To have real value in healthcare, machine learning must be actionable. Too often, IT decision makers don’t take a step back and ask if ML makes for a good business model, said Vikas Chowdhry, chief analytics and information officer at Parkland Center for Clinical Innovation.

Sometimes, big ideas are better left off the table.

"Oftentimes, people forget this," said Chowdhry, speaking this past week during the HIMSS Machine Learning and AI for Healthcare conference in Boston. "We think every problem can be solved with machine learning. That’s not the case."

For those areas where AI is appropriate, however, there are steps to take toward designing an effective approach, he said.

It may sound obvious, but defining the problem is a must-have.

If the problem is high utilization of the emergency room, for instance, perhaps giving patients appointment and prescription refill reminders could become part of the solution. Then the idea is to innovate fiercely and set up a prototype, Chowdhry said: Gather data, create cohorts, test to refine the model and create a chart of predictive performance of defined high risk.

Parkland, for instance, set up an asthma model for Medicaid patients under the age of 18 and used a control group for comparison.

During the comparison period, Parkland recorded a total 5% drop in ER visits. But for the intervention group, it realized a 31% drop in emergency room visits. This equates to about $18 million in cost savings over three years, Chowdhry said.

Through the model, Parkland set up a program to have two interns working on indoor air quality monitors in homes and classrooms. Over time, it has applied the methodology to other programs.

"Delivering value in healthcare is 90% care redesign and the other half is technology," Chowdhry said.

Understand the ROI of problem solving, be ready to change

Len D’Avolio, CEO of Cyft and an assistant professor, Harvard Medical School and Brigham and Women’s Hospital, outlined some lessons learned from his perspective – starting with demystifying the process.

When looking for potential applications where AI and ML might help, D'Avolio suggested healthcare organizations "fall in love with the problem, not the solution."

And once a problem area has been identified, change comes next – and there’s no improvement without it, he said. While change is hard, however, it can also be incremental.

It's also critical for healthcare organizations to understand the potential ROI that can be gained from solving those problems, he said. And they should plan, from the start, for how machine learning will be integrated into clinical and operational workflows.

Twitter: @SusanJMorse
Email the writer: susan.morse@himssmedia.com