Meeting minutes
ACTION: DBooth to put full zoom info on RDF mtg page, for dial-in
Introductions
Erich: dep of informatics, working on RDF 15 years, working on software for teachers, bridging to clinical . Stonybrook Univ.
Relating imaging data to clinical info
erich: Tagging onto OMOP group and, diacon group. Trying to integrate spacial info, and make heatmaps.
… I'm using RDF, but want to walk it to clinical info also.
… Any work done w diacom or OMOP?
eric: Some of the same people worked on an OMOP mapping a while ago.
… Didn't do anything between images and OMOP. Not aware of images to OMOP. Maybe Scott Marshall?
erich: Imaging and clinical domains seem to be separated. Want to bring them together.
… OMOP is tyring to use Daicom verbiage.
… Have annotations by the files are usable, but cannot bring up 3M image iunless there's spacial info.
… Cannot decode the image and put it on the screen quickly. Need spacial info.
… Same applies to deep learning pipelines.
eric: Write-up of baby use case for clinical setting?
erich: Not aware. Also work w cancer imaging archive. They get a lot of images submitted, and de-ided clincial info. People now want both.
… Diacom is noce, but unless you set standards of what should be included, you get a lot of empty boxes.
… Noe of these groups are bridgeing each other.
dbooth: Need to get folks together to duscuss?
erich: David Kume said thatere;s some diacom, but no official mapping. Series of XML docs, you need to extract info from them.
… FHIR is better, OMOP has CSV which is better. Need to convince people that this is needed.
… Easier to build something to show them the value, to get them hungry to discuss.
… Otherwise OMOP group has 100 people.
… Any image could have 1M segmentations, with lots of info associated.
… Cannot use standard RDF DB -- billions of triples.
… Need to break into separate storage, then need to get clinical date into it.
… In semweb, lots of greag components, but nobody knows what it becomes yet.
tim: little familar w diacom. Thought diacom modalities didn't seem they had depth. Have you looked at relevant FHIR resources? Visionoing more detailed modeling of those modalidties in FHIR?
erich: Looking for low hanging fruite. Collected a lot of COVID imagery 80-100M radiology images.
… Wrote software to extract diacom data and RDF-ized it. Loaded into virtuoso to query.
… Diacom allows linking by accession IDs, dates, etc.
… Brought into relational data laek, but how to connect them back? Some terms connect, but not all.
… Stonybrook purchansing scanners to scan. Say we'll need to decide what to collect on patients, etc. Otherwise you just have a big [pile of images.
dbooth: Maybe show what you have so far?
erich: Could show soon as use case. ONe odball thing: images range from 300M to 3GB, and my computer gets chewed up.
… Having the pipeline data from deep learning, in central RDF DB is not scalabl, virtuoso w TB of RAm.
… With a lot of these files, putting into central store, you're doing spacial indexing, but considering it all part of the same space.
… I partitioned out the spacial info into separate files. Originally going to use RDF HDT in separate files.
… But some pieces were missing. Did something similar: Apache jena graph, backed by apache arrow.
… RDF mapped to chunks. Predicates being loaded separately.
… That allowed performant viewing.
… There all in their separate worlds.
… Could show you in a couple of weeks. Was using W3C opena nnoatations. But recently changes to geoSPARQL.
… Creating feature collections and specifying other data, heatmaps and segmentation.
… published a paper in 2019, 7k images of several cancer types.
… Did a preload to see how it will work.
allen: https://
erich: preprint: https://
Fixing RDF examples
tim: Jim and I have PRs open that fix a lot of example. I guess I need to post in zulip.
eric: Recently working on something that ports between FHIR JSON and FHIR Turtle, to be able to compare the graph and see if they are the same.
dbooth; Would that be usable in the RDF playground for validation?
eric: No, different code base. Using it for AMIA tutorial.
… Wanted to test against official examples and see where we differ.
… Goal in playground was to not be tied to FHIR resources, to allow us of other logicla models.
… Current work depends on standard FHIR core.
https://
jim: Want to be able to gen one FHIR RDF example at a time
ACTION: tim to make github issue about gen one FHIR RDF exmape at a time
Elment summary
tim: Need a way to connect a Patient class to the StructureDefinition for Patient, for example.
… Suggestion was to put in the class definition in the FHIR RDF ontollgy.
tim: IN RDF we have the distinction bewteen the class and the StructureDef. Need to tie them together.
… Then I need to know which properties are included in the resource summary. That info is in the StructureDef but not in the class.
… That link would them together.
eric: Sounds like the right way to do it.
dbooth: What exactly would the URL be?
eric: We'd need to work on the RDF-ization of the StructreDefs. Algorithm might work out of the box, but not guaranteed.
… No sure they'd be dereferenceable, but we don't control that.
eric: Big prob w examples is that an overhaul to make them resolvable would give you a FHIR server. That would be a very helpful.
… .Would be a great opportunity to enable follow-your-nose.
ACTION: Eric to make github issue about making FHIR examples dereferenceable
tim: This looks like the right URL https://
tim: Maybe have prov:isDefinedBy https://
tim; That's an annotation property.
ACTION: Tim to show example of linking class to structure def, and also show how an objectProperty could be used instead
ADOURNED