The goal of population health is to use data to identify those who will benefit from intervention sooner, typically in an effort to prevent unnecessary hospital admissions. Machine learning introduces the potential of moving population health away from one-size-fits-all risk scores and toward matching individuals to specific interventions.
The combination of the two has enormous potential. However, many of the factors that will determine success or failure have nothing to do with technology and should be considered before investing in machine learning or population health.
Is there enough incentive?
Population health software, with or without machine learning, only produces suggestions. Getting a team to take action, particularly if that action is different, is one of the hardest things to do in healthcare. You will not succeed without executive support. Executives will not support you without significant incentive to do so.
Here's an easy surrogate for whether there is enough of that incentive: whether those executives’ jobs are in jeopardy if too many people go to the hospital. If not, the likelihood that an investment will lead to measurable improvement is minimal.
If you’ve been ordered to "do" population health, your best bet is to install a low cost risk score or have your team write a query to identify the oldest sickest people with the most readmissions. Either will return the same results more or less and your team of care managers are used to ignoring said results without rocking the boat. If there is sufficient incentive, read on.
Is the goal really the identification of the highest utilizers?
Henry Ford is credited with saying, "If I asked people what they wanted, they would have said faster horses." It’s human nature to try to apply a new technology in an old way.
Economists have named this the IT Productivity Paradox and have studied the cost of applying new technical capabilities in old ways. There are signs that healthcare organizations are unknowingly walking this plank.
For decades, risk scores were designed to identify the costliest patients with little consideration of the types of costs, the diseases they suffer, whether or not those costs are preventable, etc.
As a result, according to a systematic review of 30 risk stratification algorithms appearing in the Journal of the American Medical Association, "most current readmission risk prediction models that were designed for either comparative or clinical purposes perform poorly." A recently published study in Science also showed that prioritizing based on cost discriminates against people of color. Applying more data and better math to solve the problem in the old way is an expensive way to propagate existing shortcomings.
The opportunity now made possible is the ability to match individuals to interventions. Patients with serious mental illness that are most likely to have an inpatient psychiatric admission are very different than those with serious illnesses that might benefit from home-based palliative care. Clinicians wouldn’t treat them the same, neither should our approach to prioritization.
However, you will need to design for this and clinical teams should be prepared for the repercussions. Patients identified with rising risk (as opposed to peak utilization) will not seem as sick.
Clinical teams trained to triage may feel like they’re not doing their jobs if the patients aren’t as obviously acute. It’s important to discuss these repercussions and prepare in advance of the introduction of new technology.
Does your intervention work before introducing new technology?
Using technology to send more of the right people into a program that doesn’t have an impact only adds to the cost of an already failing program. Surprisingly, very few programs have ever measured the impact of their interventions.
Those that have, often rely on measuring patients before and after they enter into care management programs which is misleading and biased on many levels.
If you are not confident that the existing program makes a difference, invest in measuring and improving the existing program’s performance before investing additional resources. A good read on the pros and cons of different approaches to measuring impact is here.
Starting with a program of measurement can create a culture of measurement, improvement, and accountability - a great foundation for a pop health effort. Involving the clinical team in the definition of measures that matter will go a long way.
Another important consideration is whether your intervention is costly to deliver. The more costly it is to steer resources toward the wrong people, the more likely your program is to benefit from smarter prioritization.
For both reasons above – if your program is entirely telephonic and targets older people with chronic complex diseases, you may want to invest in program design and measurement before investing in stratification technology.
If you’ve made it this far...
You’re in great shape, and your odds of success are exponentially higher. You’re also better informed, as you and the team shift focus to decisions such as whether to build versus partner, what unique data you collect that can be used to your advantage and how you’ll measure algorithm and program performance.