These days, every care management / value-based care organization has a risk score to help target interventions. Unfortunately, these risk scores often frustrate clinicians by directing them to people who cannot benefit from an intervention – either the person is not actually headed for trouble, or the clinician already knew about that person.

Why is that? It turns out most risk scores were never intended to do this job. And that has real repercussions.

The Beginning: Actuarial Risk Pricing

Risk Scores got their start in actuarial science, where they are widely used to help anticipate the cost of covering different individuals. These actuarial groups needed to be highly accurate on the population level so that health plans would price their plan competitively.

As long as the actuarial group uses risk scores to set appropriate premiums for members and the health plan remains competitive, everyone is happy. Though useful for population assessment, risk scores over price the healthiest people and underprice the sickest people[1]. They simply aren't designed to consider the risks of individuals, if the risk is impending, or whether the risk is preventable - really important foundations of care management decision making.

Risk Scores Move to Care Management

A health plan’s primary competitive advantage is their ability to keep their members healthy at lower costs. This falls primarily to the care management department.

The obvious place to start for the care managers is to reduce unnecessary utilization among the highest cost individuals. Logically, many of these teams do this by focusing their attention on members with the highest past Total Medical Expense (TME). In this way, TME is often considered a type of risk score. However, studies have shown that the costs within this group are likely to subside naturally as spikes in cost subside, members get healthier, leave the plan, or die. As one study puts it – for many patients who use large amounts of healthcare services, the need is "intense yet temporary" [2].

Another approach employed by many vendors and plans is the adoption of actuarial-base risk scores.

Unfortunately, care management has very different success criteria for predicting risk than actuarial groups.

Predicting aggregate risk of the population is not particularly helpful in care management. Instead, the goal is to accurately identify individual risk. Furthermore, simply identifying high-risk members is not nearly as helpful as identifying the members most likely to benefit from receiving a specific type of intervention such as a polypharmacy review, substance abuse counseling, or falls risk prevention.

Another reason that risk scores struggle to make good predictions for care management is that they only use claims data – which was designed only for billing. Conversely, the information care managers produce and rely on is spread across notes, care management systems, customer service calls, prescription data, and beyond. More on that here: The Dangers of Claims Based on Claims.

Moving Beyond Retro-Fitted Approaches

Advances in analytics including machine learning and natural language processing, now make it possible to use all available data in its native format, including free text, to develop highly personalized intervention targeting. Unfortunately, these approaches have largely remained the tools of researchers or those with considerable R&D budgets. As of very recently, this has begun to change.

The real potential of these approaches is personalization - a move beyond one size fits all, population-based statistics toward something far more targeted.

Finally, we can combine clinical evidence teaching us what matters with advanced analytics capable of helping us mine this signal from the noise of health data.

For example, at Cyft we're working with care management teams responsible for Medicare Advantage populations. Older adults have different pathways toward "risk." In fact, several studies have shown that a number of symptoms and events - cognitive impairment, impaired mobility, incontinence, etc. - known as geriatric syndrome are far more predictive than chronic conditions in predicting outcomes [3]. Evidence of geriatric syndrome is spread throughout several data sources from durable medical equipment to nurses notes - not usually in claims data.

In effect, we can stop pretending that "risk" faced by a complicated pregnancy can be measured with the same data and math as the risk of dealing with advanced diabetes at an advanced age. We can model the likelihood of relapse of substance use disorder with a very different and far more applicable approach than finding people that are advancing toward end-stage renal disease (for more on behavioral health, see our Cyft / Beacon webinar).

To be clear, these are not experimental approaches that may one day prove capable. They are grounded in 30+ years of thousands of clinical studies empirically proving what should now seem overly obvious - when it comes to finding people "at risk" in healthcare, considering more and more relevant data to identify more actionable sub-populations is better than applying less data and one size fits all math.

One need to look no further than nearly every other industry for evidence of just how transformative such approaches can be. From Target to Netflix to Amazon - the story of progress for the last 20 years is one of using more data to better understand who is likely to benefit from what specific offering. Now instead of selling more books, ads, and movies, we can get to work applying these approaches toward improving the health of people.

Leonard D’Avolio, PhD

CEO & Co-founder, Cyft

  1. "Accuracy of Claims-Based Risk Scoring Models - Society of Actuaries". Accessed 28 Jun. 2017.
  2. Health Aff (Millwood). 2015 Aug;34(8):1312-9. doi: 10.1377/hlthaff.2014.1186.