In response to the COVID 19 pandemic, it is heartening to see efforts at all levels to contain and manage the spread of this virus. Individuals, businesses, communities, the healthcare industry, and the government are all working hard to chart a safe path into an uncertain future. Attempts to reduce that uncertainty through various forms of predictive modeling can help all stakeholders to gauge their responses to this evolving challenge.
Our nation’s health protection agency, the CDC, is using the DELPHI group at Carnegie Mellon as a trusted source to predict the course of COVID-19 in the U.S. 1,2 Led by Dr. Roni Rosenfeld, DELPHI employs a combination of machine learning and crowd sourcing for estimating viral spread. Machine learning makes predictions based on extrapolation of patterns identified in historical data. Crowd sourcing consists of giving available historical data to a random sample of people and asking them to predict future trend. Employing these methods DELPHI has consistently outperformed competitors in predicting annual influenza spread.3,4
There are a variety of other approaches to modeling pandemics such as purely mathematical calculations, biologically based studies focused on understanding the disease, and even behavioral models that project individual responses to changes in the environment.
Researchers at Johns Hopkins’ Coronavirus Resource Center created their own flu pandemic model. With data available as of Jan. 31, they used it to predict the spread of coronavirus.5 They made their best estimates of actual incidence, average number of people infected by 1 contagious person(R0), incubation period, susceptibility, and duration of infectivity of COVID-19. Their model was specifically designed to assess the impact of air travel on predicting “hot spots” for spread of the virus in order to guide policymaking for travel restrictions and airport screening programs.
The model’s estimate of the top ten countries for initial spread from Wuhan was only off by one, with Malaysia predicted to be 11th and actually ranking 8th. The research went on to predict which U.S. cities would be most impacted, with New York, Los Angeles, San Francisco, and Washington D.C. most affected, and Seattle, Chicago, Houston, and Honolulu not far behind. Currently available data, while somewhat unreliable given the challenges with testing, suggest this model’s estimate was accurate in some respects, but not entirely predictive. New York, LA, SF, Chicago, and Seattle align with the model fairly well. On the other hand, Honolulu has only 53 confirmed cases and New Orleans, at 675, has more reported cases than LA (662), but is not even on the model’s radar.
Even as our experts and leaders use predictive models to help guide us through this pandemic, we all need to acknowledge the significant limitations in trying to predict the future. Dr. Rosenfeld at DELPHI cautioned, “You could be accurate because of luck, you could be inaccurate because of bad luck. You cannot draw many conclusions from a single season."4 With that in mind, let’s all take a deep breath and continue to “plan for the worst and hope for the best.”
3 Go to https://delphi.cmu.edu/epicast if you want to participate in crowd sourcing.