Artificial intelligence for modelling infectious disease epidemics

Abstract

Infectious disease threats to individual and public health are numerous, varied and frequently unexpected. Artificial intelligence (AI) and related technologies, which are already supporting human decision making in economics, medicine and social science, have the potential to transform the scope and power of infectious disease epidemiology. Here we consider the application to infectious disease modelling of AI systems that combine machine learning, computational statistics, information retrieval and data science. We first outline how recent advances in AI can accelerate breakthroughs in answering key epidemiological questions and we discuss specific AI methods that can be applied to routinely collected infectious disease surveillance data. Second, we elaborate on the social context of AI for infectious disease epidemiology, including issues such as explainability, safety, accountability and ethics. Finally, we summarize some limitations of AI applications in this field and provide recommendations for how infectious disease epidemiology can harness most effectively current and future developments in AI.

Figure 1:. AI approaches to tackling key epidemiological questions.

(A) Time-series analysis using foundation models (FMs). Pre-training large-scale FMs with transformer architecture using observations (empirical or simulated) from time-series epidemic data allows zero-shot applications (Box 1) for forward prediction, data imputation, and anomaly detection. Output from these models can be used to evaluate the impact of public interventions and guide capacity planning for future outbreaks. (B) Modelling infectious disease spread using Graph Neural Networks (GNNs). Pathogen dissemination can be represented via annotated graphs, where nodes correspond to locations or individuals and edges represent potential transmission pathways (e.g., human or vector interactions); each node is associated with a set of features (e.g., case incidence, population size) that are either indicators or drivers of spread. GNNs are capable of learning complex patterns from such data, enabling node classification (predicting disease prevalence), community detection (identifying infection clusters), and link prediction (revealing cryptic transmission pathways). These insights can provide a detailed understanding of the underlying mechanism of disease spread and inform resource allocation and targeted interventions. (C) Predicting immune escape mutations using biologically-informed deep learning models. By leveraging recent advances in protein structure prediction (e.g. AlphaFold, ESMFold), these models could enable the early detection of pathogens or variants likely to develop mutations conferring resistance to existing vaccines or therapeutics through antibody-binding disruption. Such predictions can inform the design of next-generation vaccines and guide the prioritisation of containment strategies targeted at emerging lineages.

Figure 2:. Classification of data types to investigate infectious diseases.

Data used to inform epidemiological modelling are placed in terms of their accessibility to the research community and their population level coverage. Data are classified based on their data types, including epidemiological (red), environmental (brown), social-behavioural (blue), and biological (green).

Figure 3:. Iterative approach to public health decision making.

Conjectured use of AI to optimise the design and implementation of effective control measures during a hypothetical multi-country disease outbreak. Data collected from disease surveillance is processed and analysed by an AI agent, potentially integrating other AI models for parameter estimation, nowcasting, and forecasting of epidemic trajectories, as well as reconstructing historical transmission events based on current observations (top row). Prior to deployment of the AI model, it is trained using an outbreak simulator which simulates the spread of infectious diseases and the effects of different control measures; the model learns from these simulations by updating its model parameters according to the outcomes of the simulations and corresponding rewards (bottom right). Insights from the empirical data are used by the trained model to inform and recommend the most effective control measures, with validation and feedback from human policymakers and stakeholders to fine-tune alignment between model objectives and societally beneficial criteria. This is followed by the execution of public health actions (for example, whether to implement recommended or alternative control measures, or deploy further disease surveillance efforts) made ultimately by human policymakers (bottom left).