top of page

The COVID-19 Pandemic: Machine-Learning for Dilemmas at the Front Line

Two short weeks ago, Prof. Dimitris Bertsimas (Sloan School of Management) and nearly two dozen doctoral students began developing a suite of tools for confronting the COVID-19 pandemic. Each of the three projects - which tackle distinct problems - are made possible by the power of machine-learning.

The first project addresses dilemmas at the front line. As the medical system awaits a tidal wave of COVID-19 patients, hospitals anticipate a need for more supplies and equipment. Protective gear must go to healthcare workers, ventilators to critically ill patients. In response, Bertsimas and the group are developing an epidemiological model meant to track the progression of COVID-19 in a community. The hope is for hospitals to dynamically predict a surge of activity across their floors, so they can allocate resources in the best way.

To hone the model, the group started with data published in Wuhan, Italy, Spain and the United States. They included the infection and death rate - as well as data coming from patients in the ICU and the effects of social isolation. Through the process, the researchers are testing the model against incoming data to see if it makes a good prediction. At the same time, they continue to add new data and use machine-learning to make it more accurate.

“The model is quite accurate because it learns from the data,” Bertsimas says.

Only a few days ago, the team linked up with longtime collaborators at Hartford Hospital to deploy the model, helping the network of seven campuses to assess their needs. They expect to launch a national version that could in theory help the federal government allot ventilators to states, although that level of coordination would be unlikely. No matter what, different regions will hit their peak number of cases at different times, meaning their need for supplies will fluctuate. The model could be helpful in shaping future public policy.

“We’re not talking about months, we’re talking about weeks differences,” Bertsimas says.

The second project focuses on the individual patient. The group is building a mortality and disease progression calculator to predict whether someone has the virus, and whether they need hospitalization or even more intensive care. So far, Bertsimas says, advice for COVID-19 patients is “at best based on age, and perhaps some symptoms.”

By probing the severity of the disease in a patient, he says, “it can actually guide the advice in a much better way.”

Similar to the capacity planning model, the calculator uses machine-learning to make stronger predictions based on the patient's symptoms, demographics and geography. But right now, not much data on individual patients exists. To work around this problem, the group used a simulation experiment to generate patient data. They continue to add data from new studies to the calculator all the time.

An additional project focuses on a convenient test for COVID-19. Researchers at the Mohammed VI Polytechnic University in Morocco are developing the test which involves using spectroscopy data to analyze samples of saliva. They hope for a test that can be processed locally - as opposed to being shipped to a lab - and completed in minutes.

Bertsimas and his group recently received data from about 100 samples from Morocco. Their group is using machine-learning to augment the test, coming up with results that are more precise than they would be otherwise. Currently, the model can accurately detect the virus in patients around 90% of the time, while false positives are low. In theory, a future version could be deployed in countries where resources are even scarcer than in the US.

The backbone for each of these machine-learning projects is data. For now, much of that data is locked away in the large number of papers being published each day. Only two weeks ago, twelve people in the group hunkered down to extract data from over 120 papers, producing a database that could be used for their models. With no trick to speed up the work, the team read paper after paper - a work-intensive process that took about four days. They released the database - which they continue to update - on their website, to be plugged into any number of models assembled by other researchers.

“Our objective is to release all the data to make it available for other people too,” Bertsimas says. “We do our own analysis, but if other people have better ideas, we welcome it.”


[1] COVIDAnalytics,

bottom of page