AI to predict if you will get sicker? A CT health care system has it; what it means for patient care
The Connecticut health system soon will use the artificial intelligence in more cases and is slowly implementing the overall technology into its care.
The artificial intelligence can be trained to predict whether you will get sicker.
Hartford HealthCare calls it “predicting patient deterioration.”
The Connecticut health system soon will use the artificial intelligence in more cases and is slowly implementing the overall technology into its care.
The process has been several years in the making.
Among many uses, the algorithm can predict clinical deterioration in patients, Hartford HealthCare doctors said. For now, it’s being studied and piloted, and the science of a ‘length of stay’ algorithm is being used, and the hope is to eventually have AI assist all patients with valuable and potential life-saving information, according to health care system doctors.
‘Patient safety is always first’
“Over the last seven years we have assembled many of the pieces of the puzzle so that we can double down on the unlocking (of) the transformative power of AI to our patients in a safe, trustworthy way,” said Dr. Barry Stein, Hartford HealthCare’s chief clinical innovation officer and leader of its Center for AI.
Hartford Hospital Associate Vice President of Medical Affairs Dr. Daniel Kombert was among the first working with a Massachusetts Institute of Technology team that developed an algorithm to improve patient care. Kombert said he and others collaborated with Dimmitris Bertsimas, associate dean of business analytics at MIT-Sloan.
On the ‘length of stay’ algorithm, Kombert said, its job is to help predict when a patient is ready to be discharged.
“It’s a national problem where patients are staying at hospitals longer than they should be. Longer stays increase the risks of falls and infections,” Kombert said.
The experts looked at the hospital’s historical data base and “the tool was rolled out slowly and we made sure it was safe and valid. Since it’s been implemented, we’ve dropped seven percent in hospital stays. It’s enormous savings. Our goals are typically to drop 1 to 2 percent per year,” Kombert said.
“When we rolled it out, we told the doctors this was just a tool. It’s a discharge consult and ultimately, it’s still up to the doctor, nurses and case manager to decide when to discharge,” he said. “But we use this AI in our daily communication and on all of our medial reports across the system and we’ve seen improvement across the board.”
Kombert said daily rounds are done and “doctors, case managers and nurses get together and communicate which patients are ready to go home.”
“What are the barriers to go home?,” he said. “If a patient will go home in three days or more they are marked as red. We’ve had cases where the AI has marked them green where they have a higher likelihood of going home in a day. That information is discussed on the rounds and they can be reassessed whether they can go home safely. We always have cases where the doctor says it’s not time to go home yet. There are others where they get the patients home in a timely manner.”
Dr. Melissa Boisjoli-Langlois, the assistant director of Hospital Medicine at Hartford Hospital has seen real examples of the effectiveness of the ‘length of stay’ tool.
In one case, a patient who came into the hospital with blood in his stool was admitted, she said.
“The GI team was consulted, and the patient had a stable blood count,” Boisjoli-Langlois said. “They didn’t need a blood transfusion. But an endoscopy and colonoscopy were needed.”
“On rounds that day, we received a green alert, and we reviewed it,” she said. “That triggered a collaborative discussion. After that discussion, the providers concluded that he could be transitioned out the next day if everything turned out OK with the procedures.”
The next morning the patient’s endoscopy and colonoscopy were moved earlier in the day and he was discharged later that day, a full day earlier than originally planned.
In another example, a female had signs of congestive heart failure with shortness of breath and swelling in her legs. A cardiologist was consulted, and she was treated with diuretics through an IV, Boisjoli-Langlois said.
The cardiologist ordered an echocardiogram.
“After a day, the patent was doing much better and the treatment was working,” Boisjoli-Langlois said. “She was not retaining as much fluid and wasn’t short of breath and the green light came up on the tracker. We had a discussion and expedited the echocardiogram, and it came up normal. We were able to let the patient go home later that evening a day before we had originally planned.”
Boisjoli-Langlois emphasized that physicians don’t go by what the tool says alone, but it starts an in-depth conversation to see if there are any barriers and to make sure everyone agrees.
“It’s been helpful,” Boisjoli said. “We are very busy in the hospital, and this allows us to standardize and work as a team more efficiently.”
Kombert said the artificial intelligence that predicts patient deterioration has not yet been rolled out. It is being piloted.
“The first thing is to identify an opportunity and we want to have the physicians understand earlier than they normally would when a patient is showing signs that may not be obvious,” Kombert said, of one aspect of the model. “Patient safety is always first and we looked for a high-quality algorithm that is trusted and solves the problem we are asking of it.”
“The next step was bringing the clinical people altogether to work with the MIT team to get exactly what they want,” Kombert said. “That doesn’t happen over a week. We are talking about weekly meetings with very busy people and getting their time and effort and help the MIT team to improve the algorithm.”
“This went on for over a year having one or two meetings a week, tweaking it to where we needed it to be,” Kombert added. “Now you have a tool that is answering the questions you need and is accurate. But you can’t just roll it out. It’s a process.”
Dr. Howard Haronian, the vice president and chief quality and innovation officer at Hartford HealthCare Heart & Vascular Institute and has also been involved in the AI research with MIT. Oxford University, Google Cloud are other partners, said the most important part was ensuring safety.
“We realized MIT is great at analyzing data but when it comes to patient care, they don’t work in the hospital,” Haronian said. “They don’t see patients. That’s why it’s on physicians to make sure this is 100 percent safe. We’ve been able to set up a world class governance. It’s best in class. Less than 20 percent of healthcare organizations currently have an organized process for AI governance. We are leading in that and I’m proud that.
“We offer the opportunity for myself of my family to have a heads up that I’m going to get sick the next day,” he said, on predicting patient progress.
“Wouldn’t it be nice with all of the craziness in the hospital and all of the patients, that…my doctor had an early warning,” Haronian said. By the time we release this to the general public of our health system, we’ve already piloted it, tested it and refined it. There’s no room for any safety mishap.”
Opportunities for new information
Hartford Hospital Director in Chief of Hospital Medicine Dr. Gagan Singh, who has been involved in the project since the developmental phase, said a pilot process of AI for detecting clinical deterioration has been put into place on a small unit at Hartford Hospital.
“We identified a small team that will get the signal,” Singh said. “The team will talk to the clinician, and they will talk about what needs to be done. Patients will get care in real time. Sometimes the clinicians will say I’m already aware of it and I’m on it. Sometimes there will be an opportunity where it will be new information and they can get in front of it.”
“That’s the power of this tool,” he said. “It will not only save us time, but it will save lives.
“There could be some kind of clinical deterioration and you caught it six hours in advance. You see what can happen and you are acting on it…But we are taking a small approach because our goal is patient safety,” Singh said. “We are being careful how we do it and that’s just being diligent.”
Kombert said said the process is moving forward methodically. Also among what the artificial intelligence can being used to detect are: COVID-19 related events, predicting secondary stroke events, nursing scheduling needs and efficients uses of operating rooms.
It also is designed to predict outcomes for transcatheter aortic valve surgeries and joint replacement surgeries, according to the health care system.
“We have the units and champions ready to move forward,” Kombert said. “In the next two weeks we expect to begin the slow rollout. To give our teams the opportunity for feedback. It’s over in about a month and then it’s a slow rollout. We will put it on three units and slowly add other units. We want to make sure there is value added and the workflow is improved as well as patient safety. The slow rollout has been effective in other areas.”
Jenifer Ash, an APRN in the Department of Medicine at Hartford Hospital, said quality and safety are paramount.
“Looking at the evidence, it shows that many patients show signs of deterioration in about 24 to 48 hours to that actual event,” Ash said. “For the team, the time of the intervention really matters and for us to equip the frontline team with the tools to be able to do that would improve patient outcomes.”
Stein said it is all a team effort.
“The philosophy of AI in healthcare – before we put it into clinical practice – we go through the same diligence if we were adding a new medication or device into a system,” Stein said. “There is a back and forth with the science and the clinicians and you are learning. It’s powerful. It’s almost a force multiplying effect.”