Stream: Orders and Observation WG
Topic: AI Algorithm as performer
Diana_Ovelgoenne (Oct 11 2021 at 11:22):
Hi, I need to profile an Observation where the perfomer is an AI Algorithm and I don't find my ways around it. I thought about using Device, but from a workflow perspective indeed the legal performer is the Algorithm itself which runs automatically through all data generated and wasn't ordered to be observed by a practitioner. Do you have any suggestions how to resolve this?
Mareike Przysucha (Oct 11 2021 at 11:24):
A similar question was discussed here. Did they come to a solution?
Diana_Ovelgoenne (Oct 11 2021 at 11:43):
@Hans Buitendijk what would be your recommendation?
Lloyd McKenzie (Oct 11 2021 at 13:20):
I would suggest submitting a change request to Observation asking for Device to be supported as an Observation.performer. It's absolutely possible for the device to be the sole generator of the information, and it's a different semantic from "device used by performer to make the measurement".
Hans Buitendijk (Oct 11 2021 at 21:10):
I put in a comment on the other thread, suggesting that considering a Device as a .performer choice is not appropriate. We have to consider that the performer is about responsibility, while the device is not. The device was used to create the result. The analogy with Lab would be very informative to consider where the analyzer does not become a performer. The medical director would be documented as such. The question is therefore for AI software as to who the medical provider is who owns managing and maintaining the software that yielded the results, ensure calibration, etc., etc. and when the results are inaccurate is held accountable. Whether an organization or person, that is the .performer. All else is a .device.
Lloyd McKenzie (Oct 11 2021 at 21:43):
We allow Device to be responsible in other places. A Device can perform procedures, assert Conditions, author QuestionnaireResponses and do various other things. Saying they can't perform Observations is inconsistent with a lot of other FHIR resources.
Hans Buitendijk (Oct 12 2021 at 22:11):
Or other resources are inconsistent with Observation, so a good discussion to be had using the JIRA to help drive that to establish a consistent approach to capture the participation of a device and the responsibility of a person and/or organizaiton.
Lloyd McKenzie (Oct 13 2021 at 02:41):
Probably best place to have the conversation around consistency is the workflow project. The patterns currently recognize Device as a Request.author and an Event.performer.
Diana_Ovelgoenne (Oct 14 2021 at 11:59):
From the perspective of mapping DICOM SR to FHIR, on DICOM exists as an Observer either a Person or a Device. It is not needed that when there is a Device, there is a Person behind. If the Device creates the corresponding observations, there should be an asserter if at all, who can confirm whether the measurements created by the device are correct or not, still doesn't mean that those measurements were created by the Person.
Lloyd McKenzie (Oct 14 2021 at 12:26):
I think you meant Practitioner rather than Person? (Person is only allowed as a linking resource, never as a subject or other participant)
Diana_Ovelgoenne (Oct 14 2021 at 13:11):
in DICOM is called Person Observer, but yes, I mean the Practitioner on a FHIR lexicon
Andrea Pitkus, PhD, MLS(ASCP)CM, CSM (Dec 23 2021 at 03:50):
Lab data in the LIS has used autoverification methods for over 20 years. Most LISs indicate lab results are verified by the specific instrument upon which they are performed (due to accreditation requirements). In FHIR, would expect the performer of the result/observation would be the unique identifier for the instrument/device to meet accreditation requirements. Suspect a similar approach can be taken with other "AI" especially in cases where FDA regulates them.
Last updated: Apr 12 2022 at 19:14 UTC
