New forecasting tool improves accuracy of epidemic peak and hospital demand predictions

During an epidemic, some of the most critical questions for health care decision-makers are the hardest to answer: when will the epidemic peak, how many people will need treatment at once, and how long will this peak in demand for care last? Rapid responses can help hospital administrators, community leaders and clinics decide how to deploy staff and other resources most efficiently. Unfortunately, many epidemiological forecasting models tend to have difficulty accurately predicting cases and hospitalizations around peaks.

A new approach described in the review Proceedings of the National Academy of Sciences and led by researchers at the University of Texas at Austin, integrates a critical element of epidemiological understanding into forecasting models to address these long-standing issues. Rather than simply extrapolating current epidemic trends, the approach, known as “epimodulation,” gives models a more intuitive sense of how epidemics generally tend to evolve.

“This tells the model, in effect, ‘We expect the curve to bend as immunity builds,’ so the model can look for early signs of this slowdown while continuing to learn from the data,” said Lauren Ancel Meyers, Cooley Centennial Professor in UT’s Department of Integrative Biology and director of epiENGAGE, a national center of excellence in forecasting and analysis of epidemics. “The result is a better forecast that provides real-time information to hospitals and communities when it matters most. »

The team tested their approach on a wide range of models and with real data from past outbreaks of influenza and COVID-19. They found that the approach increased model accuracy by up to 55% during epidemic peaks for forecasting hospital admissions, without reducing accuracy outside peak times. Epimodulation has also improved the accuracy of ensemble models, which combine multiple models into a single forecast. The results suggest that this may be a powerful new tool for health systems to adapt to rapidly evolving epidemics.

According to Meyers, this approach could be applied to many infectious diseases that spread in waves, including avian flu, Ebola, Mpox and even new pathogens that have not yet emerged. Such wave patterns often appear as immunity develops within a population, individuals change behavior, or environmental conditions change.

“Outbreaks tend to follow recognizable patterns. They grow very quickly at first, then slow down as more people become immune or change their behavior, eventually peaking and fading,” Meyers said. “These dynamics reflect basic epidemiological principles: how infections spread, how immunity builds, and how people respond when risk increases.”

Most forecasting models, especially those based solely on machine learning, do not “know” any of these epidemiological principles. They basically look at recent data and project the trend forward, like extending a line on a graph. They often get good results when cases rise (or fall), but miss the turning point when growth slows or reverses. Epimodulation can help make predictions around the peak more realistic.