Craig: syndromic and case report can complement
EricS: until 4-5 years ago, it was all done with forms
... forms with "does this person present with symptom X?"
... lots of yess and nos
... recently, there's been a desire for a more detailed picture
... the states aren't prepared to collect data at that level of detail
... a richer presentation of a patient's profile would present a clearer picture for public health
... for syndromic (admissions, tests, complaints)
... , there's a desire for richer presentation, but there's no mandate
... so current representation is electronic form of yess and noes
egombocz: i think we can do substantial work on peices of overlap?
... i'll be presenting a poster at Bio-IT
... i can go over the formal ontologies from the last couple meetings and find join points
... i'm interested in prediction/outbreak prevention
... we're looking at bacterial and viral pathogens
... we create ontologies on the fly and see where we can connect them
<egombocz> http://www.io-informatics.com/news/pdfs/BioIT2012_Poster.pdf
EricS: the V2 approach intends to be very flexible but ends up brittle
... we translate stuff to RDF for processing and epedimiological studies
... whether or not we could get folks to report RDF, there's still value to the RDF for processing
... e.g. taking a V2 message, converting it to RDF and then use it for classification
Craig: i'm not sure we could have much impact (c.f. V2-to-V3 impedence), but that doesn't keep us from working on this
... another piece that interests us is taking the pieces of the puzzle for inference
... e.g. if we get over-the-counter drugs sales for upper-represpiratory issues correlated with case reports
... how can we put that stuff togheter in a common way beyond classification into probablility and correlation
... we've built expert systems which rely on weightings from a belief system (e.g. bayesian)
... it's easy to say "here are the five relevent symtoms" obtained from the reports
... so we have a graham stain analysis recommended by probabilities
ericP: so reification with weights?
Craig: using the AI paradigm of "blackboards"
<mscottm2> IE -> AI
Craig: we modeled the blackboard architecture in RDF
... whatever we plug into that, e.g. JESS, registers with the RDF repo
EricS: we're mapping to DROOLS
[discussion of what's doable with SPARQL and without JESS]
mscottm2: [discussion of bayes over unambiguous identifiers]
EricS: sometimes there are no coded representations
... sometimes we have to have to look at the ontology, e.g. SNOMED's term may be at wrong granularity
Craig: we use bayes for e.g. associating fever with chills
... our typical challenges are in drawing from the natural language, e.g. cheif complaint
... we're processing small bodies of text