New AI Tool Outperforms Top Models in Predicting Infectious Disease Spread
A cutting-edge AI model developed by researchers at Johns Hopkins and Duke University is setting a new benchmark in forecasting the spread of infectious diseases like COVID-19 and influenza. Backed by federal funding, this tool significantly surpasses current state-of-the-art prediction systems, offering a breakthrough in how health agencies can prepare for and respond to disease outbreaks.
The AI system—named PandemicLLM—introduces a revolutionary shift in disease modeling. Unlike traditional models that primarily rely on mathematical formulas, this new approach uses large language models (LLMs)—the same type of technology behind generative AI like ChatGPT. This allows it to “reason” through complex data sets, taking into account real-time developments such as emerging variants, public health measures, and shifts in behavior.
“COVID-19 highlighted the limitations of existing forecasting models, especially when conditions were volatile,” said Dr. Lauren Gardner, Johns Hopkins expert in disease modeling and the creator of the widely-used COVID-19 dashboard. “Our inability to adapt to new variables quickly made us poor at predicting outcomes. PandemicLLM addresses that challenge head-on.”
The research was recently published in Nature Computational Science.
A Smarter, More Adaptive Forecasting System
PandemicLLM was trained on a broad set of data sources, many of which had never been used before in epidemiological forecasting:
-
State-level spatial data: Demographics, political affiliations, and healthcare infrastructure.
-
Epidemiological time series: Daily case counts, hospitalizations, and vaccination rates.
-
Public health policy data: Government interventions, lockdowns, and mask mandates.
-
Genomic surveillance: Tracking and understanding the prevalence and behavior of emerging variants.
By integrating these diverse streams of information, PandemicLLM delivers forecasts up to three weeks in advance with greater accuracy than any current model, including those submitted to the CDC’s CovidHub. It was particularly effective during unstable phases of the pandemic when traditional models often failed.
“Most models look backward to predict forward,” explained Dr. Hao “Frank” Yang, an assistant professor at Johns Hopkins. “But our model uses dynamic, real-time data to understand what’s currently driving changes in disease trends.”
Ready for Future Pandemics
The tool was validated through a retrospective application to COVID-19 data across all 50 U.S. states over a 19-month period. It consistently outperformed existing forecasting tools, especially during periods of policy change or variant emergence.
Even more promising is the model’s flexibility. PandemicLLM can be adapted to other infectious threats like avian flu, monkeypox, and RSV, provided the necessary data is available.
Researchers are now exploring the next frontier: using AI to simulate how individuals make health-related decisions, such as whether to get vaccinated or wear a mask. This could help policymakers create interventions that are not only effective but also grounded in realistic human behavior.
“We need tools that allow us to act swiftly and intelligently in the face of evolving threats,” said Gardner. “This kind of AI-powered framework will be essential when the next pandemic comes.”
The research team includes Johns Hopkins PhD students Hongru Du, Shaochong Xu, and Yang Zhao; Jianan Zhao of the University of Montreal; Prof. Yiran Chen of Duke University; and Xihong Lin of Harvard University.
