TLDR: Mantis is a new AI model for disease forecasting that learns entirely from simulated outbreaks, not real-world data. It can accurately predict disease spread, hospitalizations, and deaths up to 8 weeks in advance, outperforming many expert-tuned models. Mantis is unique because it works “out-of-the-box” for various diseases and regions, even with limited historical data, and provides explanations for its predictions by linking them to underlying disease mechanisms. This makes it a valuable tool for public health planning, especially in new or under-resourced outbreak situations.
In the critical field of public health, accurately predicting the spread of infectious diseases is paramount for effective response and resource allocation. However, traditional forecasting models often fall short, especially during novel outbreaks or in areas with limited resources. These models typically demand vast amounts of disease-specific historical data, require extensive manual tuning by experts, and struggle to generalize across different diseases or locations. Furthermore, their forecasting horizons are usually quite short, often only predicting 1-4 weeks ahead, which is insufficient for proactive public health planning that requires 6-8 weeks of lead time for major decisions like hospital capacity or vaccination campaigns.
Adding to these challenges, many high-performing models, particularly those based on complex neural networks, operate as ‘black boxes.’ Their predictions are often opaque, making it difficult for public health decision-makers to understand the underlying reasons for a forecast, which can hinder trust and timely action. There’s a clear need for models that are not only accurate but also interpretable in epidemiological terms.
Introducing Mantis: A New Paradigm in Disease Forecasting
A groundbreaking solution has emerged with the introduction of Mantis, a foundation model designed to overcome these limitations. Mantis is built on the innovative Simulation-Grounded Neural Network (SGNN) paradigm, meaning it learns entirely from large-scale, highly realistic mechanistic simulations rather than real-world historical data. This unique training approach allows Mantis to forecast disease dynamics out-of-the-box, even in settings with scarce or delayed real-world data.
Mantis was trained on an immense dataset comprising over 400 million simulated days of outbreak dynamics. These simulations cover a wide array of pathogens, transmission modes (human-to-human, vector-borne, environmental), intervention strategies, immunity profiles, and even surveillance artifacts like underreporting and delays. By learning from these diverse synthetic outbreaks, Mantis internalizes fundamental contagion dynamics, enabling it to generalize to real-world scenarios it has never encountered before.
Unprecedented Performance and Generalization
The performance of Mantis has been rigorously evaluated against 39 expert-tuned models across six real-world diseases: seasonal influenza, early COVID-19, dengue, hepatitis B, smallpox, and scarlet fever. These evaluations spanned various transmission modes, outbreak phases, surveillance systems, and historical eras. Remarkably, Mantis consistently outperformed all baselines in every comparison, including all models in the CDC’s COVID-19 Forecast Hub. For instance, at 8-week horizons, Mantis achieved lower relative error than the CDC’s ensemble of COVID-19 forecasts at 4 weeks, despite not using any COVID-specific data or covariates during its training.
Mantis demonstrates exceptional generalization capabilities. It can forecast diseases with transmission mechanisms not explicitly represented in its training simulations, such as chronic blood-borne infections like hepatitis B. It also adapts to new input data modalities; for example, a Mantis model trained without syndromic surveillance data (like influenza-like illness percentages) can still accurately forecast these indicators. This level of adaptability underscores Mantis’s capacity as a true foundation model, capable of immediate deployment in diverse contexts without the need for retraining or expert tuning.
Mechanistic Interpretability for Actionable Insights
Beyond accuracy, Mantis addresses the critical need for interpretability through a feature called “back-to-simulation attribution.” This method provides mechanistic explanations for its forecasts by identifying the underlying dynamics the model believes are driving an outbreak. When given a real-world time series, Mantis searches its vast library of simulated scenarios to find those that are most similar, not just in appearance, but in their underlying mechanistic processes (e.g., similar transmission rates or immunity profiles). The known parameters of these retrieved simulations are then aggregated to explain the forecast in epidemiological terms. This transparency allows public health decision-makers to understand and trust the model’s outputs, enabling more informed and proactive responses.
Also Read:
- RadarQA: Enhancing Weather Forecast Evaluation with Multi-modal AI
- AI Framework Enhances Emergency Response and Resource Equity in African Communities
A Foundation for Future Public Health Systems
Mantis represents a significant leap forward for infectious disease forecasting. Its ability to deliver accurate, long-range, and mechanistically interpretable forecasts without relying on real-world training data positions it as a practical and general-purpose tool for public health, particularly in low-resource or emerging outbreak settings where data is often limited or delayed. By learning from mechanistic priors, Mantis supports plug-and-play forecasting, eliminating the longstanding trade-off between generalization and interpretability.
The development of Mantis also signals a broader shift towards simulation-supervised machine learning in scientific domains, where mechanistic models encode prior knowledge, and neural networks learn to generalize from them. As an open-source tool, Mantis is readily available for users to generate forecasts from a single historical time series without requiring machine learning expertise. For more technical details, you can refer to the original research paper: Mantis: A Simulation-Grounded Foundation Model for Disease Forecasting.
