Meeting minutes
Introductions
Darrell Woelk: involved in semantic world, going back to MCC. Then worked on Carno, using Cyc. The started Info Sloop, Made a company in 1995, then got into healthcare, building a soft dev platform in healthcare Started w FHIR in 2015. Green Room Tech is our latest. I do FHIR consulting, both tech and business. When Chat GTP cam out, I started working on AI again. Ivolved w U Texax Information, doing lectures on FHIR, then last hyear on gene
rative AI. Had 4 gra dstudents working on gen AI, and want to do more this year.
Darrell's ideas on student involvement.
Darrell's slides: https://
darrell: Working w Dr Ding.
… KG and LLM.
… Spoke w Juan Sequeda in Data.world. They did soft dev and pub an article about KG and LLM bencharks.
… It can gen SQL, using SQL DL.
… Doing the same thing w an ont and SPARQL
… Juan did a mapping into DB
… They had a claims ont.
… data.world executes the SPARQL queries.
… They turn those SPARQL queries into SQL and return answers.
… They did a benchmark, comparing SQL-only-based vs ont-informed. It was 50% better, and they showed why.
… They used zero-shot prompt.
… They made this available for others to try.
… Proposing to do this in the healthcare domain. On the left side, that meansw we don't have to tell chatGPT what the FHIR resources are, becuase it already knows that.
… Use a FHIR API earch instead of SQL.
… On the RHS want to do something similar to what they did w insurance claim, but w phenotype ont.
… Then using FHIR RDF data.
… Or could export all my FHIR RDF data into the triplestore.
eric: Sounds fun. There might be some code that you could reuse, that does part of that.
… SPARQL query gets dismantled to match hierarchies agains a library of FHIR path. Rule head would be FHIR Path, body would be SPARQL.
… Divide up resources and test it against know FHIR paths. Launch , get results back, bind appropriate resources.
(Eric explains the code more)
eric: That's in typescript right now, Claude Nanjo and I worked on it.
eric: There's a shex schema for FHIR.
dbooth: And a FHIR ont.
darrell: Phenopackets include a lot of data. Vulcan project trying to come up w FHIR rep of phenopacket.
… Met w them yuesterday.
jim: HPO folks have a transformation to FHIR ont.
… They're at least workin on it.
… Someone who works for Chris Chute working on it.
darrell: These are grad students. Class is AI and health. Want a group of 2-3 students. How to intro FHIR RDF?
https://
dbooth: That hasn't been updated to R5 yet.
eric: Did the HPI queries assume that the closures implied by the ont were already instantiated in the SQL data?
darrell: IDK.
eric: could either assume no inference, or could add subsumptions in tables, and join those also.
… Or you could transform the query to add the subsumptions.
darrell: They product is a catalog. Guessing that they're not pushing on inference, but I'm interested in it.
… This would be a first step toward building a test system.
jim: I think the work by Joe Flack on ontology-to-FHIR has been folded into OAK python (Ontology Access Kit): https://
eric: Queries were supposed to go over their info mdel.
… You could infer the structure from the DDL, or maybe from an ont.
… But sounds like they used both. How did they use the ont?
darrell: Their product maps from KG to individual DBs.
eric: Could ask a FHIR query, either 1. "I'm looking for sombody w a procedure performed on the kidney", and that's turned into a query for procedure performed on a kidney, or on kidney lining, etc., using the closure over the subclasses to expand the query.
… 2. Issue the original SPARQL query, but have the infra on the FHIR side, HAPI FhIR server goes to the term server to do that closure.
darrell: And HPO work is doing that?
jim: Joe Flack is the developer working on that at HPO. He works for Chris Chute at Johns Hopkins, but also working on Monarch initiative.
… The OAK python toolkit does lots of things w ont. IDK if they're hosting the HPO ont someplace as FHIR.
eric: Wonder if you could use shex to steer GPT4.
gaurav: You could also give it a shex schema and tell GPT4 to "using this schema, give me ... "
eric: But a well-informed person still needs to read the GPT4 responses to tell it to correct them.. Maybe use shex for that.
eric: Will be hard to make it guess a good query.
darrell: Benchmarks that data.world did are not perfect, an dthey discuss them.
… How to prompt it to correct the mistakes?
eric: What's "included" in the FHIR queries?
darrell: it joins other info to your results.
Concept IRIs
gaurav: The submission part of jira is down right now. Cannot edit it right now.
… But I'll keep an eye on it.
ADJOURNED