Calculating condition periods (condition_end_date)

[Suranga]

Hi!

In populating the condition_occurrence table, we’ve been trying to figure out how to best determine accurate condition_periods. Basically, we’re trying to determine the best way to populate the condition_end_date column of the condition_occurrence table.

Issues we’re trying to figure out are instances such as:
If patient A reports condition X during his 1st and 3rd visits, but not his 2nd, how do we report condition_end_date, etc.

I’m told that @jon_duke may have some experience in setting up algorithms to do this for RMRS, and was hoping if he could clarify what he’d done?

[Christian_Reich]

@Suranga:

You do - nothing. Unless you have specific information when a condition ended (was resolved, the infection healed or anything like that) you leave the condition_end_date empty. That way, it becomes a snapshot diagnosis or recording at that time.

The CONDITION_ERA table is what will try to infer the actual duration of a condition that is recorded multiple times over time. It uses a standardized algorithm for doing that found here or here.

[Gowtham_Rao]

Unless you have specific information when a condition ended (was resolved, the infection healed or anything like that) you leave the condition_end_date empty. That way, it becomes a snapshot diagnosis or recording at that time.

Based on the standardized algorithm - is it correct then that:

• if end-date is not specified, then it assumed to be 30 days after?
• Is it reasonable to apply this convention to chronic conditions like Diabetes Mellitus and Hypertension? Would it imply such a condition has ended? Would it be reasonable to apply longer days for such chronic conditions - atleast 1 year?

Thank you

[Christian_Reich]

@Gowtham_Rao:

Nope. As discussed before: The guesswork is in the CONDITION_ERA, only. There, the 30 days is the persistence window, which means, your are looking within 30 days after the first Condition record for another one with the same Concept, and if you find one you extend the era to that one (the start date of the second becomes the end date of the era). You repeat that process till you can’t find any more Concepts within 30 days of the latest addition, but it ends at the last one, not 30 days thereafter. So, you don’t add 30 days to anything.

The CONDITION_OCCURRENCE doesn’t assume anything. But analytical tools generally interprete an empty end date to be the same as the start date, like you would for an acute condition. Ie., you had a heart attack that day.

Obviously, for chronic conditions that is known to be a suboptimal assumption. You can’t have type-II diabetes or hypertension for a day. Folks always point that out, however, we haven’t heard from anyone about a systematic or algorithmic method how to know when to expect a condition to end. If you have an idea bring it on.

[Gowtham_Rao]

Thank you - agree that the assumption is sub-optimal for chronic condition. Documentation of end of a condition is very rare; it is unlikely to find persistence in documentation every 30 days for Diabetes Mellitus or Hypertension - where the expected rate of visits with PCP is maybe every six months.

Regarding systematic algorithm (apologies if this was already discussed): Would it be reasonable to have different persistence rules by condition. Example: use six months for chronic conditions and 30 days for acute conditions. Another alternative – use, drug_era in association with condition_occurrence that indicated the drug, i.e. persistence of treatment, as an indicator of condition persistence.

Thank you

[Christian_Reich]

@Gowtham_Rao:

It would be very reasonable. Only problem: How does the ETL know which condition is chronic, and which is acute? For the usual cases (heart attack=acute, diabetes=chronic) it appears obvious, but what about things like cancer, pain, depression? Acute or chronic? Your hypertension - it can be fixed with exercise and medication. Even diabetes - get rid of your 100 extra pounds and eat healthy, and unless the pancreas has given up completely you can live without any medication. My grandfather did exactly that.

So, I don’t know how to make those guesses. Until then, we will have a bunch of diabetes eras with gaps of more than 30 days between them.

[Gowtham_Rao]

Thank you. I know of one source - but it is only for ICD9.

https://www.hcup-us.ahrq.gov/toolssoftware/chronic/chronic.jsp

ICD10 version https://www.hcup-us.ahrq.gov/toolssoftware/chronic_icd10/chronic_icd10.jsp

“A chronic condition is defined as a condition that lasts 12 months or longer and meets one or both of the following tests: (a) it places limitations on self-care, independent living, and social interactions; (b) it results in the need for ongoing intervention with medical products, services, and special equipment (see Perrin et al., 1993).”

[Christian_Reich]

Sounds like a starter.

We should make an eperiment and apply this to real data:

• If a Concept is declared “acute”, use a persistent window of 30 days (or even less).
• If a Concept is declared “chronic” use a persistent window of 12 months.
• Instead of persistence window, add 12 months or whatever is left to the end of the observation period

And check the resulting end_dates of eras. For acute, it should cease when no more clinical activities is seen. The chronic should essentially follow some expected length for their resolution.

Want to do that?

[Gowtham_Rao]

Yes - definitely interested (we dont have an environment with data in OMOP CDM yet).

Being new to the community, I have been reading historical discussions: one relevant to this topic is here The community raised some good points - anything between then and now, that we could use to further this discussion? Is it possible to link/merge the two threads?

Here is the ICD 10 version https://www.hcup-us.ahrq.gov/toolssoftware/chronic_icd10/chronic_icd10.jsp

How exactly are eras generated? (e.g., Rx vs. Dx)

Considering diagnoses:If someone has repeatedly diagnosis of osteoporosis, I abstract an era lasting 5 years.However, how is it done for urinary track infection (which usually lasts weeks)?Is it the same rule for all diagnoses or is it diagnosis specific?How do I find the gap for osteoporosis?

How exactly are eras generated? (e.g., Rx vs. Dx)

Gents:

Why couldn’t the data tell that? We could for each product generate the distribution of distance between individual admistrations (in case of drug) or diagnoses (in case of condition) and try to infer what a good persistence window would be.

How exactly are eras generated? (e.g., Rx vs. Dx)

Yes, I am currently exploring the use of hidden markov models for inferring
condition eras and drug eras. The hope is that we could learn
condition/drug specific parameters that would both characterize the
chronicity of the condition/drug and allow for inference of condition/drug
eras that are robust to noise.

[Christian_Reich]

@Gowtham_Rao:

You can’t merge. You can just leave a forwarding link at the end of a stale debate to the fresh one.

The reasons it went stale are these:

• Nobody had time and money to do anything
• It is not clear why it is important. While on the surface correct eras would be a good thing to have, the actual analytics don’t care in most cases. For example, if you count things up for chronic patients (co-morbidity, provider specialties etc.) or you do causal inferencing (drug safety studies) it actually doesn’t matter if you have one long diabetes era or 5 smaller ones. The result comes up the same. The only metric that wouldn’t give the right result is “length of disease”.

C

[Jerrod.George]

Sorry to bring up an old thread…Being new the community I wanted to throw out a question on if the persistence window for CONDITION_ERA was ever “solved”? At the moment we plan to go with a 30 day window but has anyone implemented anything else with Ambulatory data.