On Fri, 6 Jul 2001, Adrian Midgley wrote: ... > I wrote a tool some years ago for a particular purpose that offered > half a dozen phrases that separately or in combination served most of > its restricted set of purposes and were selected with a click on each > that applied. It was quicker and less boring than typing them. ... > Imagine a daemon that watches what sort of things you type, and what > sort of events occur beore you type them, and stacks them up ready for > you. A hinting engine should be able to do that as well as be ready > to suggest what treatment or investigation is generally agreed to be > useful. Adrian, Thanks for clarifying. This sounds like Philippe's Odyssee system. I believe Odyssee auto-completes physician notes with terms from the semantic tree as text input occurs. ... > An unsophisticate might think it was telepathic. This may indeed be quicker than free text and yet not require checkboxes. The challenge is to have a sufficient knowledge base which may require a much larger semantic tree such as the one being constructed by Cycorp (www.cyc.com). Remember, it is not just a matter of having the correct spelling - the system must correctly _understand_ each term by precise classifiction within the semantic network. Philippe has taken the approach of starting with a very specific type of notes. It sounds like your system took the same approach. My understanding (also repeatedly emphasized by Doug Lenat of Cycorp) is that the biggest challenge is to scale up these type of "AI" systems (http://www.cyc.com/hpkb/proposal-summary-hpkb.html). This involves the ability to "learn" and to use what has been learned. Maybe systems with narrow domains and no ability to "learn" new meanings and concepts are sufficiently useful - but then what happens if the clinician wishes to include a concept that the system does not understand? > As far as the computer doing the talking goes, I don't think Eliza and > her friends are up to taking over just yet, assuming all the Turing > +ves on here are carbon-based, and TV is probably better at presenting > narratives for people to try out as their lives for the moment. While talking computers may not be ready yet, giving computers the capability to ask clarifying questions is helpful to automated semantic classification. Maybe Philippe can tell us more about this (and whether he uses this in Odyssee). > The benefit of observing language and then presenting chunks of it for > re-use is that the computer then has a map for the ontology or other > representation of what is picked, whereas if one freetexts it from > scratch the degrees of freedom are inconveniently large for a > computational approach to determining meaning. I think for sematic classfication/coding the large chunks must be parsed into smaller pieces anyways. Again, the analogy of the "word processors" may be somewhat misleading. Word processors don't really provide semantic classification - only spell checking (and maybe syntactical/grammar suggestions). Using the "word processor" _interface_ as an analogy, the semantic classification "editor" will force the clinician to select among valid "meanings" of each term as they are typed in: "Traversing" the semantic tree to select the "Concept/Term". Again, Philippe's Odyssee demonstrates this approach. > I think this also > models some of the socialisation of students into the medical > profession, where one excuse for the professional vocabulary is that > it carries agreed meanings precisely between professionals where > ordinary language would not do so. I think it is only partly true in > that context, but if one is going to do it, lets get mechanical > assistance with precison and with the layered explanations of the > agreed meaning of phrases that are easier to add formally to an > electronic system. The challenge is to make this scalable and efficient. Since you are interested in this approach, I would love to read your views on Philippe's work and Cycorp's approach. As I previously mentioned on this mailing list, Cycorp is releasing portions of their knowledge base and methodology under open source license (LGPL). We should look into how best to benefit from their work. > To the extent that the medical record is a narrative, lets have some > narrative tools. Narrative understanding/classification happens to be one of the major challenges in AI research. We are going to have a good time working on this :-). Cheers, Andrew --- Andrew P. Ho, M.D. OIO: Open Infrastructure for Outcomes TxOutcome.Org (hosting OIO Library #1) Assistant Clinical Professor Department of Psychiatry, Harbor-UCLA Medical Center University of California, Los Angeles
