See also: IRC log
JoshMandel: re: FHIR resource creation
... modelers work in excel, props in column a, cardinality in B, desc in C, "rdfs:range" in D
... they generate all the docs from the excel
... if we can generate some way RDF adds value, e.g. detecting inconsistencies, we can persuade them to play
ericP: original hope was that we'd be able to use the RDF representation to detect overlaps or inconsistencies between resrouces
JoshMandel: i think that they're largely human-readable, when present
ericP: i thought they were XPaths over RIM/ITS
JoshMandel: that may have been true for lloyd's original set
... they've committed to human-readable by draft standard
ericP: we like RDF 'cause of the ease of integration. is that a persuasive story?
emory: a visual rep allows me as a medical domain expert to define and create representations of the world
... i've tried to communicate my domain expertise in UML etc.
... largely too verbose
... this [SemWeb tools like WebProtege] works better for me
... like the HL7 FHIR XML pipeline, i can create models which translate to javacode etc.
JoshMandel: one of OpenEHR's core goals is to enable clinicians to participate in the modeling
emory: i looked at it years ago. since then, DOD has driven me to be HL7-centric
... that said, having so many models is testimony to the fact that we don't have an adequate interface for rule authoring
JoshMandel: you expect FHIR mapped to plain java objects, over which you execute the rules?
emory: Virtual Medical Record (HL7 CDS WG) intended to support rule authoring but there are few commodity tools which handle nested rules
... most of the rule engines key on a hash of the attribues
JoshMandel: so those attributes are flattened out on the subclasses
emory: OpenCDS, the VMR tool for CDS, initially compiles by flattening the rules 'cause OpenCDS can't handle the inheritance in a performant level
... so there are perfromance issues around hierarchies, and issues around retractions
... when you assert a deeply-nested object, it becomes difficult to get the performance of the Redi algorithm
JoshMandel: my guess is that FHIR takes nested VMR objects and references them instead of embedding them
emory: my guys are much happier with FHIR; easier to follow a link than to traverse a tree.
... the other advantage of FHIR is the organization of attributes is more intuitive.
... most clinicians have no idea what a mood is.
ericP: would those clinicians object to duplication between a prescription and an injection?
emory: yes, that fits their cognitive models.
... RIM designed to be a class model but has difficulty capturing workflow state
... e.g. desribing in RIM the state changes associated with a substance administration
ericP: is the problem that it tries?
emory: yes, i argued that VMR should just be a state object. workflow should be separate
... in net, i agree that we want to feed the ontology to the FHIR toolchain
JoshMandel: each resource defn has a dir which has a examples in XML, value sets, intros, model defintions, etc.
... the build tool reads the examples, parses them into java domain objects, and serializes them as XML and JSON
... the build tool generates the XML and JSON serializers
JoshMandel: pulling an extension into the core means the intance data has to change
... there is a principled distinciton between the core and the extensions
Mike_: we've been speaking of the design-time advantages
... re: run time benefits, we can use SPARQL 1.1 federation
next meeting: Tuesdays 15 Oct 11a US Eastern