As the coronavirus pandemic brings floods of people to hospital emergency rooms around the world, physicians are struggling to triage patients, trying to determine which ones will need intensive care. Volunteer doctors and nurses with no special pulmonary training must assess the condition of patients’ lungs. In Italy, at the peak of that country’s crisis, doctors faced terrible decisions about who should receive vital resources.
Artificial intelligence can help. AI systems that have been trained via machine learning to offer “clinical decision support” may play an important role in the COVID-19 crisis, helping to keep hospitals functional and patients alive. They may also help with decisions about how to ration care, should it come to that.
Around the United States, some medical centers are repurposing existing AI systems meant to predict the course of patients’ illnesses, retooling them to predict specific COVID-19 outcomes such as intubation.
These AI systems have learned about patterns of illness by ingesting data from thousands of patient records—and while there isn’t enough data from COVID patients yet to create entirely new prediction tools, researchers are checking to see if existing tools can be customized to help with COVID.
IEEE Spectrum spoke with doctors and researchers at the University of Chicago, Stanford University, and Johns Hopkins University, where projects have been fast-tracked and are nearly ready for use.
University of Chicago
At the University of Chicago Medical Center, physician Dana Edelson is working on an upgrade to an AI system called eCART, which the hospital launched in 2015. “We’re testing the new version right now, and we expect to push it out in the next few weeks,” says Edelson, who cofounded the company AgileMD to commercialize eCART.
ECART is already in use at more than 20 hospitals around the United States. The existing version predicts which patients will have a cardiac arrest, which will need transfer to the intensive care unit (ICU), and which will die—all within the next eight hours. It uses data from patients’ electronic medical records (such as vital signs, lab results, and demographic information) to provide real-time risk scores for patients.
The upgraded version will still predict those outcomes, and it will also “predict more specifically the need for intubation and the need for IV medications that increase blood pressure, both of which require the patient to be in the ICU,” Edelson says. The project was already underway when the pandemic began, but the team has accelerated the work in the past months.
In the current version, which uses about 30 variables in its calculations, the most predictive feature was the patient’s respiratory rate. In the new version, which uses about 100 variables, they found that the amount of supplementary oxygen required to keep a patient’s blood-oxygen level up was the most predictive signal that a patient’s condition was deteriorating. Edelson says that increasing oxygen requirement is a particularly important signal for COVID patients whose lungs are failing.
Currently, the team is working with other institutions to validate the new eCART version on COVID patients, making sure it accurately predicts COVID-specific outcomes. Matthew Churpek of the University of Wisconsin, who is leading the validation effort, says the collaboration will include data from at least seven hospitals across multiple states.
What can doctors do with eight hours’ warning that a patient will need to be intubated or transferred to the ICU? While they probably can’t prevent those patients’ conditions from deteriorating, Edelson says, they can try different treatments immediately, like affixing nasal cannulas that deliver supplemental oxygen through the nose.
Ultimately, Edelson believes the predictions can improve health outcomes for both patients and hospital staff. “Our goal is to minimize emergencies. In the TV shows, you see people running into the room, and they look really heroic. But in this setting, they’re really dangerous,” says Edelson. “With COVID, we worry about both the patients and the providers.”
COVID-19 patients sometimes crash suddenly and require immediate intubation; that’s the procedure where a tube is pushed through the mouth and down the airway to allow a mechanical ventilator to assist with breathing. The procedure can spray virus-laden droplets from the patient’s airways into the air, putting the medical staff at risk of infection.
When done during an emergency, Edelson says, intubations are even more dangerous for providers. “A bunch of people swarm into a room, and there’s a risk that they’re not adequately adorned in [personal protective equipment],” she says. “And then there’s the heightened anxiety and stress, so things may not go as smoothly, and it may take several attempts to successfully intubate a patient. Every failed attempt increases the risk to providers, and every extra person that’s in the room puts more providers at risk.”
Another goal of the project is to more efficiently use resources and staff time, says University of Wisconsin’s Churpek, for the good not only of the COVID patients, but for the other patients in the hospital who also need care. “We have to be sure that we’re not leaving behind all the other patients who need critical care resources at the right time,” he says.
Stanford University’s Ron Li, a physician who leads Stanford’s efforts on integrating AI into clinical workflows, says his team is currently evaluating another AI system that could help. Stanford has previously worked with a model provided by its electronic health record vendor, Epic; the model predicts which patients are likely to need an ICU transfer, have a cardiac arrest, or die.
Now Li’s team is testing whether this “Deterioration Index” can also accurately identify COVID patients whose condition will deteriorate. It’s looking good, he says: “We’re trying to implement this in the next couple of weeks.” The San Francisco Bay Area hasn’t experienced the surge that many feared, Li says, “but we are designing this AI-enabled system to prepare for any future surges that may happen.”
Right now, assessing COVID patients requires a lot of staff time, Li says. “Currently, a dedicated triage team closely monitors patients with COVID in our hospital. It’s a highly manual and resource intensive process of chart review and in person evaluations,” he says. “But if you have hundreds of patients in the hospital, that won’t work.” What’s more, an automated system provides an objective measure for the whole care team, doctors and nurses alike, which may prevent ailing patients from falling through the cracks.
In an email, Epic tells IEEE Spectrum that its Deterioration Index was originally trained on more than 130,000 patient encounters. In March, the company provided an update that hospitals could use to evaluate the model for COVID patients. By the end of the month, “six organizations had produced results for performance across approximately 3,000 COVID-19 positive patients,” an Epic spokesperson says. “The results showed that the model performed well and didn’t need to be changed. Some hospitals are now using the model with confidence. Stanford is in the process of evaluating the model before they begin using it with their patients.”
Johns Hopkins University
But some experts say that modifying tools that predict general deterioration won’t provide the best results for COVID patients.
Bayesian Health, a stealth-mode startup that spun out of Johns Hopkins University, previously developed a targeted prediction tool for sepsis, a life-threatening infection. Recently the team has been working on an early warning system for acute respiratory distress syndrome, a type of respiratory failure that can be caused by many diseases, including sepsis, pneumonia—and COVID.
“The tool is already monitoring patients across five community and academic hospitals,” says Suchi Saria, founder and CEO of Bayesian Health and director of the machine learning and healthcare lab at Johns Hopkins. “We’ll be making this tool available for hospitals nationally to use.”
In an email, Saria explains that her team has been researching the predictive markers of respiratory failure for nearly two years, using data that’s routinely collected during hospital stays. Their respiratory models can help physicians evaluate COVID patients, predicting which patients are likely to rapidly deteriorate in the coming hours and which ones won’t. They can also help with the most fraught decisions, such as “when to consider removing a patient from a ventilator, should there be a need to ration,” Saria says. Additionally, the models can forecast a hospital’s projected need for ventilators and other critical equipment.
Physicians on the front lines of the battle with COVID have been actively discussing guidelines that they could follow in the event of equipment shortages.
Daniel Burke, Bayesian Health’s director of critical care solutions, is one of those doctors on the front lines. A critical care pulmonologist who usually works in at the University of Pittsburgh Medical Center, he volunteered to help at New York City’s Bellevue Hospital during the peak of the crisis. During a typical shift, he watches over more than 20 critically ill patients on ventilators.
Burke says that in this crisis, doctors who aren’t expert pulmonologists have to make life-and-death decisions for patients. “I feel for all these doctors who have no experience, and frankly didn’t chose to go into the field and never wanted to make these decisions,” he says. “I’m convinced that machine-learning tools can help.”
He adds that COVID is a tricky disease. “One of the most striking aspects is the ambiguity around how and when patients’ lungs begin to fail,” Burke says. “As critical care doctors, we’re used to seeing respiratory failure in all presentations. But sometimes in COVID patients it’s subtle. Their numbers look bad but they seem comfortable, and then suddenly they begin to decompensate.”
Burke says that New York City doctors haven’t been forced to ration supplies and decide which patient should get a life-saving ventilator, as occurred in Italy.
He also says the Bayesian Health tool, and others like it, can prevent such dire situations from arising. The Bayesian tool offers predictions at various time horizons, and hospital management can use the longer-term predictions to manage case loads and resources.
“Let’s say you have a predicted surge of demand for ventilators in the next 24 hours,” Burke says. “You can call in reserve staff, you can try to acquire ventilators from strategic stockpiles, and you can also plan to send patients out to other hospitals.”
Everyone hopes that there won’t be future surges in caseloads that max out hospitals across the United States. But even if the current flood of patients slows to a steady stream, AI may be the key to keeping hospitals afloat.