DeepCOVID
For COVID-19 US Mortality and Hospitalization Forecasting

About

The Centers of Disease Control and Prevention is hosting collaborative forecasting projects related to predicting the coronavirus disease spread, and anticipate the mortality and number of hospitalizations caused by the disease across the country. This critically timed project comprises a handful of teams of leading data scientists, epidemiologists, statisticians, and high-performance computing researchers from national laboratories, public universities, public health institutions, and some private sector agents. Our team is using deep learning models to forecast specific targets at the national, regional, state, and local levels. In addition to CDC data, we are incorporating many other real-time datasets such as syndromic surveillance data and point-of-care data from major providers. We combine these datasets with domain knowledge using end-to-end deep learning models to predict targets on a weekly basis. The CDC synthesizes our weekly and monthly predictions with other models to help determine policy and other planning decisions to help communities prepare for and fight the disease. We have extensive experience with disease surveillance - AdityaLab has been leading a team at the CDC FluSight challenge since 2018 (our model EpiDeep had the best performance in the HHS1 region) and have published multiple research papers on incidence prediction at major venues (see here).


Approach

We use a data-driven approach based on deep learning for forecasting mortality and hospitalizations using syndromic, clinical, demographic, mobility and point-of-care data. The model uses data sources (mobility patterns etc) which implicitly capture the interventions in place. As the model is updated each week, the effect of any new intervention is taken into account via the changes in the input data. The learned model is then used to predict mortality and hospitalization at various points in future. To compensate for the effects of noise (in data and initialization), the model is bootstrapped. The model also propagates uncertainties in the data to show the confidence intervals in the forecasts.


Latest Results

forecasting1 forecasting2

We show our current results (June 1) for US national cumulative and incidence mortality in the pictures above using the JHU dataset as ground truth. Also see our latest results on mortality prediction on the CDC website. Other results (hospitalizations etc) will be released later, depending on terms of the projects.


Contact Us

Please email us with any comments/suggestions. Emails are also welcome if you want to help, collaborate or support our work!

Funding

Our project is partially supported through (NSF CAREER, NSF CISE Expeditions, NSF IIS, NSF COVID RAPID, ORNL, GT and GTRI Rapid Seed Funds).


Team

Georgia Institute of Technology

IQVIA




University of Illinois, Urbana-Champaign

gt-logo uiuc-logo iqvia-logo