Re: Minutes from Clinical Decision Support teleconference

I wanted to first thank Matthias and the CDS group for doing a very nice 
job on some initial OWL and SPARQL rules for pharmacogenomic clinical 
decision support [1]. I had a few comments and questions regarding the 
rules...

Questions:

- Were the entries blank in "Germ-line genetic marker (curated SNPs)" 
because the 2C9 variants were not present in a curated resource?

- What is the meaning of the evidence column? I could imagine assigning 
and evidence type, an OBI description of the study from which the 
evidence was derived, or a citation. Thinking about the big picture, it 
would be neat if the evidence was a URI or SPARQL query to an annotation 
of the text (human or computer) complete with metadata on when the 
annotation was created and the expertise of the group providing the 
annotation. Perhaps there should be both an entry for the annotation and 
metadata about the scientific provenance of the data from which the rule 
was derived. In this case, the scientific provenance would be that the 
statement came from a decision made by an authoritative source (the 
FDA). It might also be interesting to link to the evidence that 
motivated that decision for which one would have to go to a different 
source e.g., <http://www.ncbi.nlm.nih.gov/pubmed/19228618> or 
<http://www.ncbi.nlm.nih.gov/pubmed/17906972>

Comments:

- It might help communicate the rules more clearly if you provide a 
label for VKORC1 variants in the column "Germ-line genetic marker (as 
appearing in source)"

- I think that the key to really making this example strong will be to 
include rules that factor in both clinical and genetic factors. Genetic 
variability is important but it is only part of the picture. Perhaps 
some rules could be derived from clinical dosing algorithms that factor 
in clinical and genetic factors such as described in 
<http://www.ncbi.nlm.nih.gov/pubmed/19228618>

- You might find it interesting that one can get a SPARQL query to the 
actual text and table from which you grabbed this variant information 
using the linkedSPLs endpoint: 
http://dbmi-icode-01.dbmi.pitt.edu/linkedSPLs/sparql:

PREFIX dailymed:<http://dbmi-icode-01.dbmi.pitt.edu/linkedSPLs/vocab/resource/>

SELECT ?dosage WHERE {
   ?s dailymed:setId "d91934a0-902e-c26c-23ca-d5accc4151b6".
   ?s dailymed:dosage ?dosage.
}


- The "level of evidence" and "clinical significance" columns are very 
challenging things to assign. The level of evidence could be some 
evidence-based ranking 
(e.g.,<http://www.fda.gov/ohrms/dockets/dailys/03/Aug03/080103/03n-0069-rpt0001-04-Attachment-b-vol4.pdf> 
) or a consensus-based scale (e.g., SORT, 
<http://www.aafp.org/afp/2004/0201/p548.html>). I think that there is no 
standard rating scale but that most will agree some kind of rating is 
important. So, it might make sense to include a triple that has the 
rating and metadata describing the scale used and who did the rating.

Clinical significance is even more challenging because no one can agree 
on what this means. In drug-drug interactions, all of the major drug 
compendia have very poor agreement on the clinical severity of an 
interaction and local sites will customize the severity rankings to fit 
the populations they work with. Some groups avoid classifying severity 
and go for an operational classification such as "avoid", "minimize 
risk", or "no special precaution" 
(<www.ncbi.nlm.nih.gov/pubmed/11297327>. Here again, it might make sense 
to include a triple that has the clinical significance (or operational) 
rating and metadata describing the scale used and who did the rating.

thanks,
-Rich


[1] 
<https://docs.google.com/spreadsheet/pub?hl=en_US&hl=en_US&key=0AiGT-vnkGcoLdFFVMEdqcFdYaDFqS0xHTnlUT0N3cEE&single=true&gid=3&output=html> 


On 04/26/2012 12:19 PM, M. Scott Marshall wrote:
> You will find the minutes for today's Clinical Decision Support
> teleconference here:
> http://www.w3.org/2012/04/26-HCLS-minutes.html
>
> Matthias showed us an interested demonstration of his SPARQL rule
> approach to pharmacogenomic data (example: matching on certain SNPs to
> produce personalized drug advice). Very nice! Many participants are
> hoping to look deeper into this in a followup call.
>
> Cheers,
> Scott
>


-- 
Richard Boyce, PhD
Assistant Professor of Biomedical Informatics
Scholar, Comparative Effectiveness Research Program
University of Pittsburgh
rdb20@pitt.edu
412-648-6768