Prof. Dimitris Bertsimas will be speaking about "Predictive and Prescriptive Analytics for the COVID-19 Pandemic" at the AI Cures Conference: Data-driven Clinical Solutions for COVID-19 on September 29th. His brief talk will be based on the innovative work he and a group of his doctoral students have developed since March of this year.
The group has created an array of machine-learning tools to predict the spread and impact of COVID-19. They collected clinical data—including symptoms and demographics—from about 10,000 patients in 44 places across the globe. Harnessing that data, the group designed models to predict the risk of both infection and death for individual patients.
These models can be used to build personalized “calculators”—for example, tools for clinicians to help them triage patients, or to screen people for testing over the phone. So far, several European hospitals have used the risk calculators to make decisions before and after triage. In South America, a major financial institution uses the risk-of-infection calculator to decide how employees can safely go back to work.
An offshoot of the project uses machine-learning to match personalized treatments to patients. The idea is that different people react differently to treatments,” Bertsimas says. The program could be a boon for patients, but with a caveat: since clinical trials for COVID-19 treatments are incomplete, it’s harder to know the positive effect of one treatment compared to another. The group is currently perfecting the model before publishing.
Understanding the disease at a population-wide level is equally as important as treating individuals. Starting in March, the group began developing a new machine-learning model, called DelPhi (“Differential Equations Leads to Predictions of Hospitalizations and Infections”) to predict the number of COVID-19 cases, hospitalizations, and deaths for all 50 US States and 130 countries. In certain regions, the group works with collaborators to make predictions at an even finer scale. A particular project in Brazil, for example, zooms in on a county level.
“It depends where we have data,” Bertsimas says.
In April, the number of COVID-19 cases in the US surpassed 200,000. The DelPhi model predicted that number would continue to soar, reaching anywhere between 1.2 and 1.4 million cases by mid-May—a prediction that came true.
By predicting how quickly the virus spreads—and where—the models can help inform crucial public health policies. What happens, for example, if a government lifts or reinstates a lockdown? DelPhi can point to how many lives could be lost or saved.
The model can also aid in research. Recently, a major pharmaceutical company used DelPhi to identify areas to conduct clinical trials. Similarly, the Bertsimas group is using DelPhi predictions to come up with a strategy for distributing vaccines. If a vaccine becomes available, and if it is like a flu shot—effective maybe 60% of the time—then deciding how it will be distributed is crucial.
“We want to allocate vaccines to the populations that are more likely to infect others,” Bertsimas says. Not only will it improve those lives, “we will lower their infection rate in the future.”
The DelPhi model and risk calculators exist online and are available to the public at covidanalytics.io. Bertsimas says the site, which is updated daily, receives thousands of hits each day.