Why analytics alone won’t bend healthcare’s cost curve
In a world of personalized experiences, from our Facebook feeds to our shopping recommendations, technology still hasn’t perfected personalized healthcare. Increasingly precise analytics, paired with domain experts, offer the best chance to make exponential progress in improving care and bending cost growth.
Patients have tried to take healthcare into their own hands with on-demand doctor appointments, fitness trackers and genetic profiling, but beyond these consumer developments, how are doctors and hospitals using technology and data to improve healthcare?
At a time when healthcare costs are in flux and just one percent of the population accounts for over one-fifth of total US healthcare expenditures, it's time we use modern analytic techniques to get as personalized as possible when it comes to patient care.
Finding specific groups that have high incidences of expensive diseases but where the cost of early intervention is low is one way to make a tangible change in healthcare. To do this, hospitals need advanced analytics with augmented intelligence—that is, technology that lets a computer do the initial work to find patterns that a human being would not see. This takes advantage of modern computing power, because while a human can only think of a few questions or hypotheses, a computer can ask millions of questions, covering many angles of research. Then, experts can augment those findings with domain knowledge to determine if a discovered pattern is worth acting upon.
To cause people to act on patterns, we have to get specific, and analytics enable us to find surprising findings that appear counterintuitive and, as a result, we wouldn’t otherwise discover.
For example, statistics show teens are four times more likely than adults to get into crashes or near-crashes when texting and driving. For a specific teen, the probability may seem too small to bother. But if we could show (hypothetically) that 18- to 20-year-old females with less than six months of driving experience, who use Phone A and drive Car B, have a 50 percent probability of crashing, then maybe that much more specific group of teens would pay more attention to the very real risk that pertains to them.
When it comes to our health, there are a multitude of unique characteristics that have an influence—such as diagnoses, treatments, age, gender, race, pre-existing conditions, current medicine use and more. Traditional data analysis led by humans can see broad trends, but these are difficult to act on widely because they seem so generic. However, with the help of advanced analytics and augmented intelligence, doctors are better equipped to find patterns specific to small groups of people so that they can have a better chance of encouraging action and change.
For example, McKinsey & Co. used advanced analytics in 2013 to analyze healthcare cost increases for 30 million patients across a million variable combinations at StrataRx Conference. Among the many findings: half of 18- to 35-year-old females with diabetic ketoacidosis were readmitted to the hospital. For that specific population of patients, this was not an abstract generic concern, but an immediate and imminent possibility.
For the female patients in the study, readmissions were primarily a result of noncompliance with insulin regimens. The physicians who McKinsey briefed on this pattern were initially surprised, because young women are typically better at taking their medications. However, doctors soon realized that they had indeed seen such cases, but had not recognized it as an actionable pattern. They eventually determined that for young women, not taking insulin might be an unorthodox weight loss strategy.
The key here was that medical researchers who knew the biological drivers of diabetes had not thought about the psychological behavior patterns of young women and had thus never thought to ask this question. McKinsey researchers themselves said they might have taken seven or eight months to manually test 250 variables, while the automated analysis evaluated millions of questions in just hours.
It’s important to realize that machine learning could not completely solve the problem on its own. It detected the specific pattern, but human experts had to use their years of domain knowledge to translate the patterns into the fact that young women were dieting by not taking insulin. They also had to figure out how to best counsel this group of patients so that better health outcomes were actually delivered.
In this case—and probably, in most within healthcare—the human and the machine working together could achieve what neither could solve alone. Once the pattern is detected, the adjustment to care is preventive, safer and less costly. Achieving all those benefits would be impossible if the pattern was not detected in the first place.
This is how clinicians and analysts will change healthcare—not through generalities, but through finding and intervening one at a time with extremely specific populations of patients.