AI Foundation Models Show Promise for Forecasting Infectious Disease Outbreaks Across Countries

Public health agencies around the world depend on accurate forecasts of infectious disease activity to trigger outbreak alerts, schedule vaccination campaigns and allocate scarce resources such as hospital beds, diagnostic tests and personnel. Yet building a reliable forecasting model for every notifiable pathogen has long been impractical. Modern deep-learning forecasters can be highly accurate, but they typically demand disease-specific training data, careful hyperparameter tuning and substantial computational resources, conditions that are rarely met for rare pathogens, newly notified diseases or surveillance programs operating with limited infrastructure. A new study published in BMC Public Health suggests that a newer class of artificial intelligence systems, known as time series foundation models, may offer a way around this bottleneck by delivering useful forecasts without any disease-specific training at all.

Time series foundation models, abbreviated TSFMs, are a product of the same paradigm that produced large language models. Instead of being trained on a single dataset for a single task, they are pre-trained on large, heterogeneous collections of time series drawn from many domains, allowing them to learn general patterns of temporal structure such as seasonality, trends and autocorrelation. Once trained, they can be deployed in a zero-shot manner, meaning they generate forecasts for a brand-new series without ever having seen that series during training. For surveillance epidemiology, this is an attractive proposition: a single pre-trained model could, in principle, forecast influenza in one country, dengue in another and an obscure enterovirus in a third, all without bespoke modeling work.

Whether that promise holds up in practice has been unclear, because no systematic multi-disease, multi-country evaluation had been carried out. To fill this gap, Peng Cai, Guoyu Huang and Ximing Xu of the Children’s Hospital of Chongqing Medical University assembled an unusually broad benchmark. They collected 289 disease–country time series covering more than 70 notifiable infectious diseases from six national surveillance systems: China, Singapore, Japan, South Korea, Ireland and the United States. The dataset comprised 106,234 weekly or monthly observations, spanning a wide range of transmission dynamics, reporting cadences and epidemiological contexts, from respiratory viruses with strong winter seasonality to infections with irregular or endemic patterns.

The researchers evaluated four time series foundation models: TimesFM, Moirai, Chronos and TimeMoE, whose parameter counts range from roughly 14 million to 314 million. These were tested strictly in zero-shot mode, with no fine-tuning on the surveillance data. They were compared against five established neural network baselines that were trained for the task: DLinear, a simple linear architecture; LSTM, a recurrent network long used in sequence modeling; PatchTST, a transformer that processes patches of a series; TimeMixer, a multi-scale mixing architecture; and TimesNet, which decomposes series into periodic components. Forecasts were generated at three horizons: one-step-ahead, medium-term and long-term, and scored using four standard metrics: mean absolute error, root mean square error, mean absolute scaled error and the coefficient of determination, known as R². For the Moirai variants, which produce probabilistic outputs, the team additionally assessed prediction intervals using the Winkler score, a measure that rewards both accuracy and well-calibrated uncertainty.

The headline result is striking. Across 36 evaluation settings, defined as six countries multiplied by three forecasting horizons multiplied by two error metrics, a foundation model variant achieved the lowest observed error in 30 of them, or 83.3 percent. The pattern was most consistent at the one-step-ahead horizon, where a TSFM variant ranked first in every country on both MAE and RMSE. Foundation model variants also ranked highest on MASE and R², indicating that their advantage was not an artifact of a particular error definition. The authors are careful to note that this is a descriptive ranking summary rather than formal evidence of statistical superiority, but the breadth and consistency of the pattern across such diverse surveillance systems is difficult to dismiss.

Equally interesting are the results that complicate a simple story. The study found that forecast errors did not decrease monotonically with model size for Moirai, Chronos or TimeMoE. In other words, the largest variant was not reliably the best; in several settings, smaller versions of the same architecture outperformed their bigger siblings. Similarly, the Moirai variant with the lowest Winkler score, indicating the best-calibrated prediction intervals, varied by country and by forecasting horizon. This has practical implications, because it suggests that surveillance agencies cannot simply assume that scaling up parameters will improve forecasts, and that modest, lightweight models may be the most sensible default for routine deployment.

That last point matters for equity in public health technology. Foundation models with tens of millions of parameters can run on modest hardware, unlike the massive models that dominate headlines in other areas of artificial intelligence. The study’s authors highlight these small variants as a particularly promising starting point for surveillance forecasting pipelines, since they combine zero-shot usability with computational demands that low-resource settings can realistically meet. For a health ministry that cannot afford a data science team to build and maintain bespoke forecasters for dozens of pathogens, a pre-trained model that produces competitive forecasts out of the box could meaningfully expand analytical capacity.

The researchers also explored a more ambitious use of the technology. In an exploratory multivariate analysis using Moirai, they jointly forecasted co-circulating respiratory pathogens rather than treating each disease in isolation. The results indicated that this joint approach may yield additional performance gains, presumably because related pathogens share seasonal drivers, compete for susceptible hosts and respond to common behavioral patterns such as school terms and holiday travel. If confirmed in future work, this would align foundation model forecasting with a growing recognition in epidemiology that multi-pathogen modeling better reflects the reality of respiratory disease seasons than single-disease silos.

The authors are candid about the limitations of their evaluation, and these caveats deserve emphasis. Each time series was assessed over a single held-out year, which means the results do not establish consistent performance across multiple years, shifting epidemic cycles or changing surveillance conditions such as altered case definitions or reporting disruptions. The comparison also pits zero-shot foundation models against neural baselines trained with fixed default configurations rather than exhaustively tuned ones, so the benchmark measures a realistic deployment scenario rather than the theoretical ceiling of each method. Within those evaluated test periods, however, the conclusion stands: zero-shot TSFMs were competitive with task-specific neural forecasting models, a remarkable outcome for systems that never saw the target data during training.

The study arrives at a moment when health agencies are actively seeking to modernize surveillance in the wake of the COVID-19 pandemic, which exposed both the value and the fragility of epidemic forecasting. By benchmarking foundation models across more than 70 diseases and six countries, this work provides one of the most comprehensive empirical pictures to date of what pre-trained time series AI can and cannot do for outbreak preparedness. The message for the field is cautiously optimistic: foundation models are not a finished solution, and their performance must still be validated over longer horizons and more varied conditions, but as a low-cost, low-effort starting point for forecasting pipelines, they appear to earn a place at the table. The work was supported by the Natural Science Foundation of Chongqing, and the full open-access article is available for public health researchers who wish to examine the benchmark in detail.

Subject of Research: Zero-shot evaluation of time series foundation models for multi-country infectious disease surveillance forecasting

Article Title: Benchmarking time series foundation models for multi-country infectious disease forecasting: a retrospective evaluation across over 70 diseases

Article References: Benchmarking time series foundation models for multi-country infectious disease forecasting: a retrospective evaluation across over 70 diseases. (n.d.). https://doi.org/10.1186/s12889-026-29918-1

Image Credits: AI Generated

DOI: 10.1186/s12889-026-29918-1

Keywords: infectious disease surveillance, outbreak forecasting, time series foundation models, zero-shot learning, machine learning, epidemiology, public health, time series analysis, deep learning, benchmarking, BMC Public Health, predictive modeling