An AI tool, created by researchers at Johns Hopkins and Duke universities, could revolutionize how public health officials predict, track and manage outbreaks of infectious diseases including flu and COVID-19.

“COVID-19 elucidated the challenge of predicting disease spread due to the interplay of complex factors that were constantly changing,” said author Lauren Gardner of Johns Hopkins, a modeling expert who created the COVID-19 dashboard that was relied upon by people worldwide during the pandemic.

“When conditions were stable, the models were fine. However, when new variants emerged or policies changed, we were terrible at predicting the outcomes because we didn’t have the modeling capabilities to include critical types of information. The new tool fills this gap.”

The work is published in Nature Computational Science.

During the coronavirus pandemic, the technology that underpins the new tool didn’t exist. The team for the first time uses large language modeling, the type of generative AI used most famously in ChatGPT, to predict the spread of disease.

Instead of treating prediction merely like a math problem, the model, which is named PandemicLLM, reasons with it, considering inputs such as recent infection spikes, new variants, and mask mandates.

The team fed the model streams of information, including data never used before in pandemic prediction tools, and found PandemicLLM could accurately predict disease patterns and hospitalization trends one to three weeks out, consistently outperforming other methods, including the highest performing ones on the CDC’s COVIDHub.

“A pressing challenge in disease prediction is trying to figure out what drives surges in infections and hospitalizations, and to build these new information streams into the modeling,” Gardner said.

The model relies on four types of data:

• State-level spatial data including information on demographics, the health care system and political affiliations.
• Epidemiological time series data such as reported cases, hospitalizations and vaccine rates.
• Public health policy data including stringency and types of government policies.
• Genomic surveillance data including information about the characteristics of disease variants and their prevalence.

After consuming this information, the model can predict how the various elements will come together to affect how the disease behaves.

To test it, the team retroactively applied it to the COVID-19 pandemic, drilling down on each U.S. state over 19 months. Compared to other models, the new tool was particularly successful when the outbreak was in flux.

“Traditionally we use the past to predict the future,” said author Hao “Frank” Yang, a Johns Hopkins assistant professor of Civil and Systems Engineering who specializes in developing reliable AI.

“But that doesn’t give the model sufficient information to understand and predict what’s happening. Instead, this framework uses new types of real-time information.”

With the necessary data, the model can be adapted for any infectious disease, including bird flu, monkeypox and RSV.

The team is now exploring the capability of LLMs to replicate how individuals make decisions about their health, hoping such a model would help officials design safer and more effective policies.

“We know from COVID-19 that we need better tools so that we can inform more effective policies,” Gardner said. “There will be another pandemic, and these types of frameworks will be crucial for supporting public health response.”