Machine Learning in Healthcare Takes Another Step

Get ready for the next wave of predictive analytics, capable of identifying future admissions and health plan disenrollments.

Machine learning is not new to healthcare, and we have IBM's Watson technology to thank for that.

Until recently, many of the machine learning applications talked about for healthcare had been used to teach computing systems enough to be able to suggest a diagnosis on a specific disease.

IBM took things further. It essentially sent Watson to medical school.

IBM had Watson ingest large amounts of medical literature to learn everything physicians are taught about patients' conditions, and then taught it to make diagnoses.

This is how Watson won at Jeopardy.

But a Harvard professor who leads a startup supplying machine learning technology to Senior Whole Health, a Medicaid managed care organization active in New York state and Massachusetts, says that machine learning will eventually power all technologies we know today as predictive analytics and population health.

Leonard D'Avolio is that professor, and his background in healthcare makes him someone to watch on this front. His startup, Cyft, specializes in creating proactive care models with all available data from EHRs, unstructured notes, pharmacy info, and more to identify and better treat patients who will be soon experiencing some kind of trauma or risk of readmission.

D'Avolio's background includes collaboration with Atul Gawande, MD, at Ariadne Labs, an innovation lab startup where Gawande serves as executive director.

"I came up as a researcher and so I knew from trying to solve medical data problems that more than 50% of what is considered clinically relevant is unstructured free text in the medical record," D'Avolio says.

Highly Inaccurate Claims Data
"Claims data can be, depending on the disease, up to 80% inaccurate, and yet when we're doing analytics in healthcare, we are relying on rules and traditional statistics, all of which have at their base assumption the idea that the data will all be in one place, well-structured, and reliable."

D'Avolio says 20 years of research shows that machine learning and natural language processing (NLP) are capable of making sense of healthcare data that traditional business intelligence analytics technology cannot.

But after publishing more than 20 papers on the topic, speaking publicly, and releasing some machine learning and NLP technology as open source for healthcare to use, D'Avolio was disappointed by the results.

Only 300 institutions downloaded it and used it as a teaching tool. "It never jumped the fence and really made a difference in the care of patients."

So D'Avolio found investors and formed Cyft, which commercialized this work for organizations such as Senior Whole Health.

"One of the things we're doing right now is not just identifying folks that are going to be readmitted, which is sort of where predictive analytics stops in most of health care today. They've got to figure out which people in my population are most likely to end up in the emergency department in the near future, and not just anyone, because again, the high risk pregnancy is very different than the geriatric patient."

Cyft is focusing on the top five condition classes faced by organizations such as this—heart disease and diabetes are two—to try to target preventable admissions. It does this by scrutinizing nontraditional data such as nurses' impressions on notes, call center transcripts, Medicaid surveys on activities of daily living, and Likert scale functional status data.

Start with the Outcome
Cyft predictions get funneled to the appropriate person who specializes in that area, and who can now justify the cost of a $150 home visit. The results of such decisions increasingly can be shown to be more cost effective than an emergency room admission, which can cost more than $12,000 a day, D'Avolio says.

The same technology can even predict which patients are most likely to disenroll from a care plan, D'Avolio says.

"What makes machine learning so different is that you start with the outcome," he says. "You're taking a group of patients that were readmitted and a group of patients that looks a lot like them except they were not readmitted, and you're letting the math learn what makes group A different than group B."

While recognizing such patterns can be done by people, computers prove to be superior in recognizing more patterns faster than humans can.

So get ready for the next wave of analytics, one that may not rely so much on those analysts as today's analytics solutions. As companies such as Cyft gain traction, the economics, as usual, favor ever more automation.