The University of Chicago Medicine and Google have joined forces to research how well historical data housed in EHRs can predict future medical events.
In recent years, researchers have proposed analyzing clinical information to identify which patient risk factors are associated with adverse outcomes. The analysis would inform early intervention for a range of issues, such as unplanned hospital readmissions, a problem that costs the U.S. as much as $17 billion each year.
The UChicago Medicine Center for Healthcare Delivery Science and Innovation was founded in 2016 to research healthcare operational challenges. Today, the center is leading the university's collaboration with Google.
Michael Howell, MD, the center's director and a practicing intensivist, says his experiences in the intensive care unit sparked his interest in patient risk factors and predictive models.
"I've always felt like the patients get to us too late," he explains. "That's why I've devoted a large portion of my career to predictive modeling. I think prediction is the cornerstone of prevention. Unless you can tell something is coming, it's very hard to avoid it."
Challenges in predictive analytics
Healthcare data is notoriously difficult to model. As a patient moves through the healthcare system, they might accumulate an "electronic footprint" with hundreds of data points, according to Dr. Howell. And while this abundance of information excites researchers, it poses its own set of challenges.
"I've been doing risk-factor predictive modeling in healthcare for more than 15 years, and one of the things that's always been really frustrating to me is we can't use the majority of the electronic health record for prediction," Dr. Howell says.
Traditional epidemiological and statistical tools — like logistic regression and multivariable modeling, for example — don't work well with datasets comprising thousands of dimensions, and traditional analytic tools can't create predictive algorithms from free-text physician notes or X-ray images.
The UChicago-Google partnership is out to squash these obstacles.
Through the collaborative project, researchers will use Google's machine learning tools to incorporate physician notes, medical images and other difficult-to-standardize data into predictive models. "The main focus of the research project is to see whether we can use the whole electronic health record, rather than just the discrete data, for prediction," Dr. Howell explains.
Streamlining data analysis with AI
UChicago Medicine is the latest institution to join Google's research on predictive models in healthcare, following UC San Francisco and Stanford University in California. Google's goal is to partner with "world-class medical researchers and bioinformaticians" to explore the role machine learning might play in improving patient outcomes, according to a Google blog post.
"Machine learning is mature enough to start accurately predicting medical events," Google researcher Katherine Chou wrote May 17, commenting on how the approach to artificial intelligence can help providers unlock the full potential of EHR data. "We believe clinical breakthroughs using machine learning will come only when the medical community and deep learning experts collaborate closely."
For Dr. Howell, applying machine learning to data analysis represents a fundamental shift in the field.
"It's amazingly gratifying to see how, in the course of my career, healthcare moved from a place where a fundamental problem was data scarcity to a place where a fundamental problem is the logistical and statistical challenges of dealing with a lot of data," he says. "It's exciting to see how machine learning can deal with the kinds of data that healthcare now generates routinely."
Next steps for UChicago Medicine
Dr. Howell's goal for the research project is to publish a peer-reviewed paper investigating how machine-learning techniques compare to traditional predictive models. "There's a huge amount of hype around predictive modeling today, and I think there's some real potential benefit there," he says. "But we have to prove that with all of the rigor you would expect of any clinical research."
The first step is to connect researchers with a large-scale clinical dataset, which will serve as the base for the predictive models. To access this type of standardized, de-identified patient data, Dr. Howell and Google are collaborating with the UChicago Center for Research Informatics, which provides computational research support to the school's biological sciences division.
The center's clinical data warehouse is of particular interest to the research collaborators, according to Sam Volchenboum, MD, PhD, director of the Center for Research Informatics. He describes the warehouse as a "research compendium" of UChicago medical data, which researchers use for clinical investigations.
"One of the reasons I think Google wanted to work with us, in particular, is we've taken a very rigorous approach to our data warehouse that is not necessarily the norm," he explains. "We've been able to take our clinical data and standardize it and clean it up in a way that makes it much more easy to analyze and to perform this type of machine learning."
After developing their predictive models, the researchers — like all teams working on data-driven efforts — will also have to validate their findings, to ensure algorithms based off one sample of patients are applicable to larger populations.
Dr. Volchenboum highlights the research collaborators must be both "cautious and optimistic."
"It's attractive to think we can try to understand what factors are likely to make a patient, say, more likely to have nausea and vomiting at home after chemotherapy," he explains. "But, at the same time, you have to be careful when you make those predictions, and realize that you're training over a limited set of data."