Virginia Tech researchers garner two major awards in COVID-19 forecasting challenges
January 29, 2021
DeepOutbreak, a team of researchers from Virginia Tech, Georgia Tech, and the University of Iowa, has taken first place in the COVID-19 Symptom Data Challenge.
The competition explores how Facebook symptom survey data can enable earlier detection and improved situational awareness of COVID-19 and flu outbreaks that can help both public health authorities and the general public make better decisions.
The first place award, announced by Catalyst @Health 2.0 in late December, and the team’s work will be featured on the Facebook Data for Good blog. Facebook was one of the sponsors of the challenge.
Earlier in December, DeepOutbreak took second place in the COVID-19 Grand Challenge, a competition hosted by the enterprise artificial intelligence (AI) firm C3.ai.
“The basis of our work for both competitions was the same though the emphasis of each competition was slightly different,” said Naren Ramakrishnan, director of the newly announced Sanghani Center for Artificial Intelligence and Data Analytics, formerly the Discovery Analytics Center. “While the C3.ai challenge primarily tasked us with integrating a diverse range of data sources to improve forecasts, Facebook provided access to its own data for the Symptom Data Challenge and also asked that we explore when that data was useful and when it was not.”
DeepOutbreak’s study, “Steering a Historical Disease Forecasting Model Under a Pandemic: Case of Flu and COVID-19,” is the first to address the problem of adapting to shifting epidemiological trends. It uses a neural transfer learning architecture that successfully steers a historical disease forecasting model to new scenarios where flu and COVID co-exist and automatically learns when it should emphasize current COVID-related signals and when it should accentuate the historical model.
The researchers showed that methods based on historical data alone cannot predict the atypical nature of COVID-I9 with multiple co-occurring phenomena like actual COVID-19 infections, corresponding shutdowns and societal lock-downs, and changes in health care-seeking public behaviors. The team’s model can accommodate multiple COVID-related data for more accurate forecasting.
In addition to Ramakrishnan, the Virginia Tech team included Nikhil Muralidhar and Anika Tabassum, two of the Sanghani Center’s Ph.D. students in computer science and research trainees in the National Science Foundation-sponsored urban computing graduate certificate program.
Other team members are Aditya Prakash, associate professor, and Alexander Rodriguez, a Ph.D. student at Georgia Tech, and Bijaya Adhikari, assistant professor at the University of Iowa, all formerly part of the Sanghani Center.
“The use of AI and data science in public health is still fairly new. That is why these kinds of competitions are great at showcasing what might be possible with new techniques and data sources and where we might need to direct more efforts as a community,” said Prakash. “Additionally, doing this in the middle of a novel ongoing epidemic gives a very unique — and challenging — setting to explore the utility of various methodologies.”
Ramakrishnan said the Sanghani Center has a strong foundation in forecasting and that faculty and students have participated in CDC and IARPA forecasting groups and programs for a number of years.
“So, applying our work to competitions like these is a logical next step and allows us to further engage in one of our main objectives, data for the public good,” said Ramakrishnan.
“As a Ph.D. student, I have been fortunate to be a part of the urban computing program, which offers opportunities to participate in projects that apply data mining to urban environments and have explicit real-world impact,” said Muralidhar. “This experience motivated me to join the DeepOutbreak team to contribute what we can, through our research, to the global effort in addressing and containing the pandemic.”
“It is always a goal of mine to contribute to society with my work. Both of these competitions had public health practitioners as judges, which allowed us to leverage sophisticated machine learning tools without losing sight of requirements from domain experts,” said Rodriguez. “Working with my advisor and collaborators toward this goal has been a great learning experience for me.”
Muralidhar said that the highly dynamic, quick turnaround time for the competitions, compared to often longer-term research time lines, requires developing and iterating solutions quickly.
“I believe these are imperative skills for working at any enterprise,” he said.
Written by Barbara L. Micale