OHDSI Home | Forums | Wiki | Github

Conventions on visit and observation _end_datetime for death

Hi all. I am implementing OHDSI for an intensive care database in the UK. From the ICU perspective, it makes sense to consider each hospital admission, a new “observation period”, with admissions to the ICU a new “visit occurrence”.

There is a logical sequence to the way in which an ICU visit could end. Please see the attached graphic for a detailed explanation. Broadly, after admission, you can either: be discharged alive, or die (of which a small percentage will be via brainstem death). My feeling is that the datetime of death (from whatever cause) should be mapped to VISIT_OCCURRENCE.visit_end_datetime as well as OBSERVATION_PERIOD. observation_period_end_date and PERSON.death_datetime.

I can think of no good analytical reasons to continue to declare the visit or observation period “open” after the confirmation of a death. It would also be hard to pin point a good source of information to determine these times beyond the time of death. However, I can’t see this convention explicitly written anywhere. I would appreciate any advice from the OHDSI community on the correct procedure here?

Many thanks,


Welcome to the community, @Doc_Ed!

There may be some data points, especially Measurements (Labs) and Conditions (cause of death, cancer diagnosis) that will be recorded after a Person dies. Lab cultures, autopsy reports, tumor identification, etc. If you end an Observation Period at the death datetime, then this data may be missed. At Colorado, we end the Observation Period with the last record for a Person. Others may do it differently. Many EHR data holders limit records to deathdate +60 days. Any records after that are thought to be an error. However, as you point out:

We as a community need to provide more guidance around this topic. You should join our EHR WG calls held every other Friday at 8am Mountain time, 3pm UK time (GMT +1). The next call is August 23rd. This would make for a good topic of discussion.

Regarding your diagram. An “inpatient visit” is the time a Person is admitted to the hospital until the Person is discharged. If the Person goes from ICU to general medical floor to discharge, all the different locations inside the hospital are part of one inpatient visit. If you want to differentiate the time a Person is in the ICU from the time they are on the general medical unit, please use the Visit Detail table. One record in the Visit Detail table will be for the ICU (this is a Care Site) and another record would be created for the general medical floor/unit/area of care (also a Care Site).

Does this answer your question? And did I create more questions?

Thanks Melanie. This has been extremely useful, thank you for helping me to understand. I’ll be stuck on a train in Europe for that next meeting, but shall endeavour to join the following one.

Thats a good point about results becoming available after a death.

To clarify, the whole inpatient stay in the hospital becomes a single visit in VISIT_OCCURANCE, with internal moves within the hospital (for example admission and discharge from ICU) recorded in VISIT_DETAIL? In which case, should the start and end datetimes in OBSERVATION_PERIOD be more symbolic rather than referring to events in the patient journey? For example, my cohort runs from 2014 to present day. So should observation_period_start_date be 01/01/2014 for everyone (since that is the theoretical time we could first see a patient) and then observation_period_start_date perhaps set to 60 days after death, or set to the present day?

I’m always happy to help :slight_smile: The complexity and ambiguity of EHR data coupled with the lack of clear conventions increase the complexity of ETL’ing to the OMOP CDM. If you send me your email address, I will add you to the EHR WG invite. It’s an open meeting, but I find it easier to have an invite on my calendar instead of digging through the forums or finding the meeting details on the OHDSI WG meeting page.


Well, there are many ways to skin a cat and many ways to ETL observational healthcare data to the OMOP CDM :slight_smile:

However, there are not any clear conventions on how one should set an Observation Period for EHR data. And I think it is time we decide as a community and write it down (@Christian_Reich I’ll add this to your ‘to do’ list since you’re working on the conventions now :slight_smile: ) In general, folks with EHR data set the start date of the Observation Period to the start date of the first event for the Person in your data. The end date is the end date for the last event in your data. UNLESS you know all the Persons in your data will always be seen at your hospital/clinics/care sites. Then you would start the Observation Period with their enrollment date or registration date and end with their exit date from your system (if you know they left & now receive care elsewhere) or the date of the last data from your EHR system (if you believe they would still seek care at your institution) or the date of the last clinical event (if they may have received care elsewhere). Here is some discussion on the topic. Documentation of the decision and reasoning behind it is important!


Excellent. Thanks Melanie.

I think I shall do as you advise here. Seems sensible to me!

I have revised the initial graphic for future reference. Hopefully this is a better reflection of the conventions as applied to an ICU centric cohort.

Hm. I am not following. The OBSERVATION_PERIOD is the time you capture data of the patient. The cohort is something you define and extract afterwards, and it is completely independent from when the patient is in your system. Or am I missing something?

Well, it is on his to-do, but it is not that easy. I agree we are not very prescriptive how to handle it in EHR systems. But that’s for a reason. Here it is:

The OBSERVATION_PERIOD is supposed to define the time during which we capture data for a patient, at least Visits, Conditions, Drugs and Procedures. And “capture” means the two axioms of observational data are true:

  1. If something happens we get to know it and have a record.
  2. If there is no record nothing happened.

We know this is not perfect: 1) is not reliable, but certainly 2) is anybody’s guess. But the question is how bad can we let it be? If we are picky we say “In a hospital EHR system we have no idea what happens to a patient outside the hospital, and we don’t even know if the patient didn’t visit a different hospital.” As a consequence, we can only make assumptions about the time the patient actually was hospitalized and EHR records were written. In this case, the VISIT_OCCURRENCE record becomes identical to the OBSERVATION_PERIOD record: This is bad, because whatever we see of a patient is only during the hospitalization time, and if we draw incidence reports we get a massively exaggerated picture. Alternatively, we say “Likelihood is high that the patient returns to “our” hospital if things go bad, plus we also record medical and prior treatment history from the patient.” In that case, we will create an Observation Period from the first Visit to the last, or even to the end of the database if we believe the patient is still with us.

Which one you pick depends on the assessment of the local situation. If in rural Colorado maybe people are loyal to “their” hospital and you can rely on a stable system like this. In this case you pick alternative #2. If not, and people are likely to shop around for different hospitals each time they have a problem you pick #1. OHDSI cannot prescribe which one. It’s a difficult decision, but it is yours, I think. In reality, most people opt for #2.

Makes sense?

You are of course right. I’m being a little imprecise with my description. 2014 is the starting date that we have permissions to collect data. So we don’t have anything before that. But yes, the study cohort is defined subsequently.

Thanks @Christian_Reich this seems to come back to a pragmatic choice, which is very reasonable. In the UK, given that our healthcare is nationalised, there is a good opportunity to learn if patients are attending different hospitals. Indeed, in my cohort, we can link to national data to understand when patients attend different hospitals that sit outside our research network. We don’t have any information regarding what happened inside the hospitals that sit outside our network (beyond a diagnosis), but at least we know if there have been attendances elsewhere.

So I think for my specific purposes (though this may not work for others due to the constraints on data capture that you highlight) OBSERVATION_PERIOD should start with the first appearance of a patient, and end either following death, or at the present day/last observation. My cohort is updated approximately every 3 months, so it might be easier to just set the OBSERVATION_PERIOD end to “today” (if the patient is thought to still be alive) to reduce the about of table updating I have to do with each data ingest.

For clarity, we don’t capture any community based events (GP attendances, pharmacy visits etc.) only hospital attendances.

Makes complete sense. I wasn’t clear and improperly used the word “convention”. Let me try again.

I believe the community should give “guidance” on Observation Period creation for EHR data. Something along the lines of what you wrote is good:

This is a topic that continues to come up. I believe written guidance would be helpful. Or we can continue to answer questions and give advice on the forums :slight_smile:

Use case similar to Doc_Ed.

Has there been a resolution to this?

We are building the ETL now and an entry in observation_period will correspond to the time period between the first and last data point in a patient.

Visit occurrences will be outpatient and inpatient encounters, and movement within hospital locations will be in visit_details. Is there a convention for a patient admitted to hospital A and transferred ICU to ICU to hospital B. Should this be one or two visit_occurrences?

Since the person is continuously receiving care, this would be one inpatient Visit Occurrence record and two Visit Detail records. Hospital A and Hospital B are Care Sites and movement between the two would be found in the Visit Detail records. Implementation of this idea might be tricky, there’s also an ambulance ride and possibly other transport between the two. I hope your data aren’t too ambiguous!

Very helpful. Makes sense. Thank you.