Donate to Science & Enterprise

S&E on Mastodon

S&E on LinkedIn

S&E on Flipboard

Please share Science & Enterprise

A.I. Employed for Local Covid-19 Prediction Models

Network illustration

(Gerd Altmann, Pixabay)

31 July 2020. Computer scientists in California are writing deep learning algorithms that generate forecasts for Covid-19 infections in local communities. Researchers from University of California in Santa Barbara developing the technology described the models earlier this month at the Computing Research Association, or CRA, Virtual Conference.

Xifeng Yan and Yu-Xiang Wang, professors of computer science at UC – Santa Barbara, are seeking better techniques for community leaders to predict the extent of Covid-19 infections in their cities and counties. Most current models forecast Covid-19 case loads and hospitalizations at national or state levels, yet communities are faced with making decisions on opening businesses or schools, for example, without data predicting infection rates for their local areas.

Current models, say Yan and Wang, also require a lot of highly granular data for local communities that may not be readily available, if at all. To deal with these real-world conditions, the computer scientists propose adapting deep learning models, a form of artificial intelligence, that find patterns in the mass of data to derive guidance on community Covid-19 infections. “The challenges of making sense of messy data,” says Wang in a university statement, “are precisely the type of problems that we deal with every day as computer scientists working in A.I. and machine learning.”

In April, the UC – Santa Barbara team received a nearly $200,000 award from National Science Foundation for the one-year project. Their approach aims to capture early experience with Covid-19 infections in some communities and apply those patterns to similar communities in the U.S.

The researchers use a deep learning model called transformer for the task. Transformer models were first designed for natural language processing, where they interpret one sequence of data, such as a string of characters or symbols, and apply the structure or meaning from that text to another sequence of data. A key part in this process is identifying more important elements in the sequence, which helps accurately predict later data sequences.

In this case, the important element in the sequence of data is the time period to find the most relevant dates to apply for predicting outcomes in another community. “If we are trying to forecast for a specific region, like Santa Barbara County,” notes Yan, “our algorithm compares the growth curves of Covid-19 cases across different regions over a period of time to determine the most-similar regions. It then weighs these regions to forecast cases in the target region.”

Census data train the algorithms

To train their algorithms, the team drew data from the American Community Survey conducted by the U.S. Census Bureau. The annual survey captures demographic, economic, health, education, housing and other community variables from 3.5 million households in the U.S. Databases generated by the survey provide highly detailed geographic data, down to census tract and block groups with 600 to 3,000 people. Cottage Health, a Santa Barbara health care system, provided clinical guidance for the project.

“When you combine [Census data] with Covid-19 data available by region,” says Wang, “it helps us transfer the knowledge learned from one region to another, which will be useful for communities that want data on the effectiveness of interventions in order to make informed decisions.”

Using Santa Barbara County in California as a test case, the team found the county’s summer 2020 spike in Covid-19 cases resembled three urban counties in North Carolina in March and April. Predictions for Santa Barbara Country using the model, say the researchers, have a mean absolute percentage error — a common measure of statistical predictive accuracy — of 0.030, lower than more widely used models.

Yan and Wang plan to collect data on more regions and communities in the U.S., and publish the model and data online. “We hope to forecast for every community in the country,” adds Yan, “because we believe that when people are well informed with local data, they will make well-informed decisions.”

More from Science & Enterprise:

*     *     *

Comments are closed.