A trio of presenters at HIMSS 2018 shared data from practical machine learning and AI-focused studies that demonstrate how the technologies could reduce costs and improve patients’ outcomes. 

These studies, conducted by teams at the Grady Health System, Kaiser Permanente, and the University of Pittsburgh Medical Center, primarily focused on flagging adult and pediatric patients whose EHRs indicated specific risks, including readmission to care or unscreened colorectal cancer. 

“These [AI-related] technologies are here to stay,” Dr. Srinivasan Suresh, chief medical information officer at the Children’s Hospital of Pittsburgh of UPMC, said during the conference’s Machine Learning & AI event. “They are very transformative. They’re also very disruptive, so you should be extremely cautious integrating these into a clinical workplace. A key point is the collaboration just between clinicians first, and then between care providers and data scientists and software engineers.”

Suresh first described machine learning technology designed to predict 30-day all cause readmissions across pediatric patients while gauging a number of variables, such as recorded medications, vitals, demographics and length of stay. After achieving a 79 percent rate of accuracy in identifying readmissions as ‘high-risk’ at discharge during a three-month silent run, Suresh and colleagues paired the predictions with an intervention for 106 patients in which they proactively reached out to the child’s caregiver. 

“Keeping in mind this is a very small [sample], … if we make three or four phone calls, the readmission rate drops down to around 10 percent, from a baseline of 16 percent,” he said. “This is a game changer. If you want to do cost data, the average cost — not charges or reimbursement but cost, is between $2,000 and $10,000 per patient.”

Suresh also highlighted another ongoing program, in which machine learning is being used to detect catastrophic events in the hospital’s cardiac intensive care unit, and a big data project designed to identify correlates of health across birth cohorts slated for next year.

Another presenter, Robin Frady, an executive director at Grady Health System, described her own team’s work incorporating AI into its mobile integrated health program in an effort to reduce hospital readmission. Applying this program to 1,720 patient encounters allowed the system to intervene and avoid 113 readmissions. With each admission saving an estimated $11,200, the team’s total return after implementation costs was more than $1.1 million, achieved in just under a year’s time.

“At the end of the day, because we were able to do some small-scale success with the patient at the center of what we were trying to achieve, not the tool, … we were able to get buy-in from our senior leaders and other parts of the organization,” Frady said. “What we learned is [you should] start with the problem … and then go in there and say ‘hey, this is something that can help you out.’”

A third presenter, Dr. Elizabeth Liles, an investigator at the Kaiser Permanente Center for Health Research, demonstrated how machine learning can parse EMR data for adult patients at increased risk of undiagnosed colorectal cancer. This study, which included 9,108 controls and 900 cancer cases among health system patients aged 40 to 89 years, frequently flagged high-risk patients’ colorectal tumors 180 to 360 days prior to usual clinical diagnosis.

“What’s required [to develop a similar program] is a comprehensive EMR, and it probably helps where it’s not completely taboo to order complete blood care for diagnostic workup or screening,” she said. “I’d say one of the strengths is that it just happens in the background. It’s not something that you have to actively reach out to patients about until they're flagged as high risk.