FHIR Chat · FHIR + NLP/OCR/AI document analysis BoF

Stream: implementers

Topic: FHIR + NLP/OCR/AI document analysis BoF

Simone Heckmann (Nov 14 2018 at 09:26):

We‘re planning on doing a BoF at DevDays about how to link automatically generated FHIR Data from unstructured documents to their source and all the associated questions and problems around this use case...

Who‘s interested...?

Jim Steel (Nov 14 2018 at 09:26):

:hand:

Simone Heckmann (Nov 14 2018 at 09:32):

@Patrik Sundberg @dr Kai U. Heitmann @Gregorio Canal ...?

Simone Heckmann (Nov 14 2018 at 09:32):

@Josh Mandel

Simone Heckmann (Nov 14 2018 at 10:10):

I have a Chalet booked for tomorrow 11:55-12:35, hope that works for everyone....

Veliyan Georgiev (Nov 14 2018 at 10:36):

Gregorio Canal (Nov 14 2018 at 10:56):

:thumbs_up:

René Spronk (Nov 14 2018 at 13:07):

Simone Heckmann (Nov 14 2018 at 13:41):

One thing I'd like to discuss is the question on how to use/extend provenance to be able to point a piece of information (e.g. a Condition) created through text analysis (e.g. NLP) to a specific part, point or word in a document.
Probably, capturing the level of certainty with which information has been recognized may be a challenge.
What additional points would you like to discuss?
@Josh Mandel : Can you provide a link to the wiki page, where you've been collecting ideas/issues around this?

Veliyan Georgiev (Nov 14 2018 at 14:48):

@Simone Heckmann I second that topic but I might be a bit behind. Just in general - I’m curious if NLP-based decomposition of a document into FHIR resources is the right way to go as a first pass. I’d like to get the group thoughts on generic _text searches. My concern (which I hope I’m just missing something) is that it gets progressively more difficult to pick the proper FHIR resources from NLP when we start talking about things like MedicationRequest, MedicationAdminstration etc.

Simone Heckmann (Nov 14 2018 at 14:55):

I think, the disambiguation between Medication*-Ressources is a general problem, not just for NLP :)
As a rule of thumb I would suggest to always use MedicationStatement, unless the intent of "request" or "administration" can be clearly determined by NLP...

Yunwei Wang (Nov 14 2018 at 15:00):

Could not go to Dev Days. :cry:
Can we have the same topic in January's WGM?

Simone Heckmann (Nov 14 2018 at 15:24):

We sure can! (unless we solve all our problems in tomorrows session) :laughing:

John Moehrke (Nov 14 2018 at 15:55):

so the idea is an extension on Provenance.entity that can hold a region identification? (characters 99-118)

John Moehrke (Nov 14 2018 at 15:58):

as in: Provenance.target is a Observation that was derived from a text document, character region 99-118. So Provenance.entity.what points at a DocumentReference that describes the text document, and DocumentReference.content.attachment.url is a Binary which is the text document. Thus the extension needs to be clear it is a character region within the attachment Binary...