AI Moves From Lab to Front Line of Outbreak Response, WHO Forum Hears

Artificial intelligence is expanding what public health systems can do to detect infectious disease signals, forecast outbreaks, analyse pathogen evolution and support operational decisions, but experts gathering at a recent World Health Organization forum warned that technical performance alone is not enough to make these tools useful in practice. The thirteenth session of the WHO Pandemic and Epidemic Intelligence Innovation Forum brought together researchers from the University of Oxford, Harvard University, Google Research, the Centre for Epidemic Response and Innovation, the University of California Berkeley, and the University of Copenhagen and Imperial College London to examine how AI can strengthen pandemic and epidemic preparedness. The meeting report, published in BMC Proceedings, captures a shared recognition among participants that AI tools must be trusted, auditable, transferable across settings and actionable within the tight timeframes and constraints of real public health decision-making.

The session opened with the question of early warning. Contributors described how AI can contribute across four foundational areas of epidemic intelligence: identifying risk, forecasting trajectories and imputing missing data, characterising pathogen evolution, and strengthening surveillance through adaptive, data-efficient approaches such as active learning. At the individual level, AI is already used to detect early and ongoing symptoms through wearable devices that collect health data, often before people present to care, facilitating early detection and prompting behaviour changes that can prevent worsening illness or onward transmission. The harder challenge, speakers argued, is scaling these insights to support population-level decisions, particularly in data-scarce settings where tools must adapt to evolving outbreaks, integrate diverse data types and produce timely, context-aware outputs.

One presenter described a multi-layered system combining real-time genomic surveillance, ecological modelling and machine learning to support both rapid response and longer-term forecasting. The approach exploits the ability of machine learning and deep learning methods to capture complex, non-linear relationships between environmental and demographic change and pathogen spread, enabling more accurate short- to medium-term predictions at local scales. Reported examples include predicting the importation risk of arboviruses such as dengue and modelling the ecological drivers of Oropouche virus transmission in South America. In the group’s integrated workflow, data sources such as pathogen genomes, epidemic time series and climatic and demographic covariates feed into a diverse set of modelling approaches, with a key strategy being to map classical analytical frameworks and identify where AI can enhance speed, scalability or resolution.

That hybrid philosophy echoed through the panel discussion. Several contributors noted that the most useful methods often do not replace traditional tools but combine them with AI in ways that build on their strengths. By integrating statistical models, mechanistic frameworks and expert-driven inference with machine and deep learning techniques, these hybrid systems can deliver more robust, interpretable and operationally relevant outputs. In this view, AI acts not as a substitute for established epidemiological methods but as a force multiplier, extending the power of existing approaches to meet the complexity and speed of modern epidemic intelligence needs.

Genomic intelligence featured prominently as an area where AI is reshaping practice. Beyond lineage assignment and phylogenetic reconstruction, emerging tools are beginning to predict how mutations may affect viral properties such as transmissibility or immune escape. By using deep learning and structural data to anticipate pathogen evolution, researchers hope to inform public health decisions in advance, although speakers stressed this remains an emerging area. Such approaches could eventually help prioritise variants for laboratory validation, guide vaccine and therapeutic development, and support earlier interventions. In one applied example, a participant described AI methods used across the genomic surveillance pipeline, including anomaly detection applied to molecular clock data to flag sequences with unexpected divergence, a potential early signal of variant emergence. The team plans to apply neural networks to predict pathogen dispersal from phylogeographic data and is exploring AI support for therapeutic target design directly from pathogen genomes.

Access to these capabilities, however, remains deeply unequal. In many settings, public health actors in low- and middle-income countries lack dedicated analytical teams or specialised expertise and depend on external partners for data interpretation. Relying on external expertise can mean losing control over the analytic process, particularly during emergencies when speed, autonomy and trust are essential. In response, one group is developing a platform that supports genomic analysis in low-resource settings using large language models to guide users through workflows such as sequence identification, phylogenetic interpretation and novelty detection via natural-language queries. A central design principle is to enable and facilitate rather than dictate, ensuring users retain control over their data and analysis. The platform runs locally or in secure environments and is being co-developed with partners in LMICs to ensure usability and relevance.

Geospatial AI emerged as another frontier for translating coarse or incomplete data into localised insight. The Population Dynamics Foundation Model, created by Google, integrates aggregated signals including search trends, mobility, weather and air quality data into a graph-based framework that produces high-resolution geospatial embeddings. These embeddings can be used for disease forecasting, data interpolation and super-resolution estimation, aiming to infer fine-grained health indicators, such as ZIP-code-level disease burden, from coarser national or county-level statistics. Recent examples include interpolation of health metrics in Malawi and super-resolution estimates of vaccination coverage in the United States. Other researchers have applied similar methods, combining coarse administrative data with high-dimensional embeddings of search, mobility and environmental variables to identify localised gaps in coverage and inform more targeted vaccination outreach.

AI is also being deployed to support decisions directly, from frontline operations to long-term planning. In partnership with the Virginia Department of Health during the COVID-19 pandemic, one team created interactive dashboards integrating mobility data, point-of-interest metadata and vaccine uptake to guide reopening policies and optimise vaccine site placement, giving decision-makers granular, real-time insight into community behaviour and access barriers. The same researchers developed models estimating vaccine uptake at the ZIP code level by combining coarse administrative data with embeddings from search trends, and showed that these search-based estimates were not only more spatially granular but preceded official CDC vaccination reports by one to two weeks, potentially enabling more targeted and pre-emptive outreach where coverage data are missing or delayed. The team is now using generative AI, including diffusion models and LLM-based agents, to simulate human behaviours such as mobility trajectories, social networks and public opinions, creating synthetic populations that allow interventions like reopening timelines or school closures to be tested in silico before implementation.

Agent-based modelling offered a striking illustration of how simulation can reshape operational thinking. Built with reinforcement learning and optimisation algorithms, these models replicate interactions between individuals, communities and health systems under real-world constraints such as varying supply levels, staff shortages and geographic barriers. During the COVID-19 pandemic, such models were used to compare testing strategies and demonstrated that frequent use of rapid antigen tests could outperform less frequent PCR testing in reducing transmission, despite lower test sensitivity. The key insight was that the speed and frequency of testing could have a greater population-level impact than diagnostic precision alone. The same simulation approach has been applied to maternal health and HIV prevention in India, testing how service delivery strategies perform under constraints including limited staff, transportation barriers and social stigma.

Across all sessions, participants returned to the barriers that keep high-performing models out of practice. Attendees raised concerns about trust in black-box models, noting that approaches based on known epidemiological methods are often easier to explain to decision-makers; speakers suggested that combining classical models with AI components may help teams feel confident without losing control over decisions. Contributors emphasised transparency through clear documentation, meaningful uncertainty estimates and interpretable interfaces, and warned that models developed in high-income, data-rich contexts often fail to generalise to settings with different population structures, reporting systems or data realities. Proposed remedies include robust benchmarking, using real-world interventions as quasi-experiments and building shared libraries of validation tasks. Even well-benchmarked models can fail if they depend on infrastructure, bandwidth or expertise unavailable in lower-resource or emergency settings, making simpler, frugal models that run offline and update incrementally more impactful in practice. Looking ahead, speakers pointed to generative AI and digital twins for policy testing, LLMs as orchestration engines linking datasets and decision-support logic, structural modelling from genomic data, graph neural networks and AI-accelerated calibration of agent-based models, alongside a proposed cross-organisational survey to identify which high-value AI use cases deliver the greatest public health impact. Realising that vision, participants concluded, will require sustained interdisciplinary collaboration, human oversight, robust validation and equitable design at every stage of development and deployment.

Subject of Research: Applications of artificial intelligence in infectious disease modelling and public health decision-making discussed at a WHO innovation forum

Article Title: Leveraging AI for infectious disease modelling and public health decision making

Article References: Kraemer, M. U. G., Tornimbene, B., Bhatt, S., Chang, S., Prasad, G., Tambe, M., Tegally, H., & Morgan, O. (2026). Leveraging AI for infectious disease modelling and public health decision making. BMC Proceedings, 19(S4), Article 41. https://doi.org/10.1186/s12919-026-00398-w

Image Credits: AI Generated

DOI: 10.1186/s12919-026-00398-w

Keywords: artificial intelligence, infectious disease modelling, WHO, pandemic preparedness, genomic surveillance, outbreak forecasting, machine learning, public health, agent-based models, geospatial modelling, large language models, epidemic intelligence