the flu season (when things are chaotic), the filter sometimes struggles to pick the right marbles, and the original "static" bag might actually be safer.

The Bottom Line
This paper proposes a way to make long-term disease forecasts more useful without needing a supercomputer or a team of chefs to re-cook the meal every week. By simply filtering the best predictions from a pre-made list based on what's happening right now, public health officials can get a clearer, more accurate picture of the epidemic's future. It's a way to turn a static map of all possible roads into a dynamic GPS that guides you down the road you are actually traveling.

Technical Summary: Adaptive Multi-Model Ensembles for Improved Epidemic Projections and Decision Support

Problem Statement
Multi-model ensemble projections have become a standard method in infectious disease modeling to account for structural uncertainties and integrate diverse modeling approaches. However, traditional ensemble methods, particularly those used in collaborative hubs like the U.S. Flu Scenario Modeling Hub (SMH), rely on static aggregation of pre-season projections. These static ensembles do not update as new empirical data becomes available during an epidemic. Consequently, they lack the ability to refine trajectories, update the selection of plausible scenarios, or expand the assessment of scenarios in real-time. This static nature is resource-intensive to maintain and limits the accuracy of long-term projections when underlying epidemic dynamics shift.

Methodology
The authors propose an adaptive ensemble approach that dynamically selects individual model trajectories based on observed data throughout the epidemic projection period. Analogous to a multi-model particle filtering or Sequential Monte Carlo (SMC) update, the method filters a pre-existing pool of trajectories in real-time.

The core methodology involves:

Trajectory Pool: Starting with a large pool of stochastic trajectories generated by multiple modeling teams prior to the season under predefined scenarios (e.g., varying dominant subtypes and vaccination coverage levels).
Iterative Selection: At each epidemiological week, the method extends a calibration window by one week. It evaluates all trajectories against observed data using a loss function (specifically weighted Mean Absolute Percentage Error, wMAPE, for scenario projections; a generalized wMAPE with linear decay for short-term forecasting).
Filtering: The top $X\%$ of best-performing trajectories are selected from each contributing model (for scenario projections) or from the entire pool (for short-term forecasting).
Ensemble Construction: The selected trajectories are aggregated using a Linear Opinion Pool (LOP) approach to generate an updated adaptive ensemble.
Scenario Probability Estimation: By tracking the fraction of selected trajectories originating from each specific scenario, the method derives a proxy for the posterior probability of alternative epidemic scenarios.

The study applies this methodology to two distinct tasks:

Long-term Scenario Projections: Updating projections for the remainder of the season based on data available up to the current week.
Short-term Forecasting: Generating 1-to-4-week-ahead forecasts of influenza hospitalizations.

Key Results
The methodology was evaluated using data from the U.S. Flu Scenario Modeling Hub for the 2023-2024 and 2024-2025 influenza seasons, as well as the CDC FluSight forecasting challenge.

Improved Predictive Accuracy: The adaptive ensemble consistently outperformed the original static ensemble (which aggregates all trajectories without selection) in terms of Weighted Interval Scores (WIS) and Mean Absolute Errors (MAE) across national and subnational (state/territory) levels. Improvements were observed across various selection thresholds (5% to 75%), though higher selectivity (e.g., 5%) yielded the highest median improvements in MAE but with greater variance.
Scenario Identification: The adaptive approach successfully identified the most plausible epidemic scenarios early in the season. For the 2024-2025 season, the method correctly shifted probability mass toward scenarios assuming A/H1N1 dominance early in the season and later identified the co-circulation of A/H1N1 and A/H3N2 with lower-than-usual vaccination coverage as the most likely reality. Similar alignment was found for the 2023-2024 season.
Short-term Forecasting Performance:
- The adaptive ensemble consistently outperformed the FluSight baseline model across all selection percentages and scoring rules.
- Compared to the FluSight ensemble (which aggregates forecasts from up to 50 models optimized for short-term forecasting), the adaptive ensemble showed competitive performance in the 2023-2024 season but did not improve upon the FluSight ensemble in the 2024-2025 season. The authors note this underperformance in 2024-2025 may be due to the atypically severe season, which was underrepresented in the pre-season scenario trajectories.
Robustness: Sensitivity analyses confirmed that results hold across different age groups (65+), different loss functions (RMSE vs. wMAPE), and varying model representation levels.

Significance and Claims
The paper claims that the proposed adaptive ensemble approach offers an efficient, low-resource strategy to enhance the utility of multi-model epidemic projections. Its primary significance lies in three areas:

Real-time Adaptation: It provides a mechanism to update long-term scenario projections dynamically without requiring modeling teams to re-run complex models with new data, thereby reducing computational and coordination burdens.
Scenario Plausibility: It offers a structured way to assign posterior probabilities to different epidemic scenarios (e.g., specific virus strains or vaccination coverage levels), helping decision-makers understand which assumptions align best with reality as the season progresses.
Forecasting Contribution: It demonstrates the potential to reuse long-term scenario trajectories for short-term forecasting, providing a competitive model that can contribute to collaborative forecasting challenges like FluSight with a fraction of the resources required by dedicated short-term forecasting teams.

The authors conclude that while the approach shows promise for improving accuracy and supporting decision-making under uncertainty, it is not a replacement for fully re-running models with updated data. Instead, it serves as a practical, adaptive support tool for modeling teams, public health authorities, and decision-makers, particularly when resources for frequent model recalibration are limited. The paper acknowledges limitations, including reduced performance during complex epidemic phases (onset/peak) and the dependency of scenario probability estimates on the initial design of the scenarios.

More like this