The coronavirus outbreak has rapidly accelerated the nation’s slow-moving effort to incorporate artificial intelligence into medical care, as hospitals grasp onto experimental technologies to relieve an unprecedented strain on their resources.
AI has become one of the first lines of defense in the pandemic. Hospitals are using it to help screen and triage patients and identify those most likely to develop severe symptoms. They’re scanning faces to check temperatures and harnessing fitness tracker data, to zero in on individual cases and potential clusters. They are also using AI to keep tabs on the virus in their own communities. They need to know who has the disease, who is likely to get it, and what supplies are going to run out tomorrow, two weeks from now, and further down the road.
Just weeks ago, some of those efforts might have stirred a privacy backlash. Other AI tools were months from deployment because clinicians were still studying their impacts on patients. But as Covid-19 has snowballed into a global crisis, health care’s normally methodical approach to new technology has been hijacked by demands that are plainly more pressing.
There’s a crucial caveat: It’s not clear if these AI tools are going to work. Many are based on drips of data, often from patients in China with severe disease. Those data might not be applicable to people in other places or with milder disease. Hospitals are testing models for Covid-19 care that were never intended to be used in such a scenario. Some AI systems could also be susceptible to overfitting, meaning that they’ve modeled their training data so well that they have trouble analyzing new data — which is coming in constantly as cases rise.
The uptake of new technologies is moving so fast that it’s hard to keep track of which AI tools are being deployed — and how they are affecting care and hospital operations. STAT has developed a comprehensive guide to that work, broken down by how the tools are being used.
This list focuses only on AI systems being used and developed to directly aid hospitals, clinicians, and patients. It doesn’t cover the flurry of efforts to use AI to identify drug and vaccine candidates, or to track and forecast the spread of the virus.
Screen people who might be infected
This is one of the earliest and most common uses of AI. Hospitals have deployed an array of automated tools to allow patients to check their symptoms and get advice on what precautions to take and whether to seek care.
Some health systems, including Cleveland Clinic and OSF HealthCare of Illinois, have customized their own chatbots, while others are relying on symptom checkers built in partnership with Microsoft or startups such as Boston-based Buoy Health. Apple has also released its own Covid-19 screening system, created after consultation with the White House Coronavirus Task Force and public health authorities.
Developers code knowledge into those tools to deliver recommendations to patients. While nearly all of them are built using the CDC’s guidelines, they vary widely in the questions they ask and the advice they deliver.
STAT reporters recently drilled eight different chatbots about the same set of symptoms. They produced confusing patchwork of responses. Some experts on AI have cautioned that these tools — while well-intentioned — are a poor substitute for a more detailed conversation with a clinician. And given the shifting knowledge-base surrounding Covid-19, these chatbots also require regular updates.
“If you don’t really know how good the tool is, it’s hard to understand if you’re actually helping or hurting from a public health perspective.”
Andrew Beam, artificial intelligence researcher
“If you don’t really know how good the tool is, it’s hard to understand if you’re actually helping or hurting from a public health perspective,” said Andrew Beam, an artificial intelligence researcher in the epidemiology department at Harvard T.H. Chan School of Public Health.
Identify high-risk patients so doctors can reach out proactively
Clover, a San Francisco-based health insurance startup, is using an algorithm to identify its patients most at risk of contracting Covid-19 so that it can reach out to them proactively about potential symptoms and concerns. The algorithm uses three main sources of data: an existing algorithm the company uses to flag people at risk of hospital readmission, patients’ scores on a frailty index, and information on whether a patient has an existing condition puts them at a higher risk of dying from Covid-19.
Screen frontline health care workers
AI could also be used to catch early symptoms of the illness in health care workers, who are at particularly high risk of contracting the virus. In San Francisco, researchers at the University of California are using wearable rings made by health tech company Oura to track health care workers’ vital signs for early indications of Covid-19. If those signs — including elevated heart rate and increased temperature — show up reliably on the rings, they could be fed into an algorithm that would give hospitals a heads-up about workers who need to be isolated or receive medical care.
Detect Covid-19 — and distinguish it from other respiratory illnesses
Covid-19 testing is currently done by taking a sample from a throat or nasal swab and then looking for tiny snippets of the genetic code of the virus. But given severe shortages of those tests in many parts of the country, some AI researchers believe that algorithms could be used as an alternative.
They’re using chest images, captured via X-rays or computed tomography (CT) scans, to build AI models. Some systems aim simply to recognize Covid-19; others aim to distinguish, say, a case of Covid-19-induced pneumonia from a case caused by other viruses or bacteria. However, those models rely on patients to be scanned with imaging equipment, which creates a contamination risk.
Other efforts to detect Covid-19 are sourcing training data in creative ways — including by collecting the sound of coughs. An effort called Cough for the Cure — led by a group of San Francisco-based researchers and engineers — is asking people who have tested either negative or positive for Covid-19 to upload audio samples of their cough. They’re trying to train a model to tell the difference, though it’s not clear yet that a Covid-19 cough has unique features.
Predict which Covid-19 patients will deteriorate
Among the most urgent questions facing hospitals right now: Which of their Covid-19 patients are going to get worse, and how quickly will that happen? Researchers are racing to develop and validate predictive models that can answer those questions as rapidly as possible.
The latest algorithm comes from researchers at NYU Grossman School of Medicine, Columbia University, and two hospitals in Wenzou, China. In an article published in a computer science journal on Monday, the researchers reported that they had developed a model to predict whether patients would go on to develop acute respiratory distress syndrome or ARDS, a potentially deadly accumulation of fluid in the lungs. The researchers trained their model using data from 53 Covid-19 patients who were admitted to the Wenzhou hospitals. They found that the model was between 70% and 80% accurate in predicting whether the patients developed ARDS.
At Stanford, researchers are trying to validate an off-the-shelf AI tool to see if it can help identify which hospitalized patients may soon need to be transferred to the ICU. The model, built by the electronic health records vendor Epic, analyzes patients’ data and assigns them a score based on how sick they are and how likely they are to need escalated care. Stanford researchers are trying to validate the model — which was trained on data from patients hospitalized for other conditions — in dozens of Covid-19 patients. If it works, Stanford plans to use it as a decision-support tool in its network of hospitals and clinics.
Similar efforts are underway around the globe. In a paper posted to a preprint server that has not yet been peer-reviewed, researchers in Wuhan, China, reported that they had built models to try to predict which patients with mild Covid-19 would ultimately deteriorate. They trained their algorithms using data from 133 patients who were admitted to a hospital in Wuhan at the height of its outbreak earlier this year. And in Israel, the country’s largest hospital has deployed an AI model developed by the Israeli company EarlySense, which aims to predict which Covid-19 patients may experience respiratory failure or sepsis within the next six to eight hours.
Track hospital beds and supplies
AI is also helping to answer pressing questions about when hospitals might run out of beds, ventilators, and other resources. Definitive Healthcare and Esri, which makes mapping and spatial analytics software, have built a tool that measures hospital bed capacity across the U.S. It tracks the location and number of licensed beds and intensive care (ICU) beds, and shows the average utilization rate.
Using a flu surge model created by the CDC, Qventus is working with health systems around the country to predict when they will reach their breaking point. It has published a data visualization tracking how several metrics will change from week to week, including the number of patients on ventilators and in ICUs.
Its current projection: At peak, there will be a shortage of 9,100 ICU beds and 115,000 beds used for routine care.
Keep tabs on patients outside the hospital
To focus in-person resources on the sickest patients, many hospitals are deploying AI-driven technologies designed to monitor patients with Covid-19 and chronic conditions that require careful management. Some of these tools simply track symptoms and vital signs, and make limited use of AI. But others are designed to pull out trends in data to predict when patients are heading toward a potential crisis.
Mayo Clinic and the University of Pittsburgh Medical Center are working with Eko, the maker of a digital stethoscope and mobile EKG technology whose products can flag dangerous heart rhythm abnormalities and symptoms of Covid-19. Mayo is also teaming up with another mobile EKG company, AliveCor, to identify patients at risk of a potentially deadly heart problem associated with the use of hydroxychloroquine, a drug being evaluated for use in Covid-19.
Many developers of remote monitoring tools are scrambling to deploy them after the Food and Drug Administration published a new policy indicating it will not object to minor modifications in the use or functionality of approved products during the outbreak. That covers products such as electronic thermometers, pulse oximeters, and products designed to monitor blood pressure and respiration.
Among them is Biofourmis, a Boston-based company that developed a wearable that uses AI to flag physiological changes associated with the infection. Its product is being used to monitor Covid-19 patients in Hong Kong and three hospitals in the U.S. Current Health, which makes a similar technology, said orders from hospitals jumped 50% in a five-day span after the coronavirus began to spread widely in the U.S.
Remotely detect high temperatures and block sick people from entering public places
Several companies are exploring the use of AI-powered temperature monitors to remotely detect people with fevers and block them from entering public spaces. Tampa General Hospital in Florida recently implemented a screening system that includes thermal-scanning face cameras made by Orlando, Fla.-based company Care.ai. The cameras look for fevers, sweating, and discoloration. In Singapore, the nation’s health tech agency recently partnered with a startup called KroniKare to pilot the use of a similar device at its headquarters and at St. Andrew’s Community Hospital.
Gauge responses to experimental treatments
As experimental therapies are increasingly tested in Covid-19 patients, monitoring how they’re faring on those drugs may be the next frontier for AI systems.
A model could be trained to analyze the lung scans of patients enrolled in drug studies and determine whether those images show potential signs of improvement. That could be helpful for researchers and clinicians desperate for signal on whether a treatment is working. It’s not clear yet, however, whether imaging is the most appropriate way to measure response to drugs that are being tried for the first time on patients.
This is part of a yearlong series of articles exploring the use of artificial intelligence in health care that is partly funded by a grant from the Commonwealth Fund.