Electrical engineering graduate student’s project predicts daily COVID-19 infections
May 17, 2020 / by Nanci Hellmich
“The Kalman Filter is an algorithm that can make predictions for a dynamical model one step ahead in time—in this case, a single day. ... It tracked the total number of infections each day pretty closely."
— Matthew Thornton, master's student in electrical engineering
Since the coronavirus pandemic began, experts have been analyzing data sets to try to predict the number of people who will be infected with the virus.
Using nationwide statistics for COVID-19, a Mason electrical engineering master’s student tested a computational method to estimate new infections each day.
“The Kalman Filter is an algorithm that can make predictions for a dynamical model one step ahead in time—in this case, a single day,” says Matthew Thornton, who just finished his master’s degree in electrical and computer engineering. “It tracked the total number of infections each day pretty closely,” he says.
Thornton did this project as part of a 600-level graduate course, Kalman Filtering with Applications, taught by Professor Andre Manitius.
The Kalman Filter has been used for trajectory estimation in space flights, including lunar and Mars landings, and it is the algorithm that is the crux of the Geographical Positioning System, says Manitius. More recent applications include modeling of economic and social systems, he says.
Thornton ran the Extended Kalman Filter (EKF) on the U.S. COVID-19 data set from Johns Hopkins University from late January through the end of April.
"In the final week of April, the average error between the predicted number of new daily infections and the corresponding Johns Hopkins’ data was less than 25 percent,” he says. He used the technique to analyze data in Virginia and found the predictions to have a similar degree of accuracy.
The model also produced one-week predictions with substantially lower percentage errors. These results indicate that the EKF might be useful for short-term tracking and predictions of the pandemic’s spread, says Thornton, who works for a defense contractor, specializing in signal processing.
Superior models and parameter estimation strategies could be integrated with the EKF in the future to create a more robust prediction tool, he says.
Manitius says, “I am very impressed by the research that Matthew has done on this problem. He figured out how to estimate the crucial parameters that determine the rate at which the infection is spreading and plugged these parameters into the nonlinear model of the pandemic.
“He went much beyond the course requirements and wrote a report worth publishing. His model can be easily adapted to other studies of this pandemic,” he says.
Although the class is over, Thornton is continuing to work on the project. “I plan to plug in some more data because I want to track it at least through the curve flattening to see if the filter is useful,” he says.