OHDSI Home | Forums | Wiki | Github

Negative concepts

Vocabulary Users
#1

Hi OHDSI community!

Early we refactored History concepts and created “History of event” one-bucket concept with a rich hierarchy instead of “History of clinical finding” and “Past history of procedure” hierarchies.

But there are other SNOMED branches left that reflect historical, but negative information: “No history of procedure”, “No history of clinical finding in subject” and “Clinical finding absent”.

Of course, we know that negative concepts are not the best means for standardized analytics and almost out of the scope or support in OMOP.

But sometimes users that deal with registry/survey-like and other data don’t use Altlas and other tools widely (because of the limitations), but develop their own analytics. In this case they’re willing to save the facts of negative information and build the queries around it.

Our proposal is to use the same approach for negative concepts: isolate them in a new hierarchy and add the enhanced time-span concepts.

Given the fact that these concepts already exist in OMOP as Standard concepts, we don’t break the rules. We only want to make it more organized. It won’t be any change in conventions. This would just help to deal with the negative informations for those users who can’t use our standard approach.

@Christian_Reich @Alexdavv

#2

@Maria_Rogozhkina:

There are two different issues, both violating the Closed World model of Real World Evidence: (i) we know everything that happened, and when it happened (positive facts), and (i) if something isn’t there amongst all the records we can be sure it did not happen (only inferred negative facts). Of course, we never have perfect data like that, but we still are happy to assume that we are close enough to bank all our statistics on it. Well, almost. Sometimes the data are trying to make up for their insufficiencies:

  1. History of concepts are a makeshift solution when we know something happened, but we don’t know when. The moment we learn about it we know it has to be in the past of that moment, hence the “history” moniker. Such “step child” records only hold “half” the info they are supposed to hold, and we can only use them for a small subset of use cases (usually for exclusion criteria of past events). And that is only if we remember them, Atlas doesn’t currently remind us.

  2. Negative concepts are a makeshift solution when we know we don’t have comprehensive data, such as in registries and surveys. Instead, we only poke out something that we deem interesting. Since we can’t rely on the assumption “if we don’t have a positive fact we can assume it never happened” we need these negative concepts. They, too, have a “history of” character, because we can only not have an event in a period of time it is declared. Otherwise one would have to declare “no AMI” at one day, and the day before, and the day before that, and every day of the Observation Period.

“No history of” concepts are the same as negative concepts.

Bottom line: they violate the principles, and their value is extremely low. They can’t be used for things like rate calculations. If you want to use them in complex cohort definitions they are only useful as exclusion criteria, if you happen to have to exclude exactly that fact, and only if they are declared at or after the index date (otherwise things could have changed in the meantime). In other words almost never.

So, if you ask me - don’t add any. If folks have non-RWD data, such as surveys or registries, and they want to use OMOP for convenience, they should do so. But they are operating outside the OMOP CDM and should use 2Billionaires for this stuff.

Is it worth actively getting rid of the ones we happen to have - probably not. We have more important vocab work to do.

1 Like
t