The technology and expertise in a hospital’s intensive care unit can be lifesaving. But health workers can’t always tell when a patient’s condition is about to worsen, and patients can arrive at the ICU late. This leads to longer hospital stays and poorer outcomes. But there are times when patients could have avoided the ICU altogether had their condition been recognized and treated earlier.
That’s why Muhammed Mamdani and his team created CHARTwatch. It’s an Artificial Intelligence-driven tool that looks at 100 different variables in a patient’s chart, including lab results and vital signs, and determines whether the patient is at a low, moderate or high risk of needing ICU care. Mamdani, vice president of Data Science & Advanced Analytics at Unity Health in Toronto, led dozens of people in developing the tool and then rolling it out at St. Michael’s Hospital last summer.
The program now monitors all patients in the hospital’s internal medicine unit. The team chose the intensive care unit because patients in this unit often have serious health issues involving more than one organ.
CHARTwatch updates the risk category for every patient each hour. When the tool flags a patient as having a high risk of going to the ICU, “that triggers clinical assessment, increased monitoring … and more caregiving to that patient,” says Mamdani. “We’re already seeing quite a few cases where the algorithms picked up problems the clinicians have missed.”
One of the keys in making CHARTwatch a successful learning health project is the involvement of front-line physicians and nurses from the outset.
“What usually happens in a research project is you develop the scientific tool, and then you figure out the implementation afterward,” says Amol Verma, an internal medicine doctor and scientist at Unity Health’s St. Michael’s Hospital. “But because we involved the people who would be using the tool from the outset, it really informed the development of the tool.”
For example, the development team proposed a model that would predict patient risk at any point of their hospital stay. The clinicians immediately pushed back: Without a time element, what use would the information be since the patient’s condition could deteriorate tomorrow or a week or two into the future? That prodded the developers to focus the tool on predicting which patients were likely to need critical care within the next 48 hours, allowing clinicians to run tests and adjust treatments right away.
The science and clinical teams were able to work together closely in part because in 2019, Unity Health made Mamdani’s research centre a formal part of the hospital, giving him access to data without having to go through drawn-out privacy and legal agreements each time. Unity Health leaders have also invested in data gathering and analysis. “We have a team of data engineers and developers whose job is to build data pipelines,” says Mamdani.
In a learning health system, interventions are evaluated and adjusted after rollout.
In CHARTwatch, artificial intelligence can then analyze this complex health-care data and give decision-makers access to that information in real time. With faster access to more comprehensive data, health workers and researchers can speed up the whole learning health system cycle of designing, implementing, evaluating and adjusting interventions.
In a learning health system, interventions are evaluated and adjusted after rollout. In CHARTwatch’s case, at first only doctors and nurses seeing patients on the internal medicine unit received the alerts. In the next phase, palliative care doctors were included in the notifications on patients with a high CHARTwatch risk, enabling them to have conversations with patients before they enter the whirlwind of the ICU and may not have the time to consider, for instance, if they want to be put on a ventilator.
Jonathan Ailon, a palliative care doctor at St. Michael’s, is one of the health workers who talks to patients with CHARTwatch alerts. Together, they discuss treatments that might be offered if they do end up in the ICU, and “if they accept such treatments, what is their expected quality of life going to be,” he explains. Rather than having conversations “at a time of crisis,” patients and families can take the time and space to consider what their wishes would be in a range of health scenarios. These are conversations that might not happen without the prediction tool. And if the patient doesn’t go to the ICU, they’ll be better prepared in the future to communicate their care wishes.
The tool has had its intended effect. A soon-to-be-published study involving 1,000 patients and 150 clinicians finds that the algorithm has been 20 per cent more accurate than clinicians in predicting which patients would go to the ICU.
Shirley Bell, a nurse educator at St. Michael’s, credits the success of the CHARTwatch implementation to its roll out, with advance notice, consultation and training for all the health workers involved.
In staff meetings and focus groups, Bell addressed nurses’ concerns about how the tool would affect their workflow. “At some point, people were OK, when is this starting?’ That’s how we knew that people were ready to adapt to this change,” says Bell. “The approach that we use to learning should be fun … as opposed to ‘I’m going to tell you about it this week, and then start next week.’ ”
Nurses are seeing the benefits of the tool. For one, they don’t have to spend time alerting doctors when a patient’s blood sugar, for instance, is too high. CHARTwatch does that for them.
The tool “gives people a shared language,” says Verma. “So you can say, this patient is CHARTwatch high risk, or CHARTwatch medium risk or CHARTwatch low risk. And everyone knows what that means.”
CHARTwatch is just one of the projects of the research centre. One soon-to-be-launched initiative uses historical data on emergency department visits, weather data, city planning data and “things like Raptors games or marathons on Lakeshore Boulevard” to predict how many patients will arrive at the ED over the next month, says Mamdani. The tool not only tells staff how many patients to expect but the ratio of high acuity to lower acuity cases and the number of patients they might see with mental health issues. That information helps department leaders know how many health providers they need to schedule and what the skills mix should be.
Mamdani stresses that AI isn’t the be-all-and-end-all. Some studies have shown that algorithms can cause harm, especially if not continually evaluated. But with learning health system approaches that involve wide input and evaluation at every stage, AI can be transformative.
That’s why Mamdani’s team only works on projects that are shaped by front-line providers from the start to ensure the model will be something that’s useful and improves care.
“That fundamentally changes the game because now it’s a product that’s useful to the end user,” says Mamdani. “And they feel like they own it.”