Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Inconsistent models, mapping, interoperability, and the SUO




Rob Freeman wrote

> Rich,
>
> What kinds of rules are you looking for exactly?
>
> Can you give me an example of the kind of query you want your system to
> answer?
>
> -Rob
For example, in drug to drug interactions, there are classes of drugs that interfere with other classes.  There are at least 10,000 prescriptions drugs, and I've gotten estimates as high as 100,000.   So the number of 2-way and 3-way drug interactions is far too large to treat with keyword search alone.  Watching patient medical records for cases where the wrong combination slipped through the system is one application.  For the average hospital stay, your chances are about 30% that you will get the wrong drug at least once during your stay.  So the pharmacist's commputerized interaction checker isn't enough to head off this kind of danger.  So you need a drug classification system to specify the ones to look for, and to recognize them in real time in medical records. 
 
From a slide I made:
 

               On admission 08/22/01 resident’s height was 68 inches and weight was 143.  At the acute care hospital his height was listed as 72.4 inches and weight 153 pounds.  This discrepancy was never brought to the physician’s attention.

               Significant weight losses and gains not called to the physician.  On 10/03/01 –10 lbs MD not notified.  On 10/10/01 weekly weight gain of +3 lbs. Not called to MD.  On 10/17/01 weekly weight loss of  -3 lbs. Not called to MD.  On 11/24/01 monthly weight gain of +7 lbs not called to MD.  On 05/24/02 monthly weight loss of –6 ¾ lbs not called to MD.  Malnutrition and weight loss should have been monitored very closely.

               On 10/25/01 nurse’s documented “coughing and unable to swallow breakfast, has congestion, suctioned and kept comfortable” physician was not notified of increased coughing which required suctioning.  An MD order for suctioning PRN/excessive secretions should have been obtained.  No assessment of lung sounds documented.  Resident has history of chronic aspirations and pneumonias, which increases risk for dehydration.

               On 01/23/02 Mr. Patient was noted coughing up milky colored phlegm and some coming out of his nose.  Noted with rhonchi.  Dr. Doctor paged and message left for him at the office.  No documentation if physician ever called back.  Mr. Patient has history of chronic aspiration.

 

 
 

Markup Color Codes:

Anaphoric reference

Operational Discrepancy

Nonsyntactic Phrase

Patient Condition Alert

Lexical Dialect

Failure to Notify MD of Change of Condition

 

Fragment of Actual Medical Record

 

 

 

 

And another slide:

 

 

WordNet 2.0's definition of drug:

 

 

 

 

 

 

and finally, an example deduction from the prototype tool I'm developing:

 

 

 

 

As you can see, WordNet 2.0 models the English usage for drug naming and classification conventions, so it provides an initial charge of words, relationships among words, and knowledge about drugs that helps reach a fully capable system faster than if you just start with the words alone. 

 

HTH,

Rich

 


> On Wednesday 23 March 2005 23:51, Rich Cooper wrote:
>> Rob Freeman wrote
>>
>>     Rich. Why use WordNet? Why not use words?
>>     ...
>>
>>     -Rob
>>
>>
>>     because WordNet has a large initial charge of nouns in a class lattice
>> that reflects common usage of those nouns.  To start from scratch, I could
>> extract words and build my own list, but then I wouldn't have any of the
>> collateral information about synonymy, antonymy, hyponymy, hypernymy,
>> meronymy and all the relational database descriptions of those relations to
>> start from.
>>
>>     This approach appears to provide some initial legs up to stand on so
>> corpus analysis can be used to flesh out the sparse WordNet conceptual
>> structures.
>>
>>     Since WordNet has no deductive mechanisms, I've added an FOL rule
>> engine, but won't have any rules to put into it until I find a source of
>> those rules.  So far, I haven't found any way of automating or kick
>> starting the rule base though.  Cyc doesn't relate its conceptual
>> structures to WordNet (unless someone out there knows something about how
>> its done).  There are a few rule bases (FrameNet, VerbNet, ...) but nothing
>> as widely accepted as WordNet so far as I know.
>>
>>     I'm just trying to provide a basic capability to advance linguistic Q&A
>> technology half a step.
>>
>>
>>
>>     Thanks for the comments,
>>
>>     Rich
>