sohyeon
(Sohyeon Jeon | Seoul National University)
1
Hello,
I’ve been wondering how others handle a situation I ran into repeatedly in OMOP-based clinical trial feasibility work.
For example, consider a criterion like: “Blood glucose >180 mg/dL after reperfusion of a liver graft."
The glucose result may be available, but the reperfusion time may not exist as a distinct structured event in the local OMOP data. One site may have it, another may need a proxy, and another may have no good temporal anchor at all.
These feel like meaningfully different situations, but in practice they can still end up looking very similar in the final count. In particular, a zero from the third case can be hard to distinguish from “no eligible patients.”
What I’m most curious about is whether these differences get recorded anywhere that stays with the study. If a proxy is used, where do you usually document what was chosen and why? And in a multi-site study, how do you handle it when the same criterion is directly available at one site, proxy-dependent at another, and not really representable at a third?
I’d really appreciate hearing how others have handled this in practice.
This is tricky. The blood tests “after reperfusion” are usually done to test whether or not a transplanted liver starts working or is being rejected. So, you check liver function tests (AST, ALT, GGT etc.) and the cytokines after the blood starts flowing through the graft. The time intervalls are usually minutes and hours.
That kind of stuff is notoriously not easy in real-world data. Typically, the unit of time measure is day, which means, you will see like 20 measurements on that day of the transplantation. If that is the case, you can’t do a dynamic study thightly following the transplantation.
But the data are there, you need to ETL them. The surgery report will tell you exactly when the hepatic artery is connected. If not, the end of surgery minus, say, 30 or 60 minutes should be a good guess. The lab tests also have time stamps with hours and minutes. You would need the time of the blood draw, but if you only get the time of the test result it should be fine, too: In those cases, they take the blood to the lab and put it through the machine immediately. The difference will be minutes.
If you are not interested in a dynamic study, but want to have a patient cohort with a functional liver transplant, forget the minute game. Just check for the labs the next day or some days after. The new liver either works at that point or not.
This was very helpful. The clinical examples were especially useful, particularly the point that the reperfusion time may still be recoverable from the surgical record, and that when it is not, a practical approximation can sometimes be justified from the surrounding clinical workflow. That helped me separate what is missing from the OMOP instance from what is not available in the source data.
What I’m still turning over is what happens in a network study. If those kinds of judgments are made site by site, I wonder how much of that variation remains visible to the study team, whether through the shared protocol, site-level characterization, or simply as something worked out as the study proceeds.
Thanks again for taking the time to explain this. The ETL side is much clearer now.
Yeah, this the typical challenge of real-world evidence. The scientific question is of biological nature, and requires the assertation of some biological facts or events. The data on the other side contain information about healthcare conduct and payment justification. The bridging between these two is the hard part. In order to do that, we apply:
The standard model
The standard vocabs
The Themis conventions
When the above don’t give you what you need - phenotypes
This latter is rarely straightforward, we often have to be creative and figure out some work arounds, like in your case. That requires what is usually called “clinical experience”, but that is a misnomer, because it is the tacit knowledge of the various pitfalls and tricks.
And then there is the additional problem you mentioned: If you want to do network studies, which are much more powerful, all participating data partners have to follow the same standard, otherwise you have “variations” as you call it. And those are really painful.
At OHDSI, we really haven’t started dissecting surgeries and come up with conventions. Feel free to start a WG for that.
@Evan_Minty and friends run the Surgery and Perioperative Medicine WG. Find them here on MS Teams.
sohyeon
(Sohyeon Jeon | Seoul National University)
6
Thank you, @MPhilofsky. Both accounts are mine, sorry for the confusion! I’ll keep to sohyeon from here.
I’ll join the Teams channel and reach out to Evan and the friends.
One thing I’d like to ask the group is whether they’ve encountered a surgical event with two or more defensible definitions, say a recorded timestamp, a procedure based proxy, or a postoperative window, and whether those alternatives were ever compared across sites. Even hearing that it hasn’t come up would be useful.
sohyeon
(Sohyeon Jeon | Seoul National University)
7
Thank you, Christian. That ordering is useful, model then vocabs then Themis, and phenotypes when those run out. Melanie also pointed me to the Surgery and Perioperative Medicine WG, which seems a better place to start than anything new.
I’ve been sitting with the network part since. Running two or three prespecified variants everywhere and comparing them is something the tooling already handles, and CohortDiagnostics gives me the overlap. What I keep coming back to is a different case, where the variation isn’t one I chose but one each site’s records force. One site has the timestamp, another needs a procedure-based proxy, and a third has neither. The clinical intent is the same everywhere, but the executable definition may no longer be. Your suggestion to step back from a minute-level anchor when the study does not need it removes the problem for that criterion, and I’ve taken it. I’m less sure what to do with criteria for which there is no defensible coarser version.
The part I can’t place is what reaches the study team at the end. A true zero and a criterion that could not be executed can collapse into the same downstream result. The latter is then either reported as zero, which is wrong, or omitted, which quietly changes the definition. Neither is visible in the count alone.
What I keep sketching is small: alongside the count, a criterion-level execution status, something like directly observed, executed via a prespecified proxy, or not executed in this instance. That would let the study team distinguish a real zero from a criterion that never ran. I wonder whether this sits closer to a Themis-style convention about reporting execution than to phenotype definition itself.
I’d be glad to hear whether something like this already exists and I simply have not found it, or whether the issue is handled another way.
