Stream: implementers
Topic: Mimic Data set
Grahame Grieve (Nov 07 2019 at 09:58):
If anyone is interested, I've been working with the maintainers of the Mimic Dataset, a set of real ICU records that have been de-identified. We're going to build a mimic-> FHIR conversion into their pipeline. The code to do the conversion is in HAPI (), and I can share the publicly available set converted to FHIR (R3 or R4). It's 100 patients, 129 ICU admissions, 10k prescriptions, nearly 1million observations
Grahame Grieve (Nov 07 2019 at 10:29):
see https://github.com/hapifhir/org.hl7.fhir.core/blob/master/org.hl7.fhir.r4/src/main/java/org/hl7/fhir/r4/importers/Mimic14Importer.java for source
Michele Mottini (Nov 07 2019 at 13:50):
We would be interested in the data set - either R3 or R4 is good
Paul Church (Nov 07 2019 at 16:17):
I've worked with MIMIC3 a lot (mostly on the ETL to OMOP), it's a fantastic resource. The 100 patients is just the fully public "demo" dataset of patients known to be deceased and manually de-identified to a higher standard - the full dataset is 40k patients, 60k admissions, and hundreds of millions of observations!
Grahame Grieve (Nov 07 2019 at 19:50):
the full data set will be pretty large on fhir....
Jason Walonoski (Nov 07 2019 at 21:51):
Not MIMIC, but if you need a large dataset (1K or 1M patients) you can try Synthea downloads: https://synthea.mitre.org/downloads
Darrell Woelk (Nov 09 2019 at 01:29):
I am advising some grad students from the Dell Medical School at the University of Texas who are planning to start a project to map MIMIC III data to FHIR. Do you need some help? What is the status of your importer in github? Is there a public link to the R4 FHIR database you have created with MIMIC III data?
Grahame Grieve (Nov 09 2019 at 02:34):
http://mimic.fhir.org/r4 and http://mimic.fhir.org/r3 but they are still loading - a long way to go
Grahame Grieve (Nov 09 2019 at 02:37):
and they are also available at http://www.healthintersections.com.au/mimic-fhir-r3.zip and http://www.healthintersections.com.au/mimic-fhir-r4.zip
Grahame Grieve (Nov 09 2019 at 02:37):
(in about 10min)
Grahame Grieve (Nov 09 2019 at 02:38):
as far as status - I've done as much as I need right now; I need to clean up the LOINC coding part, do more LOINC coding, and import some more tables
Grahame Grieve (Nov 09 2019 at 02:40):
contributions to the code are very welcome....
Grahame Grieve (Nov 09 2019 at 02:40):
https://github.com/hapifhir/org.hl7.fhir.core/blob/master/org.hl7.fhir.r4/src/main/java/org/hl7/fhir/r4/importers/Mimic14Importer.java and https://github.com/hapifhir/org.hl7.fhir.core/blob/master/org.hl7.fhir.r3/src/main/java/org/hl7/fhir/r3/importers/Mimic14Importer.java
John Silva (Nov 09 2019 at 13:02):
@Grahame Grieve @David Hay - it might be nice to use the clinFHIR interface to access this Mimic dataset?
David Hay (Nov 09 2019 at 17:16):
Well, once the files are loaded it should just be a matter of adding the server in the CF front page (as the data server) and loading the viewer. Currently there are no Patient resources, and it will be interesting (!) to see how it handles the volume!
Grahame Grieve (Nov 09 2019 at 20:00):
I'll let you know when I've finished loading
David Hay (Nov 09 2019 at 20:34):
ta
Grahame Grieve (Nov 11 2019 at 04:53):
The R4 version of Mimic has finished uploading at http://mimic.fhir.org/r4
David Hay (Nov 11 2019 at 05:37):
and I'm afraid to say that the dataset for a single patient is too large to load (at least according to the architecture I currently use)...
Grahame Grieve (Nov 11 2019 at 05:37):
yep. very large. Will be a good opportunity for me to work on scale - but you'll need the bulk data interface for 10k+ observations
John Silva (Nov 11 2019 at 14:40):
@David Hay as you mentioned a week or so ago, there probably needs to be some architectural changes to do incremental loading of datasets as large as this.
David Hay (Nov 11 2019 at 19:07):
Well it's not just that - what would the display look like? representing the resources as nodes with references to other nodes will be meaningless (imagine what a Patient centred graph would look like!) You could display observations as charts for a start I guess. Other options would be to use the bulk data handling mechanism to get the data into a local more optimized store - or maybe GraphQL to restrict what you get back.
But certainly the current strategy of retrieving everything and then manipulating it on the client isn't going to fly. Any thoughts on what a useful UI would look like?
René Spronk (Nov 12 2019 at 07:12):
Timeline, or problem oriented approach?
Grahame Grieve (Nov 21 2019 at 09:01):
For those interested, James and I moved the MimicImporter code to it's own repository, so that we can grant write access to it without giving access to the entire Java code base.
Grahame Grieve (Nov 21 2019 at 09:01):
https://github.com/FHIR/MimicImporter
Darrell Woelk (Jan 16 2020 at 21:30):
Graham, I am having trouble accessing the MIMIC FHIR data at http://mimic.fhir.org/r4. Is that FHIR server still up?
Grahame Grieve (Jan 17 2020 at 00:29):
no I took it down (costing me too much). I'll work on bringing it back....
Leopold Franz (Mar 26 2020 at 11:52):
Hello everyone,
I have a technical question about how I can transform the MIMIC-III dataset into the FHIR format. You can find an overview of the MIMIC-III dataset tables here: https://mimic.physionet.org/mimictables/.
I currently have mapped the MIMIC-III tables to FHIR resources in the following manner:
mimicTables_fhirResourcses.png
I still have one problem to map the icustays table to a fhir resource. Information about this table can be found here: https://mimic.physionet.org/mimictables/icustays/. I would like to keep this table as several items in other tables are linked to certain ICU stays. My plan was to link these items by also adding the icustay_id to the resources partOf attribute. What do you think of this approach?
Thank you very much for your help.
Best regards,
Leopold Franz
Lloyd McKenzie (Mar 26 2020 at 14:19):
Would it not make sense to represent the icustay as an Encounter?
Leopold Franz (Mar 26 2020 at 16:39):
Intuitively I would have said no but going over it again I can just specify this icu encounter to be part of an admission encounter. I now need to make sure however that the identifiers don't overlap as that could cause major issues. Thank you for your help!
Last updated: Apr 12 2022 at 19:14 UTC
