As an intensive care doctor at The University of Chicago Medicine, Michael Howell, MD, MPH, sees many patients who are critically ill. To save lives, he has to make swift decisions — a challenge when the clock is ticking and the patient’s record is packed with hundreds of text notes to sort through.

But what if there were a way to automatically distill those notes and records into a concise, easy-to-read summary for one patient or to reveal patterns that could help guide care improvements at the population level? Researchers from UChicago Medicine, an association member, are teaming with Google to use machine learning to find patterns in electronic health records (EHRs) and use those patterns to predict readmissions, complications, and other hospital-acquired conditions.  The work represents a collaboration between UChicago’s Center for Research Informatics and UChicago Medicine’s Center for Healthcare Delivery Science and Innovation.

“Prediction is a cornerstone of prevention, because you need to know what’s going to happen in order to prevent it,” says Howell, who also serves as chief quality officer and director of the Center for Healthcare Delivery Science and Innovation at UChicago Medicine.

UChicago Medicine has a long history of predictive analytics research. Howell works closely with Samuel Volchenboum, MD, PhD, MS, a pediatric oncologist who serves as director and associate chief research informatics officer for the Center for Research Informatics. The Center manages the Clinical Research Data Warehouse, a repository of medical data dating back to 2006 that serves as a key research resource. Currently, the health system uses predictive models to conduct a weekly “hospital census” to predict how busy the hospital will be. The health system also uses real-time analytics to generate a list of high-risk patients and alert intensive care unit nurses to perform proactive rounds.

But traditional predictive models have limitations, Howell said. While they work for discrete data, such as numbers or categories, they typically don’t work as well with narrative data in EHRs. That’s where Google comes in.

“One of the exciting things about the collaboration with Google is that some of the newer analytic techniques around machine learning . . . can deal with those types of data in really impressive ways,” Howell says.

Many data points in the EHR are unstructured, including notes from physicians, nurses, and social workers; X-rays; and photographs. Google’s machine learning capabilities can quantify some types of data, drawing relationships between words and images and allowing researchers to measure the statistical links between word clusters and predicted outcomes. Already, Google researchers have used machine learning to detect breast cancer metastases in lymph nodes and test for diabetic retinopathy.

“One of the things that we think is really important at the University of Chicago is that the best predictive model in the world — all it does is turn a light red or green. That’s it,” Howell says. “Turning a light red or green doesn’t affect the health of any patient in any health system anywhere. It’s tremendously important to link the lights — red or green — to what the doctors and nurses and other health care providers are actually going to do in response.”

For Howell, that response is to use the data to prevent harm to patients in the health care system.

“My eventual long-term hope for this is that we will be able to figure out how to build tools that summarize the important parts of the electronic health record for clinicians, in a way that helps them to provide better care for patients,” he says. He also hopes that hospital staff can use EHR data to name a specific diagnosis and to aid billing and coding.

Howell said UChicago Medicine will safeguard patient privacy by encrypting the data, removing unnecessary data, such as names and addresses, and using secure Google Cloud infrastructure that complies with Health Insurance Portability and Accountability Act privacy rules.

“I’m excited to get to work with one of the teams that’s really, really at the leading edge of a lot of the machine learning work,” Howell says.

Howell and Volchenboum speak about using the Google Cloud platform to handle medical data at the Cloud Next ’17 conference.