Picturing the progress of COVID-19

Computer models of COVID-19 are one of the most visible aspects of media coverage of the pandemic.  As economic and political pressure grows to relax social distancing and open schools and other institutions, these models also become targets for criticism.

Researchers at Pitt’s Graduate School of Public Health create the pandemic models used by Pennsylvania policy makers. The school’s Public Health Dynamics Laboratory (PHDL) was developing a dashboard tracking the state’s opioid epidemic when the Department of Health asked for help in early March. PHDL now creates the simulations regularly referenced by Secretary of Health Rachel Levine.

In March, the CRC granted PHDL a large computing allocation for the growing volume and complexity of the data in the simulations – one of several priority allocations and consultations CRC made possible for researchers working on COVID-19 projects.

The modeling system developed by PHDL – Framework for Reconstructing Epidemiological Dynamics, or FRED – relies heavily on the computing resources of Center for Research Computing. FRED is an agent-based modeling system, which means it incorporates data on the independent actions and interactions of individuals and organizations within a larger system such as a city or state. That amounts to a lot of data.

“Large, complex agent-based models require a substantial amount of computational resources to run,” explains PHDL director Mark Roberts, MD, MPP. “Each run of the model must start from the first infection and build the epidemic forward. For the state of Pennsylvania, this means running a 12-million agent model every day for an entire season. The resources of the CRC have been extremely helpful in allowing us to provide answers and thoughts about mitigation strategies in a reasonable time frame.”

Mary Krauland, visiting research assistant professor at the PHDL, describes the calculations. “We carry out many runs with different parameter sets, doing the whole state person by person based on a synthetic population whose statistics match those of the 2010 census. We add different characteristics into calculations we run in parallel, all of which makes this very computationally expensive.”

Kim Wong, research associate professor at CRC and a faculty member of PHDL, developed the configurations of the software and hardware access needed for the FRED simulations.

“We need large memory nodes on CRC’s message passing interface cluster,” says David Galloway, research programmer at PHDL. “We work with Kim quite a bit. He set us up so that we are able to use a flexible set of software, which means we can use a wide range of data to run the simulation and generate the output file for analysis. Kim created a reservation that allowed us to jump the queue, as part of CRC’s policy to facilitate COVD-19 research. That saved a lot of time. We can run Allegheny County on a desktop – a big desktop – but without CRC resources we could not run the entire state.”

The first FRED model looked at the dynamics of an initial outbreak of COVID-19 as the disease traveled through the demographics of each county to estimate how many deaths, hospitalizations, and other consequences could occur. Data now available from the real epidemic is being used to refine the models.

“We are working on a better calibration,” says Krauland. “It’s hard with new diseases because there are so many unknowns, like how many people may be asymptomatic. The numbers we have are fuzzy and don’t really tell you how many people were infected. It is hard to model based on hospitalization rates because that shows only the most severe cases. We can’t tell how many people were infected without real data.”

The FRED team is now modeling the impact of decreasing social distancing guidelines in Pennsylvania. Conversely, these models could also demonstrate the impact of stay-at-home orders. Data now available from cell phones and Google searches help create a picture of how much time people spend in one location versus the time they are mobile. Epidemiologists generally contend that increasing mobility will mean increasing numbers of infections. Krauland estimates that 60 percent of Pennsylvanians complied with social distancing at the height of the shutdown in late March and April, and that percentage is tailing off.

“Within that increased mobility of the general population the virus is concentrated in hotspots – nursing homes, prisons, individual factories. FRED is good at capturing details of real outbreaks within specific demographics and carving that population out of models of more homogeneous populations that surround those demographics,” says Krauland.

Given the possible consequences of applying pandemic models, and the resulting positive and negative attention, does this feel like a big responsibility?

Krauland reflects. “We present a model with caveats – the model is based on certain assumptions, including a worst-case scenario. Results may differ even in the same scenario. With so many models of the pandemic available now, what we really see is an ensemble of models, with various methodologies and assumptions. We hope we can contribute to creating the understanding among people at large of what goes into studying  public health.”

Contact:
Brian Connelly
Center for Research Computing
bgc14@pitt.edu