tuva-health / tuva

Main repo including core data model, data marts, reference data, terminology, and the clinical concept library
https://thetuvaproject.com/
193 stars 51 forks source link

Build the HHS-HCC risk adjustment model #276

Open EeethB opened 1 year ago

EeethB commented 1 year ago

Here's my overview of the domain knowledge side of ACA risk adjustment. Hope it helps kick off the discussion!

Overview

Similar to the CMS-HCC model, the HHS-HCC risk adjustment model uses a hierarchical system to roll up from ICD-10 diagnosis codes (and a few others) to condition groups which represent moderate- to high-cost conditions under Patient Protection and Affordable Care Act (PPACA or just ACA, also known as Obamacare) coverage. These condition categories combine with demographic risk to create a risk score for each organization offering ACA coverage.

Motivation

The ACA ensured individuals would have access to health insurance by removing the ability of insurers to increase premium rates or deny coverage to individuals based on pre-existing conditions. This restriction may result in an issuer attracting higher risk through chance or adverse selection, which leads to higher premium rates for all members and further increases the chance of adverse selection (The so-called "death spiral" 💀 if it continues unmitigated). To protect against this selection, the risk adjustment program measures the average risk of each issuer and transfers funds from lower-risk to higher-risk issuers. The measurement is done amongst all issuers within a state for each benefit year, and payments are zero-sum.

Details

There are several pieces that go into a member's risk score. One of the best references for the different pieces is the DIY software published each year by CMS: Find it at https://www.cms.gov/marketplace/resources/regulations-guidance, search "DIY", and choose the relevant year's model.

  1. Demographics - All members get a base risk score determined by age and sex (DIY Table 5)
  2. Medical conditions
    1. Valid* ICD-10 codes roll up to condition categories (Table 3)
    2. Condition categories get a hierarchy applied, where certain more extreme categories cancel out less extreme ones (Table 4)
    3. Some HCCs are further rolled up into category groups, or contribute to an HCC count/severity factor
  3. Prescriptions
    1. NDCs and HCPCS roll up to pharmacy categories (Tables 10a and 10b)
    2. These pharmacy categories have a hierarchy applied (Table 11)
  4. Some medical condition/prescription combinations contribute an additional factor
  5. A small enrollment duration factor may be added on if a member has less than 6 months' enrollment in a year (Table 6)

*A diagnosis code is valid if falls within a member's eligibility period and is on a claim that also contains a face-to-face CPT code (Table 2)

Each of these categories listed contributes a portion to the additive risk score: Demographics + sum of rolled up medical + sum of rolled up pharmacy + interactions + enrollment duration = risk score. Relevant categories and their coefficients change depending on whether a member is an adult or child. The model for infants (Age 0 or 1) is done slightly differently, where conditions are further rolled up to a five-level severity category. The relevant categories for each model can be found in Tables 6-8, and the actual coefficients are in Table 9. If a member is enrolled in a CSR plan, this additive score gets a multiplier applied to it to account for induced demand and benefit richness.

Besides the DIY software, CMS publishes another document with these coefficients much earlier, I think it's either the Notice of Benefit and Payment Parameters or the Final Rule. I don't remember how to access that at the moment though.

aneiderhiser commented 1 year ago

@EeethB this is a great summary. Obviously a lot of differences compared to CMS-HCCs. Conceptually a lot similar though. One difference that stood out to me: NDC lookup for drug categories. That's not part of CMS-HCCs, right?

EeethB commented 1 year ago

@aneiderhiser I'm not as familiar with the CMS-HCC model, but that sounds right. I think the drug piece is a new-ish addition to the ACA model as well

sarah-tuva commented 1 year ago

@aneiderhiser @EeethB That's correct! Pharmacy/medication data is not taken into account in CMS-HCC model. There is a separate RX-HCC model that we have not built yet. That model is focused on Part D enrollment.