Philippe AMELINE wrote: > > What I meant with "discourse structure" is, for example, the formal > way you can describe a colonoscopy report. It is a huge Archetype, or > rather a set of linked Archetypes. > The reason why I call it "discourse structure" is because I really > think it is important to see any medical information as a "story > part", in some kind of fractal way : the polyp (and its description, > maybe its ablation) is a very small story inside the "colonoscopy" > larger story, inside the "patient history" journey. > So, thinking in term of "discourse structure" just means you envision > all this, and take care to be homogeneous at any level. yes - "discourse structure" is the key. This is where there are universals to be found. If you have a look at the general model of Observation in openEHR (http://svn.openehr.org/specification/TRUNK/publishing/architecture/rm/ehr_im.pdf; also see the History classes in http://svn.openehr.org/specification/TRUNK/publishing/architecture/rm/data_structures_im.pdf) you will see a model of recording an observation that is pretty much language independent (I don't mean of course that the actual recording isn't in some language, or that for purely narrative lumps of text, there is no linguistic dependence); for structured results, they can be read the same way in pretty much any language - i.e. the same archetypes can be used > > Of course, this terminology is close from the natural language one. My > vision is that a good natural language translation system must have > two levels : an "analysis layer" that transform the input discourse > into a formal language independent representation (FLIR), than a > "synthesis layer" that transform this FLIR into an output natural > language. In my opinion, our goal really is to see the medical > information we store as a FLIR. this is if fee-text or near-free-text entry is being used. If structured data capture is being used, the captured data is already in the FLIR (archetype-based) form. > > The FLIR is made from a discourse structure populated with terms from > an ontology. > This ontology can be: > a concept list (level 0 : no predicate) > a semantic network (level 1 : level 2 predicates : isA(colitis, > inflammatory disease)) > or more complex relationships ; I think that your "Ontology of > reality" is of the kind. > > Currently, I must confess that my ideas are just clear enough > concerning the semantic network. Because it is a genuine "root > component" you can use everywhere. > > From my point of view, more complex ontologies become very specific to > a given decision support system (DSS) technology. To be more accurate, > for the examples you gave, you could want to address the follow up of > this "bad polyp" by instantiating a Bayesian network or just a > decision tree. The formal representation is much different. > My idea has been to build a "formal description", like the FLIR of a > medical encyclopedia entry for this polyp. It is (quite) easy and > useful for drugs, because you can get standard chapters such as > "indications, contra-indications, bad interactions". But for symptoms > or diseases, it is much harder. > Currently I decided no longer to work on that, and use a "band" of > smart agents, each one with its specific knowledge and its specific > DSS technology, rather than building a more general expert system > using a level X ontology. we have to deal with the reality that snomed will be pushed onto all of us - WHO is involved now.
- thomas
