AEM: An AI-Powered and Epidemiology-Informed Modeling System

Research Team: Dongdong Wang, Shunpu Zhang, and Liqiang Wang

Department of Computer Science and Department of Statistics and Data Science

University of Central Florida

COVID-19 infection rates may be peaking in Orange County later this month and trending down toward December, according to the latest artificial intelligence and deep learning model developed by data scientists in the Department of Statistics and Data Science and the Department of Computer Science at the University of Central Florida.

“The model we developed is one of the finest epidemic models using deep learning techniques, and our current predictions are based on the data to date, and will be updated as more data become available,” said Shunpu Zhang, Ph.D., Chair of the Department of Statistics and Data Science, who worked on the project with Associate Professor of Computer Science, Liqiang Wang, Ph.D., and Ph.D. student Dongdong Wang.

The trio developed the projections by feeding data from Johns Hopkins University and the New York Times into 10 different deep neural networks. Each deep neural network was trained with about 50,000 simulations from classical compartmental epidemic models, including SIR/SEIR, widely accepted epidemiological models. The resulting model is a mixture of the 10 deep neural networks, which include the variables so policy makers can also see the best-case and worst-case scenarios.

Based on the observation data through August 17^th, the projection results for Florida are updated and shown in the figure below. The scenarios include:

· The daily increase of COVID-19 positive cases continues decreasing for the rest of this year and reaches 700 cases in late October.

· The number of the death cases is expected to continue decreasing, the daily deaths will reach 10 cases in late October, and the maximum rate will be approximately 14 thousand (Corrected by including the number before June 26^th).

· The maximum infection rate is expected to be reduced to 0.8 million.

Figure 1 Model prediction results for Florida. The results are derived from 7-day moving average.

https://video.wucftv.org/video/floridas-future-covid-19-trends-lvsuzt/

------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Based on the observation data through July 30, those scenarios include:

· The daily increase of COVID-19 positive cases will begin to slow around early August for Florida and the U.S., and the end of July for Orange County.

· The maximum infection rate is expected to be approximately 12 million for the U.S., 1.4 million for Florida and 55,000 for Orange County.

· The number of ICU beds occupied by COVID-19 patients will peak in early August for the U.S. and Florida, end of July in Orange County.

· There is a strong correlation between Florida and the U.S., indicating the state plays a prominent role in informing national policy. Similarly, Orange County tracks Florida’s numbers.

· The researchers also explored a hypothetical situation that extended Florida’s shutdown, instead of opening up into Phase 2 in early June. Charts show a dramatic rise in positive cases almost immediately after reopening began and the numbers tell the same story. The projected ultimate total infection rate tripled in the U.S. from 4 million to 12 million; expanded 14 times in Florida, from 100,000 to 1.4 million; and jumped 27.5 times in Orange County, from 2,000 to 55,000.

Charting COVID-19’s rise and fall took an innovative modeling approach. The novelty of the virus and its limited data are grand challenges to calibrate compartmental epidemic models, which could be well-handled by deep learning methods, Dr. Wang explained. Following guidance from Drs. Zhang and Wang, Dongdong Wang developed an approach by blending compartmental model and deep learning to more efficiently and accurately fit observed data and generate more reliable infection trajectory. “Our method is flexible and could be generalized into a variety of combinations of compartmental model and deep learning,” said Dr. Zhang. The team is actively watching the data update and calibrating the deep learning models to include new reopening events like the NBA relocating to Orlando, schools reopening in August, and tourists visiting Orange County.

The temporary closure of some testing sites in Florida due to Hurricane Isaias may cause possible data anomaly in new positive cases.