Beyond the Horizon: The Transition from Reactive Surveillance to AI-Driven Epidemic Intelligence (2020–2026)

Abstract

Background: The classic indicator-based surveillance systems of infectious disease are, by definition, reactive and often
detect outbreaks when significant transmissions in the community have occurred. The COVID-19 pandemic revealed these
structural delays and hastened the employment of artificial intelligence (AI) in epidemic intelligence. During 2020-2026, the
progress in generative AI, large language models (LLMs), digital epidemiology, and AI-powered diagnostics has radically
changed the worldwide surveillance structures.
Objective: This systematic review brings together the evidence of the development, performance, and practical use of AI-based
epidemic intelligence systems since 2020, and especially in forecasting accuracy, early warning features, and the experience
of One Health systems.
Methods: The search was carried out in the databases of PubMed/MEDLINE, Scopus, Embase, Web of Science, IEEE Xplore,
ACM Digital Library, and popular preprint archives (medRxiv, bioRxiv, arXiv). A review of grey literature by world health
public agencies was also conducted. PRISMA 2020 guidelines were used to screen the studies published in English from
January 2020 to January 2026. Eligible studies assessed AI based on the surveillance systems, such as generative AI, LLMs,
motivate accelerated machine-learning models, or AI-based diagnostic systems, and reported empirical tests or functioning
results. The information was mined regarding AI architecture, sources of data, forecast horizon, comparative performance,
and maturity of the deployment.
Results: Twenty-two studies were found to satisfy the inclusion criteria of the 590 records identified. Generative AI and LLMbased
models continued to be superior to the existing statistical methods, especially when there was epidemiological instability.
The PandemicLLM framework was shown to predict trends in hospitalization 1-3 weeks before they occurred, and this was
superior during the emergence of a new variant and policy changes compared to CDC ensemble models. Digital epidemiology
systems based on natural language processing event surveillance made outbreak signal detection more accurate, and AI-enhanced
CRISPR-Cas and graphene field-effect-transistor biosensors allowed amplification-free detection of pathogens down to the
attomolar level in 30 minutes. Population level-wide surveillance persisting AI applications were demonstrated to be feasible
at a large scale, exemplified through the implementation of an Indian system, the Integrated Health Information Platform.
Conclusions: Epidemic intelligence based on AI has ceased to be an experimental innovation and has been implemented in the
operational infrastructure of public health. The change in predictive modelling to generative systems based on reasoning is a
paradigm shift in outbreak preparedness. However, the issues of bias in data, explainability, management, and fair data sharing
remain. The future advancement will require augmented-intelligence systems, federated learning, and strong international
governance to provide ethical, transparent, and worldwide inclusive epidemic surveillance.