COVID-19 Surge Modeling

In March 2020, the HSIR team was asked to start reviewing national surge models and investigate how we could use those tools for predicting the inpatient surge at Parkview. We quickly determined that a local model – one that accounts for the unique attributes of the communities we serve, COVID-19 mandates and events occurring in our region – and the local parameters would be a better plan. Due to the lack of data, we started with simulation models – first the Susceptible-Exposed-Infectious-Recover (SEIR) and then an additional “R” (SEIRR) for “Re-infection”. The actual numbers were tracked alongside the lower and upper predicted bounds for a six-month horizon (See Figure 1). These models are known to have wide ranges of error due to their mathematical limitations. Yet, without a corpus of ground data to use to train a predictive model, this was what was available to most everyone across the world during the early stages of the COVID-19 pandemic.

As more local data was available, we were able to pivot to actual predictive modeling. This modeling looked at a shorter time frame (1-3 weeks) as depicted in Figure 2. By focusing on this time frame and using our local data including hospitalizations, infection rates, and population distributions, the model has been extremely accurate (~98%) and has continued to predict our micro-and macro-surges since early 2021. This tradeoff of long-term prediction for short-term accuracy has helped our Parkview decision makers make critical decisions about staffing, PPE, and other core resources throughout the multiple waves of the pandemic.