This article was written with Arijit Sengupta, CEO of BeyondCore. Tim and Arijit will speak at Strata Rx 2013 on the topic of this post.
Current healthcare cost prevention efforts focus on the top 1% of highest risk patients. As care coordination efforts expand to a larger set of the patient population, the critical question is: If you’re a care manager, which patients should you offer additional care to at any given point in time? Our research shows that focusing on patients with the highest risk scores or highest current costs create suboptimal roadmaps. In this article we share an approach to predict patients whose costs are about to skyrocket, using a hypothesis-free micro-segmentation analysis. From there, working with physicians and care managers, we can formulate appropriate interventions.
Health systems are increasingly taking on risk for their population’s health and deploying care managers to provide additional care where needed. Many health systems in the U.S. are either starting to take on risk for part of their patient population or are actively considering options to do so. The Centers for Medicare & Medicaid Services (CMS) has been promoting coordinated care through Accountable Care Organizations (ACOs) and as of 2013, the ACO pilots cover almost 4 million Medicare enrollees.
In addition, a number of private payer-led initiatives in some markets cover even more of the population, such as Blue Cross Blue Shield of Massachusetts’s Alternative Quality Contract program. Health systems are also taking a new look at their own employees, because even a relatively small hospital is large enough to be self-insured for all health care costs. The rallying cry for all these efforts is payment for value (that is, improvement in health) instead of for individual visits and isolated treatment events.
The goal of these value-based plans is to lower the growth of medical costs while maintaining or improving quality of care. Patient-centered medical homes (PCMHs) are a common feature that assigns every patient to a single physician, who is accountable for coordinating the entirety of their care. Many patients, though, need additional care. Urgent care and hospitalizations can dramatically increase costs, so anticipating and preventing them take on a high priority.
Teams of care managers have been deployed by some systems to provide at-home care and counseling to the top 1% highest risk patients. Health Quality Partners in Doylestown, PA was detailed in a recent Washington Post article (”The solution Medicare is shutting down” by E. Klein) that showed how care managers work with their most complex patients one at a time. They help their patients adhere to a healthy diet, make sure that they schedule follow-up visits–and even arrange rides to those visits, among many other aspects of care coordination. HQP has proven it can bend the cost curve through these programs, and many other CMS ACO pilots have had success with similar programs.
Using risk scores as roadmap to prioritize care manager interventions is suboptimal
If you’re a care manager looking to expand this type of care to further improve your community, where do you go next? Many recent discussions and industry reports have proposed to start with the top 1% highest risk patients, then using the savings to invest in the top 5%, and then the top 10%, as shown in Figure 1 (PMPM = per-member per-month costs). It souds self-evident to do so, but such a top-down approach obscures patient-specific details.
Figure 1: Population risk pyramid and average costs per tier
Such top-down risk scoring algorithms are very effective at their intended purpose: for payers to adjust payment forecasts over a large population. But if you’re a provider who is looking for a roadmap to deploy care coordination resources effectively, you need to ask a different question: where is a dollar of investment in additional care going to be most effective? Top-down approaches provide generalizations, but care managers need to identify specific micro-segments of patients–and in many cases, individual patients–who offer opportunities for actionable intervention.
Figure 2 shows the distribution of costs relative to the tier average, in per-member per-month (PMPM) expenditures. Whereas the average (the vertical red line) is a single point, the variation can be very wide. The Top 1% includes many patients that are low-cost in any given year but are still at a high risk of succumbing to high-cost events. And the Top 5% shows a split: some low, some very high cost. This shows how we need to go beyond the averages in our view of patient populations.
Figure 2: Population risk pyramid and distribution of costs per tier
A micro-segmentation approach identifies small, patient-specific groups where intervention could be most effective
Hidden patterns emerge as we filter patient populations into increasingly granular segments. Figure 3 shows how the average annual growth in total medical costs, following a hospital inpatient discharge, is $4,200 (in other words, not counting the cost of the admission, the year following the event will have higher medical costs as individuals transition through post-acute care). If we segment that group based on whether those patients have been diagnosed with heart failure at some point in the prior 12 months, the small group with heart failure averages $ 17,800 in cost increases. Segmenting again, those on anti-depressants increase annual costs to $22,000. Perhaps most telling: of that group, those who fill less than 40% of their anti-depressant prescriptions in the following year experience a cost increase of $33,000. It’s only when we look at the combination of these variables that the hidden patterns become clear, the variability declines, and actionable opportunities start to emerge.
Figure 3: Annual growth in medical costs after inpatient discharge, with increasing segmentation
In this example, could a care manager intervene in such cases where the anti-depressants are not being filled regularly with the aspiration of both improving patient outcomes and lowering a $33,000 cost increase to $17,500?
Statistical disclaimer: We confirmed that lack of adherence to antidepressants was far more predictive of this outcome than adherence to any other medications included in our analysis. We also confirmed that the use of antidepressants before the event was not significantly predictive while the use of antidepressants after the event was predictive. The findings were reviewed with experts to confirm whether there was reason to suspect a causal link in this case. While a causal link seems likely, at this point we have not clinically proven a causal link and this pattern is intended to be illustrative. It is not intended here to imply that the cost increase is avoidable by increasing anti-depressant usage, for example.
This pattern was detected without pre-conceived hypotheses using automated analysis. There are millions of similar micro-segments, and given the number of potential pre-existing conditions, diagnoses, treatments, demographics, and medicine usage, it is not possible to manually conceive of and test all possible hypotheses. We need a mechanism to evaluate them all in a hypothesis-free way and rank-order which of the segments provide the greatest opportunity for additional care.
Each pattern may also have confounding effects. For example, patients in our micro-segment might be older than the overall population. Therefore we need to statistically adjust for the other variables to check that the combinations chosen are the most predictive.
Using current medical costs as roadmap to prioritize care manager interventions is suboptimal
We discussed how using risk scores is a suboptimal approach for prioritizing patient intervention efforts. But what about prioritizing patients based on their current total medical costs? Figure 4 shows total annual medical costs for 200M commercially insured patients in 20 segments.
The top 5% by cost in any given year indeed account for most of the spending, but only the “top 1%” consistently remain high cost: the other 4% move in and out every year, as shown in Figure 5.
Figure 5: Population grouped by annual costs, showing how most of those in the top 5% change every year
An increase of over $10k per year in medical costs is enough to move from the 50th percentile ($1,300) to the 90th percentile ($13,000) in Figure 5. Every year, about 5% of patients experience increases in their costs by that amount–and a similarly sized group decrease by the same amount, as seen in the orange bars on the far left and far right of the histogram in Figure 6.
In other words, outside of the top 1%, patients’ costs jump and then re-settle. You can have a health incident and then go back to a normal life. Depression can cycle periodically. If we focus our additional care on the current top 5%, then we’d be trying to intervene when it was probably already too late. The current top 5% is not a static group.
Figure 6: Almost 10% of the population have annual costs that jump up or down by more than $10k in any given year
The next “top 5%” are the fluid set of specific patients whose costs are about to jump the most
The next top 5%–those whose costs are about to jump–is the group that care coordination can be most effectively deployed against. Part of their cost increases will be unavoidable inpatient events, but the transition back to ambulatory care and a normal life are where care coordination efforts can be most effective. This transition is where patients often need the most additional care and support. For this preventative approach to be effective, this next top 5% have to be rapidly predicted, micro-segmented and appropriate interventions have to be deployed to prevent them from actually becoming part of the top 5%.
Objective Health, a McKinsey Solution for healthcare providers, is presenting this and other insights at Strata Rx in Boston, September 27, 2013. This analysis was powered by BeyondCore, an automated solution that makes Big Data analytics accessible to business users. The data source for most analyses is Truven commercial claims data, 2009-2011.See a video related to this topic, One-click analysis: Detecting and visualizing insights automatically.