A resident in the General Internal Medicine unit of Saint Michael’s Hospital in downtown Toronto received an alert late one night from an artificial-intelligence program: a patient was at a heightened risk of either requiring transfer to the Intensive Care Unit or dying within the next 48 hours.
The resident checked on the patient, but he seemed stable, so she continued to make her rounds until around 1 AM, when the resident was called by a nurse – the patient’s medical condition was deteriorating. The resident rushed to the patient, who risked no longer being able to breathe properly. Since she had already assessed him, she carried out a second, quicker assessment and promptly arranged his transfer to the ICU. The patient’s condition stabilized, an outcome that the resident later said was largely due to the way in which the program had accelerated his care.
The program, CHARTwatch, is a tool that uses artificial intelligence (AI) to continuously monitor the medical status of patients on Saint Michael’s General Internal Medicine unit. Since the program has been trained to recognize the physiological changes that often precede a grave deterioration in a patient’s health, the program can pick up on subtle early warning signs of impending emergencies that busy doctors and nurses don’t always detect or properly interpret. When these indicators of mounting medical distress become severe enough that the patient has an approximately 30 per cent chance or higher of needing to go to the ICU or dying during their stay in hospital, CHARTwatch sends an alert through multiple platforms to a team of medical personnel. A physician can then assess the patient and decide what to do next. The program not only shows a lot of promise as a likely life-saving innovation, but also exemplifies a broader dynamic emerging between doctors and increasingly sophisticated AI health-care technology: while many doctors have voiced concern about AI replacing them, CHARTwatch and other AI health-care tools in development suggest that the rise of AI in health care may actually draw out the irreplaceable strengths of human doctors and, ultimately, make health care more human.
CHARTwatch was deployed for clinical use on the General Internal Medicine unit in late August 2020, and the team of scientists and doctors who developed it are now analyzing data about its performance. The results, while preliminary and subject to further vetting, are impressive. The program’s predictions about patients being at a heightened risk of requiring ICU treatment or dying are at least 20 per cent more accurate than the predictions of doctors, says Muhammad Mamdani, one of the scientists behind CHARTwatch and vice president of Data Science and Advanced Analytics at Unity Health Toronto. What’s more, the program appears to have played a large role in reversing a grim trend: the COVID-19 pandemic caused a huge spike in the number of deaths within the General Internal Medicine unit compared to previous years: for example, during July to September 2020, the mortality rate was 37 per cent higher than the average mortality rate of the same three-month period of the preceding four years, from 2016 to 2019. But after CHARTwatch was rolled out in late August 2020, the mortality rate during the months of October to December actually decreased by 21 per cent compared to the historical average. In other words, after the tool was deployed, the mortality rate was far lower than usual – during a pandemic.
“We’re fairly confident that we’re actually seeing a pretty positive impact on saving lives,” says Mamdani.
CHARTwatch seems to be effective not only because alerting doctors that a patient is at a heightened risk of gravely deteriorating can prompt them to administer important treatments or alter treatment plans earlier; it can also lead physicians to reassess a patient more closely and then identify underlying ailments that are easy to overlook. The program has led doctors to diagnose gall-bladder infections, intestinal infections, and blood clots in the lungs, conditions that are difficult to diagnose because their symptoms are generic, like abdominal pain or shortness of breath. Amol Verma, a physician at Saint Michael’s Hospital and clinical lead on the team that developed CHARTwatch, says that “because of the CHARTwatch alert, the clinical teams … made diagnoses of potentially life-threatening illnesses.”
The way that doctors are using CHARTwatch could be a preview of one of the main ways in which doctors interact with AI when it becomes more integrated into health care: across Canada, scientists and physicians are developing AI models that are meant to predict the probable trajectory of a patient’s illness based on important data about their current medical status, thereby allowing doctors to tend to higher-risk patients sooner. Creating AI models that can make such predictions is “where a lot of the activity in machine learning is happening,” says Verma. “The idea is that it’s going to actually change the way people deliver care.”
“We’re actually seeing a pretty positive impact on saving lives.”
Predictive AI could change how physicians care for patients by helping doctors make decisions about what treatment is needed more efficiently and accurately, says Alexander Wong, a Canada Research Chair in Artificial Intelligence and Medical Imaging at the University of Waterloo. Doctors would no longer have to draw exclusively on their own broad, varied medical education and clinical experience to determine what path a patient’s illness will likely follow and, consequently, what treatments to provide; a doctor could also use a prediction made by an AI model that has been trained on tens of thousands of cases of that patient’s particular illness.
Since AI models are so specialized in performing a specific task, keeping a “human in the loop is extremely critical,” says Wong. Predictive AI programs are limited to forecasting something that is quite narrow in scope, but “when it comes to medical practice, a lot of things are interrelated.” Physicians can place an AI model’s predictions within a broader, more well-rounded understanding of a patient’s medical profile by discerning the connections between an AI model’s predictions and other relevant medical information specific to that patient; physicians can think more laterally and also draw on their wide-ranging expertise in a given field. AI is not “this all-encompassing Wizard of Oz,” says Wong, but rather serves to “guide the clinician to look at certain things when making their final clinical decision.”
The ability of physicians to interpret an AI model’s narrowly focused predictions in light of a patient’s particular medical situation could make the care that physicians provide more personal and intimate – especially if AI also manages to automate some of the more rote tasks that physicians must perform, as Mamdani expects it will. Doctors could spend more time developing deeper relationships with their patients to tease out novel information about what aspects of their life circumstances are contributing to their health problems. “We need to really appreciate how health and medicine is not just biomedical,” Mamdani says. “It’s very social.”
The ability to “look in somebody’s eyes and understand their pain” is not just essential to being a compassionate physician – it is also a clinical skill that leads doctors to provide care that is better because it is based on a deeper, more nuanced understanding of a patient’s life, says Mamdani. By enabling physicians to spend more time engaging with patients as complicated people, and not just complicated medical problems, “AI will make medicine more human, not less.”