OHDSI Home | Forums | Wiki | Github

Prescribed and Dispensed Drugs in the DRUG_EXPOSURE table

(Marek Oja) #1


Hope that I don’t duplicate any topic.

We at STACC OÜ are currently mapping Estonian prescriptions data to OMOP CDM with the aim to map and combine three different data sources (prescriptions, claims and EHR records) in a pilot project.

Questions are related to prescription data. It’s an independent data-source, meaning there is no direct link to any inpatient or outpatient visit.

The prescriptions data-source contains:

  1. prescribing information – the date prescription was written, information about prescribed drug, provider, patient’s condition (i.e. diagnosis), etc.

  2. dispensing information – the date the drug was dispensed, information about dispensed drug, pharmacy, etc.

We would like to store in OMOP CDM both, prescribing information as well as dispensing information.

Note, that by our national regulations prescriptions can be valid up to 180 days (in some cases even more). Thus, the gap between prescription written and drug dispensed may be long, or sometimes the medication is not dispensed at all. We also have cases where we do not know if the drug will be dispensed or not (i.e. the prescription is valid, but the person has not purchased the drug yet).

The questions that we are currently struggling with:

  1. Is it OK that persons that have information about prescribing as well as dispensing can have basically two records in the DRUG_EXPOSURE table, i.e. one record about prescribing (with drug_type_concept_id ‘‘38000177 Prescription written’’) and one record about dispensing (with drug_type_concept_id ‘‘38000175 Prescription dispensed in pharmacy’’)?

If yes, then:

  1. Is it possible to maintain the connection in OMOP CDM between prescribed drug and dispensed drug?

  2. in case of a prescribing record, is it correct to set drug_exposure_start_date as a date on which the prescription was written? Even if we know based on dispensing information that drug utilization started later or the drug was never purchased?

If yes, then:

  1. doesn’t that cause misleading information in the DRUG_ERA table? I mean if the time between prescription written and drug dispensed is longer than the standard Persistence Window of 30 days it could lead to multiple drug eras instead of one?

  2. When the drug was not dispensed during the prescription validity period (person didn’t buy the drug), should the prescribing information added to the DRUG_EXPOSURE table although we know that the person didn’t took the drug prescribed with this prescription.

Any help would be appreciated.

Modelling treatment plans / care plans / intention to treat
(Don Torok) #2

Is it OK to have two record, one for the prescription being written and a second for the drug being dispensed?
No. The assumption is that a record in the drug table is the best estimate that a person received treatment. You already identified the problem, having records for both prescribing and dispensing will give misleading results. From your description, you will get a notification when the drug is dispensed, that is what should be recorded in Drug Exposure.

Your use case:
We would like to store in OMOP CDM both, prescribing information as well as dispensing information.’
is not something the OMOP CDM is designed to handle. You would not be the first to request an Orders table to hold this type of information. Also, it is not a violation of OHDSI standards to add columns or a table for local needs not supported by the CDM, so you could add a Prescription table.

The Fact_Relationship table is for maintaining one to one relationships between entities in the CDM.

(Melanie Philofsky) #3

I respectfully disagree, @DTorok, with this statement

Using the drug_type_concept_id in your SQL query or in your Atlas cohort definition will help you distinguish the prescription from the dispense record


Yes, using the type_concept_id distinguishes the record.

As Don said, use the Fact Relationship table to link the records.

Correct. You know when the Drug was ordered. And I assume you know when it was dispensed. The order record should have the date it was ordered and the dispensed record should have the date the Drug was purchased/dispensed.


You should ask in the “Researchers” forum about the “misleading” information in the DRUG_ERA table when both prescriptions and dispensed Drugs are in the Drug Exposure table. I would think you would only want the dispensed drugs to be used in the derivation of the Drug Era table. But the researchers will be able to provide more information on the “consequences” of a decision.

(Chris Knoll) #4

I would like to respectfully disagree with your disagreement :slight_smile: The information in the DRUG_EXPOSURE table should maintain information about what the patient experienced, not what a healthcare provider would ‘like to have the patient experience’. This is just my humble opinion, of course. I cant’ think of a clinical use-case where you’d need to distinguish the intent of administering a medication vs. the actual administration of the record: would it be to evaluate the performace of a drug fulfillment system? If so, isn’t the CDM oriented around patient performance, and not system fulfillment performance?

While I do recognize that we have a type concepts that we can try to dis-entangle orders from administrations, the drug exposure table is, at it’s heart, a record of what the patient experienced, right? I agree with @DTorok that if we want to represent orders separately from exposures, we should have a drug_order table.

In addition, I’m not aware of any special logic in the standardized drug era builder about distinguishing prescriptions written from drugs dispensed when calculating the error. If that’s something that is supposed to happen, there should be an issue opened up on it.

(Joe Pallas) #5

The problem with insisting that we only include what the patient experienced is one of epistemology. We generally don’t know what the patient actually experienced unless we have a record of medication being administered, which is often available in the EMR for inpatients, but for outpatients won’t include the vast majority of medications.

Did the patient fill a prescription and take it as prescribed? We don’t know. At best, there might be a pharmacy record that the prescription was filled/refilled, regular requests for renewals, and a patient report that it was taken as prescribed. Many EMRs will be lacking some or all of that.

Studies are done with the assumption that outpatients take drugs as prescribed unless there is evidence to the contrary. The question would be when and how much additional value comes from recording additional information if it is available.

(Sulev Reisberg) #6

I would like to ask whether @Patrick_Ryan has a view on this. From researcher’s perspective it really depend on the question - if you want to be sure that the patient used the drug, simply having a prescription might be not sufficient. However, it is completely sufficient to remove that patient from the control group. However, i think in most of the cases the fill data is not available and simply prescriptions are added to “drug exposure” table (yes, the word “exposure” is a bit misleading then). But in case both prescription and fill data is available, it is a question how to add them both to OMOP.

(Andrew Williams) #7

Patrick will have the better answer…

but here’s my 2 cents: It seems to me that decisions about this should be based on their impact on the measures of drug exposure, adherence, non-adherence, etc. that will be derived from the underlying representation of medication data. I’m not an expert on that and haven’t stayed current with this field. But years ago when I did a little work that required these decisions I was impressed by the thoughtfulness of the work by Marsha Raebel and colleagues on terminology and related operational definitions. It was based on useful prior systematic literature review different adherence definitions and an algorithm that improved the accuracy of adherence group assignment as well as the criteria used to select patients. They also did careful empirical work on the % of orders that result in fills and the % of fills that result in dispenses. The drop-off through that cascade was depressingly large, even at an integrated health care system with its own pharmacy that was conveniently co-located at the care site and offered reduced prices relative to external pharmacies. Unfortunately, I can’t find the publication showing exactly how large those drop-offs were, but my memory is of something huge like 30% going from order to dispense.

They seemed to make a sound case for definitions of persistence, primary and secondary non-adherence, etc. that were as accurate as possible given what is available in the source data.

But the increased accuracy from not assuming orders lead to dispenses is only one consideration. This study by Bill Vollmer and my other old colleagues who taught me about all this, examined the bias in pharmacy-based measures. They found an upward bias in adherence estimates in measures that require at least one dispensing which was increased when measures require a second dispensing event, and finally that these biases are greatest with shorter observation times.

With respect to the trade-offs between the accuracy, bias, and generalizability of measures across data sources, and how those are contingent on source data attributes, they concluded:

… requiring a dispensing to be calculated meant that these measures could not be defined for large numbers of individuals (17-32 % of participants in this study). Measurement strategies that do not require a dispensing event to be calculated appear least vulnerable to these biases and can be calculated for everyone. However they do require additional assumptions and data (e.g., pre-intervention dispensing data) to support their validity.

If this picture of the trade-offs is still relevant, this seems like a great opportunity for the OHDSI community to do some empirical work that investigates the impact of these trade-offs on the measures that the community will use in it’s studies. The OHDSI community’s rich resources in data, data modeling experts, methods development and methods evaluation experts, pharmacoepi experts is pretty ideal for that. With the right combination of those assets an OHDSI group could build on this prior work and come up with best practices for:

  • ETLing the data at sites with different levels of access to dispenses

  • Best practices algorithms that optimize the impact of ETLed data on measures of exposure, adherence, primary and secondary non adherence, etc. derived from the underlying data.

But given the fact that the OHDSI community has all those things, chances are good that these issues are already understood and accounted for in ways I should know about but don’t.

(Don Torok) #8

What does a row in Drug Exposure mean? I believe the assumption is that a row in drug exposure is the best estimate that a person was exposed to a treatment. When ETLing an EMR that did not have pharmacy claim records, the best estimate that a person was exposed to a treatment, other than direct administration, was the Rx order, and the drug type was ‘Prescription Written’ giving the analyst an indication of the level of confidence of the ETL. Yes you can use the Drug Type Concept to distinguish between Orders and Fills, yes adherence , non-adherence and worth studying. But if you are participating in a group study, what is the assumption of cohort for rows in Drug Exposure? Does the cohort check the Drug Type Concept? Should it?

(Alan Goldhammer) #9

How do we deal with drugs that are available without prescription? For COVID-19 there are a number of these drugs that may or may not offer some benefit such as famotidine, naproxen, ibuprofen, and ceterazine, to site a few. Unless the patient reports the use of these drugs, there may or may not be a record. I’m speaking from a US perspective where EMR records might be more problematic.

(Chris Knoll) #10

@jpallas I think you’re describing the challenge of ETLing the data to a standard form…and that’st he problem with standards, you have to forge your source data to follow the CDM standard. So, everythign you say is true, but at the end of the road, your data should be the best representation you can muster to represent the patient experience. If that means in some cases you use the orders, then that’st he best you can do. But I’m just disagreeing with the notion that we put data in drug_exposure that isn’t actual drug exposures.

(Chris Knoll) #11

From this description of the drug_exposure table, I think it’s not just misleading, it’s wrong:

The ‘Drug’ domain captures records about the utilization of a Drug when ingested or otherwise introduced into the body.

Later, they do reference orders:

Drug Exposure is inferred from clinical events associated with orders, prescriptions written, pharmacy dispensing, procedural administrations, and other patient-reported information

I would argue that this statement is not saying ‘drug exposures also is the record of the order’, I think they are saying 'you can infer a patent was exposed to a drug from an order for such drug if you have to.`

For @Marek_Oja case: I would make the ETL smart enough that if an order is recorded, but a subsequent dispensing is recorded, just put the dispensing in using the dispense date. If you can’t find the dispense, you have the choice of ignoring the event or record it as a drug exposure. But whatever the decision is, when it gets to drug exposure, we should interpret it as the user was exposed.

(Vojtech Huser) #12

Just a note about dispensation:
I may have picked up a drug at pharmacy (e.g., emergency inhaler for asthma) but it does not mean that I started using it the day I picked it up. I may have a different drug stockpiled and may start using my prescription later. So dispensation is also not what the patient experienced (ingesting the drug). Both prescription and dispensations are type of approximations to patient reality.

(Chris Knoll) #13

Right, that’s what we’re saying, we have to make our best guess. But the guess is about the exposure, not that someone made an order.

(Christian Reich) #14


Not sure what the discussion is about. We have drug_type_concept_id which tells you if this is a pharmacy dispensing, prescription or administration record. So, the CDM is innocent.

(Chris Knoll) #15

The discussion is about what a record in the drug_exposure table represents. It’s not clear what you are asserting above: you could say that the drug_type_concetpt_id is the type of information used to infer the drug exposure. However, the main point is: should the presence of the record in drug_exposure mean that the drug was simply ordered for the person, or does the record in drug_exposure mean that we should assume that the person had the drug administered.

(Christian Reich) #16

Hm. Not sure I follow. We know from the Type Concept what kind of evidence we have. In the absence of the precise knowledge of whether or not the drug made it into the organism we can only make an assumption based on that Type. The probability of drug exposure for ordering is lower than dispensing is lower than administration is lower than 100%. The decision whether or not we deem the probability high enough for an analytical use case is up to the researcher.

(Alexandra Orlova) #17

I think the question is about “duplicates” in source data: Physician had prescribed drug A to patient 1 and after that, patient 1 went to a pharmacy and drug A was dispensed.
In the data set we will have 2 records:
2020-06-01 - drugA - patient1 - Prescription written
2020-06-03 - drugA - patient1 - Prescription dispensed in pharmacy
And the question is should we store both records in CDM with different type_concepts or we shouldn’t store prescriptions if they have corresponding dispenses in order to avoid duplicates in CDM

(Chris Knoll) #18

Going a bit further: should we record information that we know isn’t a patient exposure to a drug: in @Alexandra_Orlova example: we can say the person couldn’t have taken the drug at 6/1 because the prescription was only dispensed on 6/3. In some databases, the prescription written is all we have, and so we will write that record into the drug_exposure record. but should we writing records to the drug exposure table for events that we know is not an actual drug exposure for purposes of doing the sort of analysis where you want to evaluate the orders for drugs vs. the administration of the drug…based on what we understand the purpose of the DRUG_EXPOSURE table is for (and we are all waiting with baited breath to hear an official ruling on this), you store the events that you can reasonably infer that the patient was exposed to the drug.

From my perspective, this isn’t about ‘should orders be written to the drug_exposure table’, this is about ‘should records in the drug_exposure represent a person’s experience of receiving a drug’?

(Christian Reich) #19

@Chris_Knoll: I agree.
@Alexandra_Orlova: So, put it into the data. Record the right Type Concepts. Declare victory.

Only few data assets have both information. You seem to be working on one of those. It actually enables asking a scientific question “What portion of the patients fill the prescription?” and “How quickly are they filling the prescription”. For the other use cases, where people are just looking for indication of exposure, let them pick their logic.

(Marek Oja) #20

Thank you for the answers and discussion.

@Alexandra_Orlova you described one of the questions we had and thank you for the example.
To have complete examples I add the example from her post also.

Basically we have three broad categories of prescribing and dispensing information.

  1. 2020-06-01 - drugA - patient1 - Prescription written
    2020-06-03 - drugA - patient1 - Prescription dispensed in pharmacy
    Of course we don’t know if the patient ever took the drug or not. Or when he started to take the drug. At least we know that patient had the drug at home.

  2. 2020-01-01 - drugA - patient2 - Prescription written
    2020-06-01 - the prescription validity ended and the patient didn’t buy it
    In this case we can be quite sure that the patient didn’t get the prescribed drug with this prescription. The question was we should store the doctors indent that the patient should take the drug although most likely the patient didn’t take the drug.

  3. 2020-06-01 - drugA - patient3 - Prescription written (and it is valid for e.g. 6 months)
    As the prescription is still valid then we don’t know if the patient will buy it or not. I think this describes the prescription in most of the datasets where it is only known that the prescription was written but nothing about the fulfilment of the prescription.

Reading all the comments I think in summary we can take one of two approaches:

If we take the recommendation and we map everything and distinguish it with right Type Concept.
Of course we should record also connection between prescribing and dispensing the drug. As recommended by @DTorok and @MPhilofsky we could use the Fact Relationship table but looking at the standard Relationship Concepts then there aren’t any suitable concepts to link these two facts. Most likely nobody else hasn’t had the need for this kind of relationship.

If we take approach of that drug_exposure table should have only records where patient was exposed to the drug (as well as we can get it from the data) then in our case we should use only prescription which have also dispensed (example 1). Also maybe “example 3” prescriptions which have not yet dispensed or cancelled - it is likely that at least some of these prescriptions will be dispensed.