Health systems and payers have begun making serious investments in machine learning, a branch of artificial intelligence that enables computer systems to learn from and make inferences about data sets without the benefit of explicit coding. When applied at scale ― it is no longer a matter of “if” – machine learning has the potential to deliver better care at the bedside by assisting clinicians in diagnosis and decision-making, improving quality and outcomes for populations across care networks, while optimizing scarce resources.
“There is a small but growing group of experts starting to focus on fundamentally new classes of machine learning and artificial intelligence that are specialized for healthcare,” said Zeeshan Syed, PhD, founder of Health at Scale Corporation and a clinical associate professor and director at Stanford Medicine. These experts are developing “a complementary set of machine learning and artificial intelligence technologies that are specialized for healthcare, using expertise spanning both healthcare and computational science.”
Machine learning’s full impact on diagnosis and care delivery may be still up the road. But providers such as St. Michael’s Hospital in Toronto (affiliated with the University of Toronto) and Geisinger Health in Danville, Pennsylvania, have launched ML initiatives that provide a glimpse of the things to come.
Geisinger Health: Three steps ahead
“We frame the clinical potential for machine learning in three steps,” said Elizabeth Clements, business architect with Geisinger Health’s Enterprise Architecture team. “First, you have the ability to gain new insights through sheer information capture. Second, you have the ability to automate predictions, and third, you have judgment generation. The first two are really where most of the development and science is today. We’re not quite to that judgement generation quite yet.”
To illustrate the first step, Geisinger Data Scientist Debdipto Misra points to the health system’s use of machine learning to analyze unstructured clinical notes and flag individual cases at risk for lung nodules, which can lead to cancer.
Given that the system creates 60,000 patient notes daily, there is no way humans could review all of them to identify the relatively few cases that describe insightful patient information for improving patient diagnostics. But, said Misra, “with a Big Data platform like Hadoop and natural language processing, it is possible to contextually parse and extract relevant data from the note.”
The algorithm identifies which patients are likely to benefit from follow-up by annotating 50,000 notes/hour, or about 1 million notes per day. The annotations are used to extract features, entities and medical events for further analysis and clinical review, eliminating the false positives and then focusing clinician contact on the remaining set for further consultations.
“This minimizes cost and improves outcomes because we are actually making contact and following up with patients at a much earlier stage,” Misra said, “before it becomes cancerous, more troublesome.”
Another Geisinger machine learning application illustrates the second level, automated prediction. Clements said hospitals in the system are using an inpatient bed-demand tool that can accurately predict census three days out.
The first iteration of the tool did not use machine learning, relying instead on traditional predictive analytics running Monte Carlo simulations. The resulting score was combined with OR scheduling data to generate future capacity predictions: red (overcapacity), yellow (at capacity) or green (below capacity).
When the system upgraded the tool with machine learning, Clements said it was able to ingest “nontraditional data, like whether it is a holiday, or a full moon ― things that you have tacit knowledge about within your facility but that you may not always think about when you are trying to predict something like that.”
The new data, as well as the increasingly refined analysis from the machine learning algorithm, has improved the accuracy of the tool by 40 percent of the mean absolute error of earlier iterations.
St. Michael’s Hospital: Future tense
St. Michael’s Hospital in Toronto also applies machine learning to generate predictions ― in one case, about emergency room capacity.
“We see an enormous variation day-to-day in the number of people walking through the doors,” said Jeremy Petch, PhD, project director for the hospital’s Quality and Performance Analytics Initiative. “It’s difficult to know when surges are coming, so when they hit, you may not have enough staff, while at other times your staff can be sitting around idle.”
St. Mike’s team of data scientists developed an advanced analytics tool to combine years of historical data with real-time weather data and upcoming city events (like marathons, concerts or hockey games) to help determine near-term ED patient volumes.
“They are now able to predict emergency room volumes a couple of days out in six-hour increments, with less than a 6 percent error rate,” Petch said. “This represents a huge improvement in our predictive power over our previous ability to respond to surges. And that, we hope, makes for a much better patient experience.”
Its team has developed several other applications that rely on artificial intelligence and machine learning, and while these are performing well in evaluation, it’s still too early to speak confidently on their impact on care and cost. “Implementation is where the rubber meets the road,” Petch said. “We’re midstream in our implementation and evaluation of these tools.”
Lessons from the leaders
Petch, Clements and Misra agree that healthcare providers would be wise to begin exploring opportunities for machine learning immediately if they haven’t started yet. Some key considerations:
1. Start small ― A Big Data platform is a significant investment. Clements suggests starting small to prove value. “When you think about machine learning, if you can’t invest that heavily in all the different platform elements, look for open source technologies. There are several out there, many are free to work with, so you can start to experiment and identify very specific use cases with your existing data sets in house, your EHR, etc. Then you demonstrate success and look for funding on a larger scale.”
2. Design a Big Data repository based on your goals ― St. Michael’s primary goal was speed. Its machine learning initiatives are aimed at providing rapid responses to critical questions. “For us, it wasn’t a difficult decision. We had a clear sense of what we wanted out of our analytics,” Petch said. “That meant going with a data warehouse where we put a lot of work upfront to structure and to clean our data to make it as analysis-ready as possible.” The alternative ― the data lake ― ingests disparate data from multiple sources without the upfront normalization, broadening its scope but requiring valuable time to cleanse the data prior to analysis.
3. Future-proof ― Pick a data platform that can grow as you scale up. Said Syed: “We’ve gone from needing significant resources for analytical projects on dedicated hardware that would be underused most of the time, insufficient when jobs really needed to be run, and outdated rapidly as Moore’s law continued to hold strong — to being able to run extremely powerful pipelines at a fraction of the cost using large amounts of computational capacity in bursts as needed on top of continuously updated and evolving infrastructure.” Whether run on premise or in the cloud, as Syed’s team does, look for a data platform that provides scalability and speed for machine learning projects.
“Whatever tools an organization decides to use, I would encourage them to think about this element of it: How are you going to set up your advanced analytics infrastructure so that you can do your analysis very quickly?” Petch said. “It’s a very narrow window. To get the full value of machine learning, you need to get your insights to decision-makers in a timely fashion.”