Tuesday, April 5, 2011

Hierarchical Condition Categories (HCC) and Accountable Care Organization Risk Adjustment: Is Prediction Accuracy of 11-12% of Expected Utilization Enough?

Take a spin on HCCs?
Addendum - DavidN below points out there is an important difference between individual (the topic below) and aggregate HCC-determined risk.  The explanatory power can  go as high as 60%, which is on par with many commercial predictive models. The DMCB appreciates the reader input and placed the new insights into the text below in bold font.

It also wonders if even a "low" 40% unexplained variation may still be too much for ACO wannabes...... 


In a prior post, the Disease Management Care Blog introduced budding Accountable Care Organizations to "ambulatory sensitive conditions" and its curious mash of of "applied" health services research and reimbursement. 

In this post, the DMCB introduces the ACO wannabes to managed care's notorious "hierarchical condition categories" (HCC).
This is important because in the ACO Proposed Rule (**NEW** link over on the right of this page), CMS is proposing that HCC be used to risk-adjust the ACO's attributed population's claims expense "benchmark" baseline.  Comparison of that baseline (plus an actuarial trend) versus the ACO's expenses will determine whether the ACO is a hero.... or, literally, a zero.

HCC use age, gender, Medicaid eligibility, Medicare status (age, disabled or end stage renal disease) and the ICD-9 claims record of prior "encounters" to assign patients to 70 different "categories."  Those categories are designed to  "bucket" patients that are not only clinically similar (based on organ systems) but follow similar cost patterns. 

The reason the system is "hierarchical" is because it reconciles the presence of multiple concurrent conditions by identifying a dominant one.  As additional conditions are added to the model - since patients often have multiple diseases simultaneously - their predicted impact is incrementally adjusted upward. The "rate" of that upward adjustment decreases as more conditions are included.  In other words, it's not additive.

HCC have been CMS' answer to the perennial complaint of hospitals and physicians that their patients are "sicker."  Confront any provider with higher than normal claims expenses, unexpectedly long lengths of stay, higher than average rates of surgery, lower satisfaction scores or substandard quality measures, the first explanation will always point to the patients, not the care.  HCC case mix risk adjustment is designed to use the Medicare patients' characteristics to mathematically reconcile the observed rate versus what would be expected so that the quality of the care can be isolated and understood.

With one significant exception, HCC, like "ambulatory sensitive conditions," have been used in health services research to compare and contrast patterns of care and inform public healthcare policy.  That exception is that HCC have also been used to risk adjust the payments from CMS to its Medicare Advantage plans.

CMS experience with MA Plans has introduced a HCC "wrinkle" in the Proposed Rule.  Reportedly, when HCCs first arrived, MA Plans responded by hiring consultants and upgrading their claims systems to increase the "accuracy" of the coding.  This also increased the "coding intensity," resulting in an upward trend in their claims profile, which made patients appear sicker.  This cleverly shifted the HCC case mix and led to higher payments in the following year.  CMS, arguing claims expenses for beneficiaries generally remain stable year to year, proposes to avoid a reprise of that trickery by using the same HCC from the baseline year for all three of the ACOs' contract years.

And, in case you're interested, new Medicare enrollees would be risk adjusted on just age, gender, Medicaid eligibility and Medicare status. Indirect medical education (IME) and  disproportional share hospital (DSH) payments would not be backed out of the case mix adjustment.

All well and good, says the DMCB, but what ACOs really need to know is how accurate the HCC system is.  The answer, from the perspective of health services research, is "pretty good."  If, however, you are betting millions of dollars of investment, the answer may be "not very."  According to this definitive paper, the individual "explanatory power" of HCC ranges between 11% and 12%.  In other words, close to 90% of the year-to-year variation in the observed vs. expected claims expense that will determine payment levels from CMS under the Proposed Rule could be unpredictable. 

If aggregated and rolled up, the explanatory power may range between 60-70%.

Managed care organizations - with hundreds of thousands, if not millions, of enrollees - are used to underwriting cycles and understand the mathematics of transferring risk.  They use underwriting, the law of large numbers and surplus to cushion claims expense unpredictability.  They can handle 90% (or 40% in aggregate) swings in unexpected utilization. 

It remains to be seen whether smaller and lower capitalized ACOs will be willing to accept that level of uncertainty.  One possible answer to that will be addressed in tomorrow's DMCB post.

Image from Wikipedia


DavidN said...

The 11-12% "explanatory power" represents the predictive ability of *an individual's* future health care expenses. This value would increase when predictions are evaluated in aggregate (which I have done for a Medicare Advantage MCO), as would be the case for an ACO - so variations of 90% in utilization is not a valid presentation in the view of HCC Risk Adjustment applicability to an ACO, or any, aggregate population.

Jaan Sidorov said...

Thank you DavidN! How much variation can be suppressed by aggregating the scores?

DavidN said...

It of course depends (that word always has a tendency of appearing, doesn't it?) on the inherent variability of the measure being predicted, but given a reasonable population size (say > 250 - I've done as low as 75), it can be significant... I have obtained correlations on admission rates through HCCs of upwards of 0.8, so >60% explanatory power for an aggregate population.

Jaan Sidorov said...

Great point. The insight is now at the top of that post.

DavidN said...

Thank you! I will add that the ~60% explanatory power I obtained was with a 'simple' application of HCCs - meaning, the weights were not adjusted from their standard model used to predict individual future costs. Apply some additional "up front" work on the HCC model to obtain updated coefficients for aggregate prediction of specific measures, and I am sure even better results could be obtained.