The 20 percent of people driving 80 percent of health costs this year aren’t the same people who will drive those costs next year.


In a population health program initiated in 2016, predictive data analytics suggested that a cohort of patients receive individual attention to head off future adverse health events. The caregivers on the ground reviewed the patients’ backgrounds and, seeing no history of high utilization or major health events, removed 100 people from the cohort.

Six months later, 90 percent of those patients had experienced one or more inpatient admissions, and half had at least one inpatient admission that was potentially avoidable. The total cost of those cases: $1.4 million.

This case study from 2016 illustrates the dual challenges faced by hospitals pursuing data-driven population health models: applying the right data to the right patients and intervening in the right ways, and overcoming protocols and “gut feelings” that cause human caregivers to put aside data-driven recommendations.

In working with 100 provider-led organizations across a broad spectrum of value-based risk contracts, retrospective analysis has repeatedly shown that the 20 percent of people driving 80 percent of health costs this year aren’t the same people who will drive those costs next year. Focusing care coordination efforts on today’s “frequent fliers” will not reduce costs as much as caregivers and hospital leadership hope, and can end up draining resources that don’t provide the expected return on investment.

Predictive data models, on the other hand, can point to who will be hospitalized next year with accuracy that frequently surprises health practitioners and challenges their expectations of where to focus preventive population health outreach initiatives.

Comparison of Traditional Versus Predictive Stratification on Future Medical Expense

Comparison of Traditional Versus Predictive Stratification on Future Medical Expense

A Necessary Shift in the Selection Process

When identifying patients for supplemental care coordination, physicians and care management staff often base their patient selections on set events, oftentimes a combination of the following criteria, a permutation of X condition plus Y visits:

  • Certain chronic conditions (diabetes, heart conditions, chronic obstructive pulmonary disease, etc.)
  • Recent acute events
  • Cost thresholds (e.g., $50,000)
  • Care delivery tallies that indicate health instability (e.g., two inpatient admissions within a timeframe or a combination of an inpatient admission plus two emergency department visits)

These patients logically make it into the care coordination pool because they are already costly, and they are top of mind for physicians because of recent treatments. This criteria-based approach has been in place for years and is still in use because the healthcare system is arranged around treating diseases. Medical directors and teams that are figuring out clinical strategies in the population health space are trained to think, “Find and target all uncontrolled diabetic patients; find and target all the frequent fliers.” However, in most cases, these teams aren’t collaborating with predictive analytics or data science teams to perform basic evaluations of these approaches.

The reality is that patients who have already had major acute events tend to stabilize, and their future utilization is not as high. You’ve already missed opportunities to reduce the costs of treating them.

On the other hand, another almost completely non-overlapping cohort of patients will soon experience these events if no interventions are executed.

A Tale of 100 Missed Opportunities

In the case study introduced above, a commercial entity responsible for 19,000 lives began a population health initiative. The data science team identified 353 people with common chronic conditions whose healthcare treatment, prescription compliance, physician visit patterns, and other variables combined to indicate they were headed for adverse health events.

Over the course of that timeframe, patients in the targeted cohort received individual attention. Typical of a time-limited outreach program, they worked with their physicians to set specific health goals, such as losing 25 pounds, and personal goals, such as being able to walk their daughter down the aisle at her wedding later that year. Care managers and nurses helped develop plans to achieve these goals, focusing on self-management as a key outcome of the project because for many people with chronic conditions, prevention and lifestyle choices have huge impacts on health.

But intervention occurred for only 253 of those patients. For the remaining 100, the care management team overrode the data analytics recommendations, determining as part of their review period of the selected cohort’s profiles and histories that those 100 people weren’t appropriate interventions.

We know what happened $1.4 million later. And while the data are not detailed enough to determine whether all the inpatient admissions were related to the conditions to be managed—the numbers do not exclude, for example, admission for a leg broken in a skiing accident—they give us one of the tidier retrospective case studies and point to both data and change management challenges. The patients who did not receive intervention ended up being 75 percent more expensive than the patients who did receive intervention in the six-month follow-up period.

The Greater Challenge: Override Intuition

Hospital leaders who initiate predictive analytics programs should be aware that such approaches may not direct them to target the cohorts they believe they should be impacting. Acting on predictive data can require a leap of faith—an informed leap—but still one that can go against the instinct to direct additional resources at people who already exhibit high utilization, rather than at people who do not yet. Furthermore, while the “a-ha moment” for initiating such plans may get intervention projects rolling at the executive and administrative levels, the “a-ha moments” that make projects work must be had at patients’ sides by clinicians on the ground—a completely different set of individuals. If those caregivers on the ground aren’t behind the chosen approaches, the patients—and the financial bottom line—will never see the anticipated outcomes.

Neither providers nor patients may feel the urgency to take action in predictive models. Once patients are admitted to hospitals for heart conditions aggravated by poor diet and exercise, the shock to their systems provides strong impetuses to make changes, but until those acute events occur, it’s hard to overcome lifestyle inertia.

On the caregiver side, gut feelings are that patients who have been hospitalized five times this year will need high-touch care next year. But it’s actually the people who haven’t been hospitalized yet who may be admitted next year. For example, patients who haven’t picked up prescriptions because their insurance doesn’t cover certain brands and who haven’t reached out to ask for alternatives may find their health conditions worsen to the point of needing hospitalization.

The Cricital Role of Line-Level Staff

It comes down to the line-level staff to ensure population health initiatives don’t fall short. Transformation happens slowly, and complaints abound about risk stratification. Physicians and their staff are used to seeing patients with a certain profile, and it’s difficult to overcome the doubt that they can better impact patients’ health outcomes—and the financial bottom line—by focusing on someone who doesn’t look or act like they need much of the care that clinicians are accustomed to providing. It’s not that their gut reactions are untrue—but it is a matter of realizing that they need to shift their lenses from acute-care providers to pre-care providers who deliver, or refer to, different sets of services.

Teams Preparation

For data-driven population health analytics programs to drive financial sustainability in value-based care paradigms, we need to do more to prepare care delivery teams before going live with investments in predictive analytics. Only with this preparation and this recognition of the power of predictive data will we be able to break the cycle of frequent fliers and achieve population health.


Shantanu Phatakwala is managing director of research and development, Evolent Health.

Discussion Starters:

Forum members: What do you think? Please share your thoughts in the comments section below.

  • Have you had any population health successes? If so, please describe them.
  • Is your organization involved in developing predictive analytics for population health? If so, please share some details.

Publication Date: Monday, January 15, 2018