Follow Datanami:
July 1, 2020

Sandia Labs Model Predicts COVID-19 Cases Over a Week into the Future


Just over the last several weeks, the United States’ COVID-19 trajectory has dramatically changed, shifting from a promising decline to a terrifying spike that may eclipse the first. Policy-makers are struggling to make decisions about reopening (or shutting down again) amid the chaos, uncertain how future trends will validate or contradict their decisions. Now, a team of researchers from Sandia National Laboratories have built a data-driven model that predicts new cases up to ten days into the future with high accuracy.

Using data from the Centers for Disease Control and Prevention (CDC), the New York Times Data Repository, Johns Hopkins University and a number of other more geographically specific sources, the researchers – Jaideep Ray and Cosmin Safta – feed a model that Ray developed in the late 2000s for tracking influenza-like illnesses. “When COVID-19 began to spread so rapidly,” Ray said, “we knew we could use the same method to help forecast the outbreak.”

“This method is a relatively easy and inexpensive way to get short-term forecasts about new coronavirus cases that decision-makers can use to allocate health care resources and response,” Safta said. “This method is much easier and cheaper to do than methods that require more robust computers and manpower.”

The model completes its forecasts within minutes without the need for high-performance resources. Even better, the forecast has generally been accurate, with cases “roughly” following the model’s predicted trends since April.

“The forecasts come with a range within which users can expect reality to lie,” Ray said. “The range changes daily depending on the data, but the model ensures that the user can have 95% confidence that reality will fall within the range.”

A simplified illustration of the model. Image courtesy of the authors.

The success of the model is a testament to how mitigation policies for COVID-19 are being transformed by the availability of high-resolution data at rapid speeds – a situation that would not have been possible even a decade ago. “For the current COVID-19 situation, having more sources of data dramatically assists our ability to create short-term forecasts to inform public health decisions,” Saft said.

“Since we are so connected today, it’s possible to get an accurate number of COVID-19 cases in a day and get it to everyone in the world within a 24-hour period,” Ray said. “Ten years ago, even five years ago, you could not get this data. In 2015, with the Ebola outbreak, by the time they got data it was pointless to try and make a forecast because it was already out of date and useless to decision-makers.”