AI spreads across hospitals, becomes an enterprise-wide solution
November 1, 2022
Once relegated to fragmented, one-off solutions, artificial intelligence (AI) in healthcare is now coming of age as an enterprise-wide strategy. All it took was a global pandemic to highlight the strategic role machine learning and algorithms can play in helping to reduce costs, optimize staffing and scheduling, improve bed management and provide enhanced care.
“The time is now for AI adoption in a healthcare setting,” said Tomi Poutanen, CEO of Signal 1, a start-up launched earlier this year to help integrate AI-driven insights into hospital workflows. “Hospitals don’t have to wait to get on the most modern electronic medical record system … it’s fully possible to make use of AI and the cloud environment to gain these benefits now.”
According to Deloitte Insights, AI is one of the biggest areas of investment as healthcare organizations look to improve efficiency of internal processes. The pandemic served as a catalyst, prompting hospitals to begin to adopt AI enterprise-wide, with 85 percent of 220 global healthcare executives surveyed saying they expect their AI investments to increase in the 2022-23 fiscal year.
“When healthcare organizations apply AI across the value chain, they can improve consumer health and well-being and support better outcomes while boosting organizational efficiency and reducing costs,” Deloitte Insights reported.
At Signal 1 – a collaboration between Unity Health Toronto, TD Bank Group and TD Bank Group’s Layer 6 AI division – the goal is to help healthcare systems deal with unprecedented challenges emerging from COVID-19 by making enterprise AI solutions attainable.
The start-up combines validated AI applications developed at St. Michael’s Hospital, part of Unity Health, with the power of Layer 6’s enterprise-ready AI engine, originally built for TD Bank. As a Layer 6 co-founder, Poutanen has wanted to work in healthcare for a long time and says he is confident the robust, secure and highly regulated environment built for the bank will transition nicely into a healthcare setting.
“Obviously the data is very different but the underlying engine we built is very effective in predicting future events that have come from data laid out over discrete time events,” he explained. Just as a bank looks at a number of data elements over time to determine the likelihood of a client going bankrupt, for example, a hospital looks at numerous interventions and data points throughout a patient’s hospital stay to understand their level of acuity and predict when their health is likely to deteriorate.
“You’re trying to predict what’s going to happen in the future. That’s what the core engine we built is very good at in the banking setting and it turns out, is very effective in a healthcare setting as well,” said Poutanen.
Since launching in April, Signal 1 has been busy customizing the Layer 6 platform for healthcare and moving applications already in use at St. Michael’s over to the new environment. One of the first to transition is ChartWatch, a real-time patient risk predictor that accurately predicts those hospital patients who are likely to deteriorate and those who are clinically stable. ChartWatch version 2.0 is now up and running as an enterprise IT service managed by Signal 1, said Poutanen.
Anonymized patient data is sent to a Microsoft Azure cloud environment in real-time where it is run through the Layer 6 machine learning algorithm to calculate an output – or prediction – which is then returned to the hospital either via email, page or alert on a dashboard for instance. Originally designed to improve patient outcomes and prevent deaths by alerting clinicians to intervene with life-saving measures, the application is also proving beneficial in helping to alleviate bed and staffing pressures as well.
Signal 1 clinical advisor Dr. Muhammad Mamdani, who also leads a dedicated applied AI program at St. Michael’s as Unity Health vice-president of Data Science and Advanced Analytics, calls it a “surprise use” that was identified by clinicians.
“Nurses were a little apprehensive at first when we were deploying ChartWatch because they felt it would create a lot more work for them, but within a few weeks they were very creative and said, ‘You know, this actually helps us a lot,’” he said. What they discovered is that the ability to accurately identify high-risk patients made it easier to assign nurses more equitably, helping to reduce stress and burnout among staff.
Another spin-off use is to that ChartWatch can help with bed management by identifying low-risk patients who will soon be ready for discharge, allowing staff to start the process early and free up needed beds sooner.
Mamdani’s team of 27 – one of the first of its kind in Canada – has developed more than 40 AI applications to date.
Some of the more enterprise-level tools include:
- one that uses historical hospital data along with publicly available information such as weather or city planning events to predict expected ER volumes three days in advance;
- an ER staff scheduler that saves four hours of manual work per day by automatically setting nursing assignments according to staffing rules;
- an application called RUSHH that identifies hospital patients at risk for hypoglycemia so nurse practitioners know who to treat before it becomes problematic; and
- a predictor for the hemodialysis unit that identifies high risk patients so more intensive care can be provided, reducing the likelihood that they will return to hospital.
The need for a company like Signal 1 was born out of necessity, he explained. During the pandemic, several Ontario hospitals expressed interest in using algorithms developed by the St. Michael’s team and though they willingly shared their code, at least five of those hospitals didn’t have the expertise required to run the models in-house.
“We ended up being their help desk for IT support and it wasn’t sustainable,” said Dr. Mamdani. “To do this in a financially responsible manner, we had to go the start-up route.”
Signal 1 – so named because doctors look for ‘signals’ in patient data to treat their patients and the company wants to be the first signal they consider – is using an AI-as-a-service model to help hospitals deploy clinical AI at scale.
The first step is a data sharing agreement, enabling Signal 1 to access unique hospital data through APIs. The company then finetunes ChartWatch to reflect the practices and workflow for each hospital before delivering insights specific to their environment.
The delivery model solves one of the biggest challenges facing hospitals as they look to benefit from enterprise AI: the need for robust machine learning operations or MLOps. Signal 1 provides the backend infrastructure and ongoing monitoring to guard the integrity of models, and hospitals need only be concerned with providing data, explained Dr. Mamdani.
“What we can say to the hospital is we’re not asking you to do anything. Just feed us the data as dirty and messy as it is, we’ll feed it through our AI engine and it will clean it up, analyse everything and spit the result out in close to real-time,” explained Dr. Mamdani.
Signal 1 is currently working with four other Ontario hospitals as it brings ChartWatch 2.0 to market. The goal is to distribute the work of Dr. Mamdani’s team, which is already having a positive impact at St. Michael’s. Their next project will be a clinical AI tool to optimize emergency room triage.
“Signal 1 is not looking to solve niche point solutions,” said Poutanen. “We’re looking to solve enterprise-wide data AI problems for a hospital and ChartWatch 2.0 is a perfect example of that.”
Another start-up company working to create enterprise AI solutions tailored specifically to the healthcare industry is Predictive Health Solutions (PHS), a joint venture between SAS reseller Pinnacle Solutions Inc. and the Center for Discovery, Innovation and Development (CDID) at RWJBarnabas Health’s Children’s Specialized Hospital in New Brunswick, N.J.
Launched in February of this year, PHS is laser focused on solving the problem of patient no-shows, a reality that is estimated to cost the U.S. healthcare industry $150 billion dollars annually.
The idea came about as both partners were exploring ways to apply analytics to create positive change. Pinnacle Solutions brings the technology expertise to PHS and CDID offers the understanding about people and processes.
“We’ve been working on the solution and talking about it with customers, but it wasn’t until we met Children’s Specialized Hospital that we realized what we were missing,” said Pinnacle director, U.S. Sales and Alliances, Elizabeth Stack.
Expected to be available to hospitals and health systems as a software-as-a-service (SaaS), based on the number of appointments managed per month, the Patient No-Show Predictor tool is supported by the SAS enterprise-grade analytics platform. Similar to what Signal 1 is doing, the approach means AI models will be robustly managed at the backend end, including the ability to auto-tune and dynamically adjust as data parameters change, and the complexity of customizing models for each individual hospital will be handled by PHS.
“Custom developing and coding a solution to one health system’s data and population doesn’t easily translate to another health system,” said SAS strategic advisor Rich Kenny. “The underappreciation for this complexity is what has stalled many other AI-related efforts in the past.”
All of the advanced math fueling the Patient No-Show Predictor – which makes predictions based on a multitude of factors as to why a patient may miss their appointment, including transportation, weather, past history, or simply forgetting – is hidden from users who are presented with an easy-to-digest visual dashboard. Each patient is scored on a scale from one to 100, with one indicating they’re not going to show and 100 that they will, and results are colour-coded red, yellow and green respectively.
Not only can schedulers see the likelihood of a patient showing up on the morning of the appointment, but they can also use the tool at the time of scheduling.
“We can see what their probability is, and what their history is of not showing up, and then if we see they are a high no-show risk … we can start to understand from the dashboard reasons why they may not show up and offer preventive measures right then,” explained CDID director Victoria Gregorio. “It highlights a scope of where the problem areas are.”
For example, if patients living within the same zip code are showing a high percentage of missed appointments, the hospital might partner with a service like Uber Health to offer rides at the time of booking. CDID is also starting to screen a subset of outpatients to determine a correlation between no-shows and social determinants of health so that broader strategies can be implemented to prevent no-shows.
“We’re working very hard to make sure the average front line worker, who’s touching patients and really has the power to make decisions, doesn’t need to understand the analytics or the math, they just need to see the information,” said Stack. “Our dashboard really puts the power of all of that into the hands of people who ordinarily might be frightened by statistics.”
Following a positive reception at Children’s Specialized Hospital, PHS is now working to improve the tool and is in conversation with prospective customers. Currently, the AI model receives a steady stream of data from the hospital and “is getting smarter literally every day,” said Gregorio, as hospital staff continue to use the analytics to identify actions that can be quickly taken to prevent no-shows, such as identifying patients who regularly miss Friday appointments and scheduling them on another day instead.
“If you have, over the course of a month, 100 appointments that went unused because someone didn’t show, that’s 100 appointments that someone else needed and didn’t have access to,” said Stack. “So this is not just about recovering the revenue, which is really important. It’s about filling those spots for patients in need.”
“I’m on a wait list for a doctor I’d love to see tomorrow but I can’t get in for three months,” added Gregorio. “We know every organization has between a 15 to 30 percent no-show rate. Wow! What that could do to serve other folks who really need to see a doctor in a shorter period of time.”