AI summary · arXiv: 2611.16935v1 · cs.AI

Universal Infectious Disease Forecasting with Mantis 2

Carson Dudley, Reiden Magdaleno, Yutong Bi, Anthony Valencia, Xiaofeng Liu, Marisa Eisenberg
Original paper Submitted to arXiv on 12 Nov. 2026 License: CC BY-NC-ND 4.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

Mantis 2 combines simulated and observed outbreak data to produce multi-variable, 16-week forecasts, with strong retrospective accuracy and simulation benefits greatest when local history is short.

Question

Can a general infectious-disease forecasting model combine mechanistic simulations with observed data to forecast across diseases and locations? The study also asks when simulation-rich training helps, particularly as the amount of local surveillance history changes.

Method

The authors trained Mantis 2 using over 1,000 real-world scenarios and 20,000 simulated scenarios, with a curriculum that decreased simulated sampling from 70% to zero. They evaluated it on 30 disease-location tasks against six comparison models, and tested 11 fixed simulation-to-observation mixtures with 8, 16, 26, 52, or 104 weeks of history.

Findings

  • Result Mantis 2 had the lowest aggregate weighted interval score among the evaluated models across short, medium, and long forecast horizons.
    See the passage in the paper
    Mantis 2 achieved the lowest aggregate log-WIS across all models for all three horizon ranges (Figure 2A-C).
  • Result Across 3,725 observed peaks, Mantis 2 missed roughly 30% fewer peaks than Chronos-2, and its successful predictions gave more than six weeks of advance warning on average.
    See the passage in the paper
    Relative to Chronos-2, Mantis 2 reduced the proportion of missed peaks by roughly 30%. Among peaks each model successfully anticipated, the mean lead time of the first correct prediction was 6.43 weeks for Mantis 2 and 6.32 weeks for Chronos-2.
  • Result Simulation-rich training improved median WIS when only eight weeks of history were available, but its advantage diminished with 104 weeks of history.
    See the passage in the paper
    With only eight weeks of observations, the 80% and 90% synthetic models improved median WIS over real-only training by 7.9% and 9.3%, respectively. With 104 weeks, their median improvements were instead $-0.5\%$ and $-1.2\%$.
  • Result Higher simulated-data fractions were associated with fewer exceptionally large forecast errors than real-only training.
    See the passage in the paper
    Across our evaluation datasets datasets, the mean fraction of forecasts that had error greater than five times the dataset’s mean real-only WIS was 0.364% for real-only training, compared with 0.261% and 0.207% for 80% and 90% synthetic training.
  • Authors’ interpretation The authors interpret simulation as useful background knowledge when local surveillance evidence is limited, while longer observed histories help the model adapt to a specific disease.
    See the passage in the paper
    Broad mechanistic training was particularly valuable when context with a specific disease was very limited, whereas longer histories often favored observational training or reduced sensitivity to the mixture.

Limitations

  • The evaluation was retrospective, so improved forecast scores do not establish that the model improves public-health decisions.
    See the passage in the paper
    Our evaluation of Mantis 2 is retrospective, and improved forecast scores do not by themselves demonstrate improved public health decisions.
  • The experiments that vary history length truncate existing series and do not directly test the conditions of a newly emerging outbreak.
    See the passage in the paper
    These experiments isolate the effect of having fewer observations, but a short history drawn from an established series does not necessarily represent the dynamics or uncertainty encountered early in a new outbreak.
  • The study's training-data mixture and coverage may not be optimal or representative for every disease, location, or surveillance system.
    See the passage in the paper
    Both the observational training data and the mechanistic simulations represent selected diseases, locations, and processes; their coverage cannot guarantee performance for every emerging pathogen or surveillance system.

Why it matters

The authors argue that one pretrained model can make multivariate, longer-horizon infectious-disease forecasts accessible without building a separate system for every disease and location. Their results also identify local surveillance history as a determinant of the value of mechanistic simulation, suggesting a role for simulation-rich training where local evidence is sparse.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: Infectious-disease forecasting; probabilistic forecasting; epidemic simulation; surveillance data; outbreak peaks

Read further

  • Comprehensive Summary
  • Layman's Summary
  • Blog article
  • Original abstract
The paper introduces Mantis 2, a general-purpose model for probabilistic infectious-disease forecasting, and examines when simulated outbreak data improve predictions compared with observational data. Its central question is whether mechanistic simulations are most useful when local surveillance history is limited, and how their contribution changes as more observations become available. Mantis 2 builds on the Chronos-2 time-series model. It was fine-tuned on more than 1,000 real-world infectious-disease scenarios and about 20,000 simulated scenarios generated with the Reservoir epidemic simulator. These simulations cover varied transmission mechanisms, interventions, and reporting processes. The training curriculum begins with 70% simulated examples and gradually reduces their share to zero, yielding an expected average of 35% simulated examples. The model accepts up to 512 weeks of history, jointly forecasts up to six surveillance variables, and generates probabilistic forecasts up to 16 weeks ahead. The authors evaluated Mantis 2 retrospectively on 30 disease-location tasks covering 23 diseases, 11 countries, and multiple transmission categories. Comparisons included Chronos-2, an LSTM, statistical models, and seasonal-naive and persistence baselines. Forecast accuracy was measured primarily using weighted interval score (WIS) on log-transformed outcomes. In separate experiments, the authors trained models with fixed simulated-data fractions ranging from zero to 100% and varied the forecast history from 8 to 104 weeks. Mantis 2 achieved the best aggregate WIS across short, medium, and long horizons: 0.209 over weeks 1–4, 0.284 over weeks 1–8, and 0.370 over weeks 1–16. The closest comparison, Chronos-2, scored 0.231, 0.311, and 0.398, respectively. Dataset-level results significantly favored Mantis 2 (Wilcoxon test, P < 0.001). It ranked first on 22 of the 29 datasets with results available for all six models and had lower reported WIS than Chronos-2 on 26 of 30 datasets. Across 3,725 observed outbreak peaks, Mantis 2 correctly anticipated more than 80%, compared with 73% for Chronos-2—about 30% fewer missed peaks. Among peaks each model successfully predicted, the mean lead time to the first correct prediction was 6.43 weeks for Mantis 2 and 6.32 weeks for Chronos-2. In a restricted comparison with the original Mantis, Mantis 2 reduced pooled eight-week WIS by 31.42% across 27 datasets and improved on every dataset. Fixed-mixture experiments showed that simulated training was particularly helpful when local history was limited. With eight weeks of context, models trained with 80% and 90% simulated data improved median WIS over real-only training by 7.9% and 9.3%, respectively. With 104 weeks of history, those mixtures instead changed median WIS by −0.5% and −1.2%. Simulation-only training improved median WIS by 8.2% with eight weeks of context but worsened it by 11.1% with 104 weeks. Simulation-rich models also produced exceptionally large errors less often: forecasts exceeding five times the real-only reference WIS accounted for 0.207% of forecasts with 90% simulation, compared with 0.364% for real-only training. The authors conclude that simulations can provide a useful prior when local evidence is sparse, while longer histories often allow observational data to better represent the specific disease and surveillance setting. Neither data source alone performed best across all tasks.

Was this synthesis useful?

Score: 0

Why do we need votes?

Votes are used to determine whether we need to re-run our summarizing tools. If the count reaches -10, our tools can be restarted.

Figures and plots

Ask about a figure by number or use its question button to request visual analysis. Up to two figures and four image panels per question. Values read from plots are approximate; check the original data.

AI assistant instructions?

Tools for this paper

Explain it differently 1 free rewrite left today
Save to a collection Group papers in collections to compare them, chat with them and export them. Continue with Google Create a free account Log in

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero Zotero guide
Citation alerts Citation alerts guide
Who works on it?

Among the 23 papers around this one, the name that comes back most: C. Dudley (2 papers)

5 more people sign several of them. The full map adds their institutions and approaches, from the papers of the last years.

Continue with Google Create a free account: your first map is free
Weekly letter Presented in "Artificial intelligence this week" of 16 November 2026. Artificial intelligence this week: What impressive scores leave out · Receive one by email for free

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

  1. Very close Builds on this paper Cited by this paper Used in depth by this paper
    Mantis: A foundation model for mechanistic disease forecasting.
    C. Dudley, Reiden Magdaleno, Christopher Harding et al. · 2025 · Proceedings of the National Academy of Sciences of the United States of America · cited 8 times · Not summarized yet

    Mantis 2 extends the simulation-only Mantis foundation model by combining simulated and observed training data to forecast across diseases and locations.

    Abstract

    Infectious disease forecasting in novel outbreaks or low-resource settings is hampered by the need for large disease and covariate datasets, bespoke training, and expert tuning, all of which can hinder rapid generation of forecasts for new settings. To address these challenges, we developed Mantis, a foundation model trained entirely on mechanistic simulations, which enables out-of-the-box forecasting across diseases, regions, and outcomes, even in settings with limited historical data. We evaluated Mantis against 78 forecasting models across 16 diseases with diverse transmission modes, assessing both point forecast accuracy [mean absolute error (MAE)] and probabilistic performance (weighted…

  2. Close Another approach to the problem Cited by this paper
    Improving outbreak forecasts through model augmentation
    Graham C Gibson, S. Fox, E. Javan et al. · 2025 · Proceedings of the National Academy of Sciences of the United States of America · cited 8 times · Not summarized yet

    Epimodulation incorporates epidemiological principles into forecasting models to improve peak forecasts for COVID-19 and influenza hospital admissions.

    Abstract

    Significance Reliable outbreak forecasting is essential for public health decision-making, yet traditional methods often falter during critical epidemiological time periods such as the peak of an epidemic. This work introduces epimodulation, a forecasting approach that integrates epidemiological principles into a wide-range of forecasting models to enhance their performance. The approach increases model accuracy by up to 55% during epidemic peaks for COVID-19 and influenza hospital admission forecasts, without reducing accuracy at other times. Due to its flexible and straightforward design, epimodulation is a powerful tool for enhancing forecast performance, boosting healthcare preparedness …

  3. Close Broader context Cited by this paper
    A collaborative multiyear, multimodel assessment of seasonal influenza forecasting in the United States
    N. Reich, Logan C. Brooks, S. Fox et al. · 2019 · Proceedings of the National Academy of Sciences of the United States of America · cited 340 times · Not summarized yet

    The assessment compares real-time seasonal influenza forecasts from five research groups in the United States and highlights data accuracy as a forecasting obstacle.

    Abstract

    Significance Accurate prediction of the size and timing of infectious disease outbreaks could help public health officials in planning an appropriate response. This paper compares approaches developed by five different research groups to forecast seasonal influenza outbreaks in real time in the United States. Many of the models show more accurate forecasts than a historical baseline. A major impediment to predictive ability was the real-time accuracy of available data. The field of infectious disease forecasting is in its infancy and we expect that innovation will spur improvements in forecasting in the coming years. Influenza infects an estimated 9–35 million individuals each year in the Un…

  4. Close Broader context Cited by this paper
    An open challenge to advance probabilistic forecasting for dengue epidemics
    M. Johansson, K. Apfeldorf, Scott Dobson et al. · 2019 · Proceedings of the National Academy of Sciences of the United States of America · cited 218 times · Not summarized yet

    The challenge compares probabilistic dengue forecasts from 16 teams and identifies early-season skill and ensemble methods as areas for improvement.

    Abstract

    Significance Forecasts routinely provide critical information for dangerous weather events but not yet for epidemics. Researchers develop computational models that can be used for infectious disease forecasting, but forecasts have not been broadly compared or tested. We collaboratively compared forecasts from 16 teams for 8 y of dengue epidemics in Peru and Puerto Rico. The comparison highlighted components that forecasts did well (e.g., situational awareness late in the season) and those that need more work (e.g., early season forecasts). It also identified key facets to improve forecasts, including using multiple model ensemble approaches to improve overall forecast skill. Future infectiou…

  5. Close Broader context Cited by this paper
    Rethinking respiratory disease forecasting: temporal heterogeneity between surveillance predictors and outcomes drives forecast instability
    H. Topazian, T. Sheets, R. Gruninger et al. · 2026 · medRxiv · preprint, not peer reviewed · cited 1 time · Not summarized yet

    The study tests whether surveillance predictors consistently forecast SARS-CoV-2 and influenza hospitalizations in Utah over time.

    Abstract

    Since the COVID-19 pandemic, forecasting hubs and non-traditional respiratory disease surveillance streams have become increasingly common. However, many forecasting approaches assume that relationships between surveillance predictors and disease outcomes remain stable over time and that incorporating additional historical data will improve forecast performance. To evaluate these assumptions in a real-world setting, we developed and evaluated forecasts of SARS-CoV-2 and influenza hospitalizations in Utah using syndromic surveillance, test positivity, and wastewater data. Rather than identifying a single, best-performing model, we examined whether relationships between surveillance predictors…

  6. Close Broader context
    Infectious Disease Forecasting via Physics-Informed Machine Learning
    Joseph Hart, Hudson Smith, L. Rennert et al. · 2026 · bioRxiv · preprint, not peer reviewed · Not summarized yet

    The paper discusses physics-informed machine learning for disease transmission forecasting, a neighboring approach to combining mechanistic knowledge with prediction.

    Abstract

    Infectious disease transmission evolves as a dynamic process shaped by biological mechanisms, population behavior, and intervention policies, yet public health responses are often driven by lagging indicators. Accurate short- and long-term disease forecasting is essential for the timely deployment of intervention strategies, healthcare capacity planning, and uncertainty-aware, risk-informed decision-making. To address this challenge, three broad classes of forecasting models have traditionally been used: statistical, machine learning, and mechanistic approaches. However, each of these modeling paradigms faces fundamental limitations. In particular, traditional statistical models often lack t…

  7. Close Another approach to the problem
    Foundation models for time series forecasting and policy evaluation in infectious disease epidemics.
    Suprabhath Kalahasti, Benjamin Faucher, Bo-Xuan Wang et al. · 2026 · Epidemics · cited 9 times · Not summarized yet

    The study evaluates transformer-based foundation models across pathogens and locations for epidemic forecasting and policy evaluation.

    Abstract

    Epidemic forecasting and public health policy evaluation rely on mathematical models, but traditional approaches struggle in data-limited settings. We evaluated whether transformer-based, foundation Artificial Intelligence models can serve as a new epidemic modeling framework. We tested five models across pathogens and locations, including influenza, respiratory syncytial virus (RSV), chickenpox, dengue. Foundation models demonstrated strong accuracy in short-term forecasts and predicted multiple epidemic waves. They outperformed standard implementations of established models on limited and irregular data. We showed foundation models can generate scenarios for policy evaluation, estimating t…

  8. Close Another approach to the problem
    Foundation time series models for forecasting and policy evaluation in infectious disease epidemics
    Suprabhath Kalahasti, Benjamin Faucher, Bo-Xuan Wang et al. · 2025 · medRxiv · preprint, not peer reviewed · cited 10 times · Not summarized yet

    The study evaluates five time-series foundation models across diseases and locations for infectious-disease forecasting and policy evaluation.

    Abstract

    Epidemic forecasting and policy evaluation rely on mathematical models to predict infectious disease trends and assess the impact of public health policies. Traditional models typically require extensive epidemiological data and may struggle in data-limited settings. Transformer-based, foundation AI models have demonstrated strong predictive capabilities in various time series applications. We investigated whether they can be the basis of a new epidemic modeling framework. We evaluated five foundation models - TabPFN-TS, TimeGPT, TimesFM, Lag-Llama, and Chronos - across diverse pathogens, diseases and locations, including influenza-like illness, RSV, chickenpox, dengue, COVID-19 and neonatal…

Prior work it builds on

  1. Close Broader context
    Applying infectious disease forecasting to public health: a path forward using influenza forecasting examples
    Chelsea S. Lutz, Mimi P. Huynh, Monica Schroeder et al. · 2019 · BMC Public Health · cited 158 times · Not summarized yet

    The paper discusses how influenza forecasts can support public-health decisions and describes CDC collaborative forecasting challenges.

    Abstract

    Infectious disease forecasting aims to predict characteristics of both seasonal epidemics and future pandemics. Accurate and timely infectious disease forecasts could aid public health responses by informing key preparation and mitigation efforts. For forecasts to be fully integrated into public health decision-making, federal, state, and local officials must understand how forecasts were made, how to interpret forecasts, and how well the forecasts have performed in the past. Since the 2013–14 influenza season, the Influenza Division at the Centers for Disease Control and Prevention (CDC) has hosted collaborative challenges to forecast the timing, intensity, and short-term trajectory of infl…

  2. Close Broader context
    Real-time Epidemic Forecasting: Challenges and Opportunities
    A. Desai, M. Kraemer, S. Bhatia et al. · 2019 · Health Security · cited 141 times · Not summarized yet

    The review discusses real-time epidemic forecasting opportunities and constraints, including surveillance data needs and reporting limitations.

    Abstract

    Infectious disease outbreaks play an important role in global morbidity and mortality. Real-time epidemic forecasting provides an opportunity to predict geographic disease spread as well as case counts to better inform public health interventions when outbreaks occur. Challenges and recent advances in predictive modeling are discussed here. We identified data needs in the areas of epidemic surveillance, mobility, host and environmental susceptibility, pathogen transmissibility, population density, and healthcare capacity. Constraints in standardized case definitions and timely data sharing can limit the precision of predictive models. Resource-limited settings present particular challenges f…

  3. Related Shared data, tool or method
    Reservoir: A Large-Scale Simulated Dataset for Training and Evaluating Epidemiological Models
    C. Dudley, Reiden Magdaleno, Marisa C. Eisenberg · 2026 · cited 1 time · Not summarized yet

    Reservoir provides an open simulator and large simulated epidemic dataset that can support training and evaluating epidemiological models.

    Abstract

    Large-scale, standardized datasets have driven many advances in AI-based scientific modeling, from protein structure prediction to natural language processing. Infectious disease epidemiology is increasingly adopting AI methods for forecasting, surveillance, and outbreak analytics, but the time-series data available to train them remains orders of magnitude smaller than the corpora behind the advances seen in other fields. Because the scope of real-world epidemiological data cannot practically reach the scale needed to train truly large-scale AI methods, simulated data provides a possible alternative. Here we introduce Reservoir, a large open simulator and dataset of realistic epidemic simul…

  4. Related Broader context
    Appropriate Models for the Management of Infectious Diseases
    H. Wearing, P. Rohani, M. Keeling · 2005 · PLoS Medicine · cited 613 times · Not summarized yet

    The paper analyzes assumptions in mathematical infectious-disease models and their effects on estimates of transmission and control.

    Abstract

    Background Mathematical models have become invaluable management tools for epidemiologists, both shedding light on the mechanisms underlying observed dynamics as well as making quantitative predictions on the effectiveness of different control measures. Here, we explain how substantial biases are introduced by two important, yet largely ignored, assumptions at the core of the vast majority of such models. Methods and Findings First, we use analytical methods to show that (i) ignoring the latent period or (ii) making the common assumption of exponentially distributed latent and infectious periods (when including the latent period) always results in underestimating the basic reproductive ratio…

  5. Related Broader context
    A modelling approach for correcting reporting delays in disease surveillance data
    L. Bastos, T. Economou, M. Gomes et al. · 2017 · Statistics in Medicine · cited 121 times · Not summarized yet

    The paper proposes a Bayesian model to correct reporting delays in dengue surveillance data and quantify uncertainty.

    Abstract

    One difficulty for real‐time tracking of epidemics is related to reporting delay. The reporting delay may be due to laboratory confirmation, logistical problems, infrastructure difficulties, and so on. The ability to correct the available information as quickly as possible is crucial, in terms of decision making such as issuing warnings to the public and local authorities. A Bayesian hierarchical modelling approach is proposed as a flexible way of correcting the reporting delays and to quantify the associated uncertainty. Implementation of the model is fast due to the use of the integrated nested Laplace approximation. The approach is illustrated on dengue fever incidence data in Rio de Jane…

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 16 November 2026.

The illustrated guide to the papers around a paper

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.