In 1871, when the Sisters of Mercy opened a 25-bed infirmary for women and children in St. Louis, they got to know each patient intimately, says Joe Kelly, Mercy’s executive vice president in the Office of Transformation. Kelly is talking about a recent Mercy initiative that he believes will continue this legacy, but it can only be described as futuristic: Together with the Mayo Clinic and three recently announced health care organizations in Israel, Brazil, and Canada, Mercy has entered into a partnership in which it will use artificial intelligence and de-identified patient data to try to improve outcomes for sick patients. Between Mercy and Mayo Clinic, Mercy will have access to secure data on 25 million patients and 900 million “transactions”—that’s any interaction a patient has had with a health care provider. If you stuffed all of that information into filing cabinets, the storage drawers would wrap around the world three times, says Kelly; Mercy believes it’s one of the largest longitudinal clinical data sets globally. But what does it mean, office supply imagery aside? Instead of having to consult medical journals or hunt for information on their own, Mercy physicians presented with, say, a patient suffering from a rare condition will be able to consult this de-identified patient information, find maybe 10,000 other patients worldwide with the same disorder and genetic profile, and see which drugs or treatments did or didn’t work for them.
“This is going to enable us to get to the point where we know how to best serve every patient by knowing what’s worked and what hasn’t worked based upon genetic predisposition, lifestyle choices, and ethnicity,” Kelly says.
Your guide to a healthier, happier you
Sign up for the St. Louis Wellness newsletter and get practical tips for a balanced, healthy life in St. Louis.
But if you read the phrase “artificial intelligence” and felt the hairs on the back of your neck stand up, you’re not alone. As questions surround the popular—and unregulated—AI bots, such as ChatGPT, the medical community is exploring how it can use different AI tools to improve outcomes for patients while guarding their data, privacy, and trust that a human will call the shots when it comes to their care.
Rising Star
THE STAR PROGRAM AIMS TO FOSTER FUTURE LEADERS IN HIV RESEARCH.
A new program is paving the way for better strategies to prevent HIV transmission in young people—and students are leading the charge. Dr. Juliet Iwelunmor, a professor at Saint Louis University’s College for Public Health and Social Justice, created the STAR (Stimulating Training and Access to HIV Research Experiences) program in the hope of empowering young people to become leaders in HIV research. The program is funded by a five-year, $1.76 million grant from the National Institute of Allergy and Infectious Diseases. STAR will use such initiatives as crowdsourcing to search for diverse youth engagement in HIV prevention research. Each year, 10 students will build and implement their own crowdsourced research project, receive research and career mentorship, and participate in experiential learning activities. The goal: to develop innovative strategies to advance HIV interventions for underrepresented minority young adults—and, in the process, provide a space for students to grow and become leaders.
The first thing you need to understand about AI is that it’s only math, says Dr. Henry Randall, a professor of surgery in the Department of Surgery, Division of Abdominal Transplant Surgery, at Saint Louis University School of Medicine. You feed a computer an equation, you feed it some data, and it gives you a response to the question you’re asking. But it’s the human using the algorithm who must decide whether the data is significant, whether the equation needs to be refined, and—a biggie—whether the data is biased.
Randall’s lab is partnering with Missouri S&T to investigate how AI can help improve the distribution of donor organs, specifically kidneys. Last year, they were awarded a $1.8 million grant from the National Science Foundation to support the work. Nationally, more people are signing up to become organ donors, and Randall says they are transplanting more organs each year. But organ transplants are dependent on time dependent—kidneys can be shipped across the country with the use of cold storage or a special pump and transplanted up to 40 hours after they’re removed from their donor. A heart? Four hours. Randall’s lab wants to improve how quickly they can execute these transplants. To do that, his team will use AI to assess how humans make decisions for patients and then will try to replicate that in a computer model. Randall and the S&T team are feeding in the specifics of previous transplants into the model—data such as patient medical history and donor information—and comparing the decision that the computer makes to the decision that a human made.
Humans are still necessary, though. “We don’t want that inferior decision-making, and we need to be able to recognize when the computer gives us inferior information,” he says. “We, as health care providers, have to make decisions about whether this data is fair. You have to make sure that it’s safe, that it doesn’t harm people, that it doesn’t disadvantage people,” he says.
Every single model has a bias, says Dr. Thomas Maddox, executive director of the Healthcare Innovation Lab at BJC HealthCare and Washington University School of Medicine. “There’s a limit to how much it can characterize a given population,” he says. “Because at the end of the day, you’ll never be able to build a model for the entire world.”
There might not necessarily be anything nefarious about it, but, Maddox says, “it does mean that you need to be clear-eyed about how accurate the model might be in your population.” Historically, people who participate in medical research have been “healthier, wealthier, and whiter,” Maddox says. It’s important for physicians to look at a model and ask where it was developed and then compare that with the patient in front of them.
BJC saw this firsthand when it used a model built by Epic Systems that used AI to detect bloodstream infection. Epic used three health care systems, which didn’t include BJC, to build the model. The model looked at factors that predicted infection and alerted a patient’s clinical team when those factors presented in a patient. The patient would then be given antibiotics. But when BJC tried the model, providers found that it wasn’t terribly accurate, partly because Barnes-Jewish sees more complicated cases than the systems used to build the model, Maddox says. “What we needed to do was to start with that baseline model and then use our data to train that model and say, ‘Here are the specifics of St. Louis patients being seen at BJC,’” he says.
BJC is working on a tougher prediction problem using an AI-driven model now: It’s trying to understand the risk of near-term death for people who have a chronic and incurable condition, such as end-stage heart failure, lung disease, and cancer, so that end-of-life talks can happen. The goal is to maximize a patient’s quality of life. If the risk of dying within the next month is high for a BJC patient with an end-stage condition, the team can reach out to the patient’s physician to make sure they’ve asked if the person is doing what they want to do, with the people that matter to them, in a place where they feel comfortable. If it’s important for a patient to be at home, then BJC can arrange hospice and other care services.
Maddox says only 16 percent of patients who are near death have this kind of conversation. “It’s really underutilized, and a lot of times it’s because we don’t have a great sense of how long somebody has to live,” he says. “That’s a really hard prediction problem. We haven’t solved it, but we think we are able to be more accurate if we integrate a huge variety of data… It’s that kind of problem where AI algorithms seem to be more powerful.”
An end-stage, incurable disease is a pretty obvious predictor of death, but that’s why it’s important to forecast more accurately when someone might pass away: because time is short. Plus, other details can help doctors understand a patient’s timeline. Medicines they’re taking, treatments that are no longer working, and certain vital signs are variables that the model can analyze to give doctors more accurate predictions. The model learns the constellation of factors most associated with impending death, and then doctors can take that information into consideration.
In a different department of Barnes-Jewish, Washington University doctors are using machine learning to predict adverse outcomes in patients during surgeries. Dr. Michael Avidan, Washington University Physicians’ head of the anesthesiology department, and his colleagues use an Anesthesiology Control Tower to monitor a surgical patient’s stats while their operation is happening. Using data from thousands of other patients, they’re able to employ machine learning to predict negative outcomes while the patient is still on the operating table.
“The clinicians working in the Anesthesiology Control Tower actually get readouts of which patients are at risk for things like death or kidney failure or respiratory failure, and also the risk changes over time,” Avidan says. “If somebody were requiring more and more oxygen over time during the course of a surgery, that could suggest they’re running into difficulties with their breathing and would be at increasing risk of respiratory failure.”
Avidan doesn’t want to give the impression that this replaces a clinician’s assessment during surgery. Surgeons strategize all the time during procedures. Maybe a patient lost a lot of blood, or their blood pressure steadily dropped, or they stopped making urine during a procedure. They might send someone to recover in the intensive care unit and change their path of care based on a number of events. But, Avidan says, “what we’re trying to do is to keep improving what we do and to supplement or complement our expertise with state-of-the-art algorithms, machine learning, and artificial intelligence.”
Organ transplants, blood infections, end-stage disease, and surgeries are all examples of scenarios where a health care system could use AI to help patients when they’re hospitalized. In a hospital setting, patient stats are gathered routinely and often, and it’s taken by trained professionals. But providers are already using AI to monitor patients at home, hoping to prevent hospital visits in the first place.
Mercy’s first machine-learning program was a program called vEngagement, which is still in place today. Ten years ago, the health care system identified the patients with multiple comorbidities or chronic conditions who might end up in a hospital’s emergency department. To try and anticipate when someone might get sick and to avoid an emergency situation, Mercy outfitted patients with tablets, pulse oximeters, blood pressure cuffs, and scales, to monitor them in their homes. Now, Mercy monitors 5,000 patients via daily check-ins with patient navigators.
“Patients are thrilled that they’re in the comfort of their own home,” Kelly says. “They feel safe, and they feel monitored. When we first started the program, we couldn’t have anticipated this, but our patients have told us that they feel like they’ve built a personal relationship with their care navigator.”
BJC has a similar program in which patients with heart failure who are healthy enough to be at home are outfitted with a Bluetooth-enabled blood pressure cuff and scale, as well as sensors that rest under the patient’s mattress. Data on the patient’s heart rate, breathing rate, sleep patterns, and weight are sent back to BJC for monitoring.
Sometimes people experiencing heart failure retain fluid, and this edema can build up in their legs or lungs. If a patient’s vitals start looking bad and BJC senses that they’re about to develop this edema, then a physician can intervene.
Buddy System
HOW A NEW VOLUNTEER PROGRAM AT SSM HEALTH BEHAVIORAL HEALTH AT DEPAUL HOSPITAL IS SUPPORTING LONG-STAY PATIENTS
At a holiday retreat with several colleagues, Stacie Estes explained the struggles of some long-stay patients: As a manager at SSM Health Behavioral Health at DePaul Hospital, she’d noticed post-pandemic that an increasing number of patients would remain at the hospital for longer periods of time because they didn’t have a safe place to go to after they were discharged. Often, they’d wait without family, friends, or a strong support system. One of Estes’ peers offered to make weekly 30-minute visits to help support patients. From that conversation, Estes and two colleagues developed The Buddy System, a volunteer initiative staffed by DePaul Hospital employees, to provide community for long-stay patients. Since beginning in January, the program has matched every eligible candidate with a buddy. “One buddy stopped me this week and said, ‘I’m so glad today happened—it was the first time I’ve gotten to genuinely talk with [my buddy] and see her smile,’” Estes says. “They had a really good time just letting down their walls. [The Buddy System] allows [long-stay patients] to build trust back with people and have a meaningful relationship that’s therapeutic.”
“We have a lot of data, and with that, you can start to build the models to say, ‘Hey, I think Joe’s getting into a little bit of trouble. Let’s give him a call,’” Maddox says. “We were finding that we would see abnormalities in data even before Joe felt bad. So we would sometimes call Joe and say, ‘Hey, things aren’t looking good.’ And he would go, ‘What do you mean? I feel fine.’ But there’s actually a risk in the next five days, [he’s] not going to feel fine.”
Is it…icky to be monitored like that in your own home? On the one hand, to be able to access treatment to prevent an adverse health event before it even happens can feel like magic. On the other, it also feels a little Big Brother.
“We talk about that a lot: How do we do this in a non-creepy way?” Maddox says. “I don’t want that, either. I’m not interested in having someone snooping over my shoulder.” Maddox points out that none of the models are used on patients who don’t want the option; it’s always the patient’s choice, and even if someone participates, they still have the say in whether to take an action that a doctor recommends based on the model. “We’re fully transparent,” Maddox says. “We tell them: ‘Here’s what we’re doing. These are the tools we have. In the case of this model, we need to have some sensors in the home to collect this information, but here’s what we do with it. Here’s how we protect that data.’ We need to have integrity and transparency.”
Physicians decide whether to act. “We give information to the physician and it’s their judgment—this is simply an assist,” Maddox says. “And if they’re seeing something that they think is different, or they don’t buy it, that’s their call. I think as long as we’re operating with transparency and preserving autonomy of the patient and the treating physician, that hopefully diminishes any creep factor.”
Health care systems are also feeling the obligation to guard patients’ data and protect their privacy. “It’s not our data,” Kelly of Mercy says. “It’s our patients’ data.” Mercy pays for a third party to certify the data’s de-identification—even if parts of a patient’s data fell into the wrong hands, it couldn’t be put back together and the person identified.
And because Mercy is a faith-based organization, the system believes it answers to a higher power. “We still have four active sisters on our board who are interested in making sure we’re doing the right thing for our patients,” Kelly says. “We also write an annual report to the Vatican. We really do believe that protecting patient privacy is the most important thing we can do.”
