As discussed in the community call Sept 15, @jposada and I have been conceptualizing a study that would look at a commonly applied non pharmacologic therapy being applied in sick hospitalized non intubated COVID-19 patients: prone positioning. We’re hoping that other centers that are using their NLP table (or possibly have confidence in capture of proning within flowsheet documentation) would be interested in participating. We heard about interest from @scottduvall , @HuaXu , and @Andrew, and others.
Proning has precedent in an intensive care setting, with good evidence of efficacy in ARDS. There is widespread current application in covid-19 respiratory disease in the ICU. The invasive monitoring of patients in ICU leaves a lot of possible surrogate outcomes (and the data density would make this a challenging thing to study within the OHDSI Network, although possible to explore if there’s interest, and perhaps @scottduvall we could connect if you’d like to expand into that population with another study). Proning an intubated patient is also a lot of work in an ICU setting (and as a consequence, there is likely more of an obersvation / condition / procedure occurrence signal in that setting). We have started with a non intubated patient population here.
Less is known about the ultimate benefit of proning the non icu patient. It’s comparatively very easy and low cost, but also harder to detect in structured data. Anecdotal experience suggests that proning can have an impact on gas exchange in the short term. But there’s a risk that it could conceivably lead to harm. If the patient is running out of reserve on their back, they get flipped over, improve a bit, and run out of reserve on their stomach, you may be in a bad situation if you do need to ultimately intubate them.
We’d like to propose a study that examines the incidence of proning in the non intubated hospitalized patient within the OHDSI network. We’ve found there’s very little in the way of codified data / concepts that would allow the identification of proned patients as conditions or procedures, but have found, in Stanford data, that we do see these concepts captured in the NLP table. An alternative, in sites that have confidence in their flowsheet data, is that may often be captured there as well (nursing documentation practices may result in position documentation when vital signs are taken). Sites could also consider regex matching in their note table. @jposada has built out a modification to a cohort definition that substitutes in an NLP table query in the inclusion critieria, this is detailed below.
This study therefore sits at a nexus of a clinical question, and a question of technical feasibility in using the NLP concepts in its cohort definition in an network study. If we are able to detect proning concepts, we’ve also detailed a possible population level estimation question that could examine outcomes of intubation, and in-hospital cardiac / respiratory arres / death comparing prone and non proned non intubated patients.
We’ve prepared a feasibility cohort for those centers that are interested in participating. The definition is here:
There is a ‘placeholder’ for the NLP table query (in inclusion criteria #6). To implement an NLP table search, the following (param SQL) substitution would be made at “-- Begin Drug Exposure Criteria”:
-- Begin Drug Exposure Criteria select C.person_id, C.drug_exposure_id as event_id, C.drug_exposure_start_date as start_date, COALESCE(C.drug_exposure_end_date, DATEADD(day, 1, C.drug_exposure_start_date)) as end_date, C.drug_concept_id as TARGET_CONCEPT_ID, C.visit_occurrence_id, C.drug_exposure_start_date as sort_date from ( select de.note_nlp_source_concept_id as drug_concept_id, note.person_id, note.note_id as drug_exposure_id, note.note_date as drug_exposure_start_date, note.note_date as drug_exposure_end_date, note.visit_occurrence_id FROM @cdm_database_schema.NOTE_NLP de JOIN @cdm_database_schema.NOTE note ON note.note_id = de.note_id JOIN #Codesets codesets on ((de.note_nlp_source_concept_id = codesets.concept_id and codesets.codeset_id = 11)) ) C -- End Drug Exposure Criteria
The full .sql code is in the google drive below, as is a draft study protocol:
What we’d need from the community to start is:
- Feedback on the definition if something doesn’t look right.
- The protocol is acknowledged as incomplete, but happy to take in feeback on the document.
- A feasibility test of the definition, with the NLP table search (presumed needed), to see what kind of counts you’d have at your centre. You can use the SQL provided in the google drive, translated to your flavour using SQL render as a starting point.
Thanks for any thoughts!