OHDSI Home | Forums | Wiki | Github

Using ATC to classify drug use in children

Researchers
#1

Hi all,

For our study on drug utilization in children we are running into the following problem: We’d like to look at drug use classified by high-level ATC category, but some prevalent drugs appear in many categories, even when that doesn’t make sense in a pediatric population.

For example, ibuprofen has several ATC codes, one being C01EB16 which falls under “other cardiac preparations”. This means we see many prescriptions in the “cardiovascular” category in children. But this doesn’t really make sense; Most prescriptions would in reality fall under M01AE01 (anti-inflammatory).

Does anybody have any suggestions for limiting the number of classes under which we count a drug?

#2

@schuemie:

Not sure what you are trying to achieve. That every drug appears only in one classification? That’s not how it is defined. Classes are not unique or one-to-one for drugs. The fact that there are two ATCs doesn’t help you. Nobody enforces to use one ATC over the other one depending on the indication.

And think about it. The cardiovascular and the anti-inflammatory indication are based on the same effects. In one it inhibits platelets, in the other one cells of the immune system and central nervous system (controlling the temperature). So, you really cannot distinguish them.

So, use the OMOP concept for ibuprofen (there is only one) and count all drug classifications. If you really want indication you need to look for Conditions. Patrick did a nice DP-based estimator for that some time ago.

#3

What do you mean by high level. We had the same problem in picking best ATC level. Is ATC3 high enough?

(or you mean ATC2 as your high level target?)

Also, what is your starting point - Drug Exposure table with Clinical+Branded drugs?

Vojtech

#4

What we’re trying to do is exactly what was done in this article. We’re trying to summarize drug utilization at the 1-digit and 3-digit ATC code.

I know in theory ibuprofen also has cardiovascular activity in children, but it is just not very helpful for a utilization study. (Doctors prescribe ibuprofen to children primarily to suppress fever) To formalize what I have in mind: for every ingredient, I would like to determine the probability of an ATC code being applicable for that ingredient in the population of interest, and remove every ATC code with a probability below x% (e.g. 10%).

#5

@schuemie:

I think you are mixing things up here (and so did Miriam, but she didn’t seem to care, she just pushed out the histograms): Drug classes and indications (or what you call “populations”, because your logic is that children don’t have cardiovascular disease to treat). There is only one ibuprofen. It can be used for several indications. Sometimes, the drug classes used in clinic use indications as a method of taxonomy, sometimes chemical classes, sometimes mechanisms of action:

  • ACE-inhibitors are based on mechanisms of action, even though that mechanism is used predominantly, but not always, to treat cardiovascular diseases.
  • Opiods are based on a chemical classification, and as pain killers they can have a gazillion of indications (unless you call “pain” an indication, which you could do).
  • The best are NSAIDS: They are a mixture of chemical (“non-steroidal”) and indication (“anti-inflammatory”) classifications.

Sorry, but I think what you want to do is impossible. At least with ATC. You can try NDF-RT. Because that one was explicitly designed to do exactly what you want to do: To create a set of clearly defined, mutually exclusive drug classifications that are either chemical, mechanistic or indication-based. It is also free and available for all drugs in the OMOP vocabulary.

#6

I just want to point out that assuming that “children don’t have cardiovascular disease to treat” is not necessarily correct. Many children are being diagnosed with congenital heart conditions these days and therefore it is possible they are being giving ibuprofen to treat a congenital or acquired pediatric cardiovascular condition. It might be better to develop a method that is robust to the possibility even if it is less likely then say ibuprofen being used for fever.
Also I think the bias that ibuprofen is used more often to treat fever (then to treat cardiovascular disease) probably also applies in older populations especially when you consider how often ibuprofen is given.

#7

Thanks so much for the discussion everyone!

I know I’m mixing up indications and drug classes, and I realize now that even in children ibuprofen could be used for cardiovascular diseases. But I’m still left with a need to classify drugs to a higher level of concepts for communication purposes.

After the suggestion of @Christian_Reich I looked at NDF-RT in the vocab. I see several types of classification:

  • Therapeutic class (30)
  • Pharmacological class (584)
  • Mechanism of action (609)
  • Physiologic effect (1,866)

Most of these have too many classes to exhaustively list in a paper. I like ‘Therapeutic class’ (maybe because it most closely resembles the indication :wink: ), but hardly anything is mapped to that classification. For example, no ibuprofen concept has any relationship to any concept with concept_class_id ‘Therapeutic class’. Why is that?

Note that the ATC Drug Dictionary does have a notion of ‘most common use’, preferring one ATC code per drug over all others (at the trade name level).

#8

An update on my search for a good drug classification system: I’ve been playing with the NDF-RT ‘Pharmacological classes’, which appears to be the main classification. However, the classification again is far complete. Roughly 40% of the RxNorm ingredients used in Japan does not have a classification.

Examples of drugs with very high prevalence but no pharmacological class are:

  • Loxoprofen (an NSAID)
  • Carbocisteine (a mucolytic)
  • Ambroxol (another mucolytic)

It seems a bit silly that we don’t have a good classification that covers all drugs. Any suggestions will again be greatly appreciated…

#9

@schuemie:

Sigh. The classifications are not comprehensive, that’s true. ATC is the best from that perspective, but it has the other problems discussed above.

This is a problem that would require a systematic but manual approach. I don’t have anything else to offer. :frowning:

For the new drug vocabs we intend to do that, by stealing from ATC and suggesting the other classifications.

#10

For what it is worth: We ran into this same issue and we ended up creating customized NDF-RT style indication/contraindication linkages (“may_treat”, “may_induce”, “may_prevent”). We also defined these relationships as “approved/primary indication” and “off label or unapproved”.

NDF-RT is horrible for newly approved / pending drugs – very slow updates in terms of novel indications/classes coming online.

One note about the ATC (we use this too) - It is critical to consider the route of admin & notes/additional info from ATC when looking across multiple categories to know exactly how they intend it to be used in cases of ambiguity.

We can see that C01EB16 is assigned a route of parenteral only — not oral - and “Products containing indometacin or ibuprofen, which are only used for closing the ductus arteriosus in premature infants, are classified here.” (http://www.whocc.no/atc_ddd_index/?code=C01EB16)

In contrast, ibuprofen M01AE01is oral and rectal: http://www.whocc.no/atc_ddd_index/?code=M01AE01.

#11

Hi all. I’m still stuck with this question, but have a (probably stupid) idea for solving it.

Short recap: we’re doing a drug utilization study and would like to summarize drug use at a high level. We can’t use ATC for that because many prevalent drugs belong to multiple high level ATC classes, and therefore the results show pretty much the same numbers in every class (e.g. ibuprofen falls in the ‘cardiovascular’, ‘musculoskeletal’, ‘genito urinary’, and ‘respiratory’ classes). We can’t use NDF-RT because many drugs are not classified (yet) in NDF-RT, and also there are too many NDF-RT classes and no hierarchy.

While trying to come with a new very high level drug classification scheme, I ran into the one used by Drugs.com. I like this one because (a) the highest level classes are decomposed into clear, mostly unambiguous lower drug classes, and (b) there are only a few top level classes, and I think they are more informative than the ATC classes:

• allergenics
• anti-infectives
• antineoplastics
• biologicals
• cardiovascular agents
• central nervous system agents
• coagulation modifiers
• gastrointestinal agents
• genitourinary tract agents
• hormones
• immunologic agents
• metabolic agents
• psychotherapeutic agents
• respiratory agents
• topical agents
Classes I might not use in our paper:
• alternative medicines
• medical gas
• miscellaneous agents
• nutritional products
• plasma expanders
• radiologic agents

I can think of several ways to semi-automatically map all drugs in the vocab to this classification. The question is: does anybody know where Drugs.com got the classification from? Does anybody know if we could use it? Does anybody know an alternative?

#12

@schuemie:

Sorr,y but it won’t help you. Because the categories, even though there are fewer, are still organized by different principles, and therefore have to overlap. Unless drugs.com just ignores that and assigns drugs only to one class, but that would be wrong. For example, biologics (organized by chemical structure) are also anti-infectives, allergenics and immunologic agents (all by indication). If you take biologics out of the those classes (sounds like that’s what they may have done) then the numbers are useless. For example, rheumatoid arthritis today gets treated mostly with biologic/immunologic agents. So, either count would be bogus, and the sum of all counts will and should never match the total number of drug records in your data.

There are several things you can do to avoid the ridiculous overcounting or arbitrary numbers:

  • The drug classes actually know whether they, say, a topical antibiotic cream for the eye, or a systemic iv antibiotic. In the concept_class we smush this over. In other words, if the drug contains an antibiotic ingredient, it will get connected to all antibiotic classes disregarding these distinctions. We are fixing this as we speak. The next release will contain a corrected concept_ancestor table.
  • You could “improve” the classes. You could remove classes that have massive overlap, and only leave “clean” ones. That will reduce the overcounting, but not eliminate it.
  • You actually look at the indication. It’s not that hard. We have an indication class, and it is linked to Conditions concepts, and you could create a simple heuristic to make those choices.

Let me know if you need help with that.

#13

Ah, and drugs.com: They created their own proprietary taxonomy to run the website. I would not use that for your purpose. They don’t care about wrong counts at all.

#14

Hi @Christian_Reich, I appreciate you’re wanting to help, but

  1. Even though the source data contained the distinction between topical and oral etc., we lost most of that information when we went to the CDM. (Lots of drugs in Asia do not have corresponding clinical drugs in the vocab and had to be mapped to the ingredient level.)

  2. There’s at least one database where we only have drugs, no conditions, so looking up indications is not an option there.

I’m actually ok if there is overlap in the classes, I just want to avoid classifications that require information we currently do no have, like indication or even route of administration. Going back to the ibuprofen example, ‘NSAID’ would be a good class, but ‘cardiovascular’ wouldn’t because not all ibuprofen prescriptions are for cardiovascular use (in fact, most aren’t). In that sense the NDF-RT Established Pharmacologic Classes are ok, but there are just too many of them for our purposes and I need a higher level grouping.

#15

@schuemie:

Let me try to be a little more helpful than so far. :smile:

  1. We will fix that too, at least partially.
  2. Look: If you have only ingredient than you have no way of knowing which class you should use. Any one is equally correct. Or how would you know that “NSAID” is good and “Cardiovascular drug” is bad, if you haven’t any knowledge on to whom it was given and how?

Therapeutic class is discontinued, I believe.

Most common use - Martijn. The most common use of aspirin is a common cold. Then all your patients who take it for their aterosclerosis will be misclassified. Don’t touch that.

#16

@schuemie - You mentioned NDF-RT EPC (which is almost entirely driven by FDA SPL btw) being a bit too granular for your purposes – have you looked at NDF-RT Therapeutic Categories ?

Might be a bit more what you are looking for:

SELECT t.* FROM public.concept t 
--  we call them Therapeutic Class in CDM
--  NDFRT calls them Therapeutic Categories
WHERE concept_class_id = 'Therapeutic Class' and invalid_reason IS NULL

BioPortal Browser:
http://bioportal.bioontology.org/ontologies/NDFRT/?p=classes&conceptid=http%3A%2F%2Fpurl.bioontology.org%2Fontology%2FNDFRT%2FN0000178293#visualization

#17

Nevermind - failed to see you already looked at that.

#18

@herrcerd:

Yes, we call EPC “Pharmacologic Class”. It’s all in the vocabularies:

select * from concept where vocabulary_id='NDFRT' and concept_class_id = 'Pharmacologic Class'
t