Unlike other areas of day to day life, health care has been late to adopt artificial intelligence as a tool for improved efficiency, satisfaction, and outcomes. Yet three Toronto hospitals are now using machine learning to predict and prevent overcrowding in their emergency departments and test artificial intelligence for direct patient care.
For more than a year, the Hospital for Sick Children has used its machine learning algorithm to predict patient surges in its emergency waiting room. Programmed with three years of historical data, the prediction software runs continuously in the background of day-to-day operations, weighing variables that include patient numbers, available beds, the time of day, and day of the week.
“We used to be overwhelmed when our emergency department would exceed 200 patients per day,” says Dr. Tania Principi, a Sick Kids emergency physician and director of strategic operations. “By the time a surge hit, it was too late. Now we see over 250 patients a day. Our tipping point has changed, and machine learning is one of several tools we are using to be resilient.”
The algorithm gives the physician and nurse in charge of the emergency department a two-hour warning of a surge, which lets them bring in additional physicians and make more treatment spaces available by discharging or moving patients to inpatient wards.
The system also relieves clinicians of directing the traffic, says Dr. Jason Fischer, the division head of emergency medicine at Sick Kids.
“We were relying on our frontline staff to make judgments about how our resources were being used in the moment, and that was distracting them in their clinical work.”
The application has been met with enthusiasm, according to Fischer, and is one of several tools they are using for crowding, including capacity and demand staffing, increasing situational awareness, and standardizing protocols. He acknowledges these are very early days and states, “it gives us the taste of the potential. At this stage we are using AI for operations and for improving patient flow, but clearly the future is going see us use it in a more clinical way.”
Meanwhile, at North York General Hospital, artificial intelligence is being used to forecast crowding months in advance at one of Canada’s largest volume emergency departments.
“The Christmas holiday period is when we are most vulnerable,” says director of emergency medicine Dr. Paul Hannam. “If it overlaps with the flu season at all, then it puts a lot of strain on everyone. This fall, our report predicted the busiest day on record as December 27th, and we saw a record-breaking 380 patients. It was surprisingly accurate.”
Extra physicians, extra nurses, and extra resources from the geriatric care and allied health teams helped weather the storm.
And just knowing extra resources are coming helps with psychological resilience preparation for the team, says Hannam.
“For us to be able to quantify what is coming at us and then to tell the team ‘this is the week it is going to busy, this is the number of extra people we are going to see. We’ve got this. We’ve got these extra supports in place. This is going to last for three weeks. This is why we are here, and we are going to work as a team.’”
St. Michael’s Hospital of Unity Health in Toronto now has a staff of 20 artificial intelligence and data analytics scientists developing machine learning applications for many areas of the hospital.
Dr. Muhammad Mamdani, vice president of data sciences and health analytics, believes the key to the successful implementation of artificial intelligence in hospitals is involving frontline health care providers. “Our data scientists do not ask the questions. The doctors, the nurses, the management folks are the ones who ask the questions.”
According to Mamdani, the use of machine learning to anticipate emergency crowding was an obvious entry point into health care, as emergency department crowding is highly predictable. He adds, “it is up to us how we are going to act on these predictions to make it easier for our patients.”
The Unity Health data analytics team worked with frontline clinicians, like emergency physician Dr. Sam Vaillancourt, to develop prediction algorithms that also take into account factors outside of the hospital, including snowstorms and city events, like Raptors games.
Vaillancourt says the challenge of developing artificial intelligence is the data and “hospitals have huge stores of data that if put together are really powerful. If separate, they are useless, but there is potential benefit that is unrealized across health care.”
What started as a research study to predict patient volume based on weather has now become much more sophisticated with the creation of linkages between existing hospital data sources. The algorithm can now determine more than a week in advance the case mix of patients in the emergency department, their likelihood for needing admission, and the proportion requiring complex care.
Vaillancourt says, however, it is important not to be blinded by the technology.
“It is really important for clinicians to be involved. We also need to make sure that it benefits patients.”
There are limitations in machine learning applications, he adds, not least the current coronavirus outbreak.
“In terms of crowding and predicting, COVID-19 is a great example of something we cannot predict, as we do not have a prior reference point for it. Machine learning can only predict based on events that have happened in the past. When you have a new event, you can’t fit it into the model.”
Mamdani acknowledges the main constraints are around data. “Do we have the right data, the right quality data, the right amount of data to feed these algorithms to work well?”
Vaillancourt says direct patient care is not currently influenced by artificial intelligence in their emergency department. Yet he and Mamdani think this is rapidly changing and that most patients will encounter artificial intelligence in visits to St. Michael’s Hospital within five years.
Mamdani describes a project nearing its launch whereby patients admitted to general internal medicine floors are monitored by artificial intelligence to predict death or intensive care unit admission within 24 to 48 hours.
The algorithm uses medications, orders, labs, word patterns in text notes, and interactions with nurses to predict a critical deterioration. A text message is sent 24 to 48 hours prior to the predicted deterioration to the chief medical resident and the charge nurse. This prompts a bedside meeting to begin the detective work of how to prevent the patient’s death or a critical decline.
These machine learning applications in the emergency department and internal medicine units are just two examples of the twenty machine learning projects currently underway at Unity Health.
Artificial intelligence brings with it optimism and potential for resilience in strained health care systems. Frontline clinicians’ input and pragmatism are critical in the adoption of this technology for patient care.