OHDSI Home | Forums | Wiki | Github

OHDSI - Medicare researchers looking for guidance

I had the opportunity to chat with our OHDSI colleagues (@marciero @BriOlivieriMui) today about health services research projects using the OMOP CDM. Mike’s interests are around predicting individual costs using the Hierarchical Condition Category approach in a Medicare population.

It occurred to me that there’s a LOT of oral history here that isn’t easily findable for a newcomer. I’m starting this discussion on the forum to increase transparency to: 1) create a thread where people can see the state of what’s been done before and the opportunities for future research! AND… 2) help newcomers see there are people who do work outside of causal inference :wink:

To do this, I need some help from the community to help with “what’s been done”.

Going forward, one other ask:

  • @RossW - Mike is looking for some help understanding a bit about creating custom features in the PLP framework. Could you or some other PLP expert help orient him to this? (Probably easier to do in the R package than ATLAS.)

I really appreciate the community’s help in this! Would love to see this research come to life.


Just a few issues to keep in mind in any cost study.

  • Year will be important for inflating costs to a more current year, and to have consistent costs across years.
  • If you want to look at costs by type, you will have to decide what to do about things like physician costs which might be incurred in both inpatient and outpatient settings.
  • Don’t forget that people can have 0 costs. It is obvious, but it may require that you impute 0 for people otherwise you will have missing records for people. Keep track of your patient counts over time.
  • Censoring is important to keep track of – there are methods for inverse probability weighting for costs.
  • Death is usually not a problem since you have captured all of the costs for people who die – don’t ignore the difference between death and censoring in a cost analysis.

NCQA has a long history of using HCCs in our RA measures and I am currently working with some HL7 folks on the use of CQL to “map” CCs to HCCs for the HEDIS plan all cause readmission quality measure. I think this would be a great point of intersection with the conversations we have been having in the OMOP on FHIR measure use case regarding using CQL to “convert” ATLAS cohort artifacts. Both Bryn and I agree that reusable cohort definitions in a shareable repository would be a huge leap forward in the quality world.

While we have not worked around HCCs specifically (though thanks for the shout out @krfeeney ), I should note for @bnhamlin and anyone else interested in the Atlas-CQL relationship, a recent presentation and tooling from Michael Riley at GT on an Atlas Cohort Definition to CQL converter. Still needs a lot of love to get to community grade, but would love to work with others if there is interest.


1 Like

Michael’s presentation is the basis for what Floyd, Brynn and I are discussing as the first priority on the list of potential OMOP on FHIR dQM pilot projects. I think that HCCs might present a bit more complex of a problem than we would tackle in the first round, but we would be sure to keep whatever we test as extensible to the HCC concept without a complete redesign.

From the EHR WG perspective, the Cost table isn’t something that has been brought up in our discussions. I’m unsure how many have implemented this table. At Colorado University, we do not receive any Cost data. Dollar figures can be a sensitive subject with healthcare systems.

Just one other thing to be careful of – duplicated costs. From claims data, you may get claim level (rolled up) and line level (itemized) costs for the same visit.

This is why we ended up with an entire post-processing chain to select and identify relevant costs for a particular cohort. And even then it is difficult because different data sources may report different combinations of claim and line costs.

Sorry i missed this, happy to orient on creating custom covariates, this might be a good place to start: Populating the study package • SkeletonExistingPredictionModelStudy

Thanks for the entree @krfeeney. I completely missed this! Need to adjust my settings and check here more often. There are a few threads to pull on here. The idea is to model financial risk for medicare patients. Predicting things like total annual healthcare spending and medicare reimbursement might be parts of that. We also want to explore what factors drive cost and resource consumption from a provider perspective. I need to learn what kinds of outcomes the CDM is amenable to. Regarding HCCs-that is one possibility but there are a number of ICD9/10 based systems that have been used for modeling and prediction. These include CCS categories from CMS as well as a number of proprietary models such as Clinical Risk Groups (CRG) from 3M and ACG from Johns Hopkins.
An initial idea was to develop our own CRG type categories.
Most of what I have seen use categories, along with demographic, socioeconomic factors as candidate predictors and employ various statistical/ML approaches.

So I need to learn which OHDSI tools might support this project. I’ve done a lot of statistics and ML using R but have no expertise in other parts of the pipeline so for example using ATLAS would make things easier. The Book of OHDSI has been helpful in understanding what these tools do, as have the various working groups, community calls, and events.

Thanks to all who replied above. I will check out some of these ideas more closely, and will likely want to reach out to some of you.