Re: [Gnumed-devel] Lab import (data) matching

[James Busser]

On 3-Feb-08, at 7:43 PM, James Busser wrote:

> Of interest, a registry study that tested match keys based on  
> various combinations (name substrings, date of birth, gender, and  
> health center) to link encounter and laboratory files found that an  
> optimal match key was the first two letters of the last name with  
> the date of birth, which had a sensitivity of 92.7% and a positive  
> predictive value of 99.5% (even across ethnicity). This would  
> suffice for auto-matching for research, but not for clinical care,  
> with one obvious vulnerability being a multiple-birth within a  
> family. This implies that we should perhaps consider multiple-birth  
> as an attribute of a GNUmed person identity
>         http://www.citeulike.org/group/98/article/408895

Karsten and I have been working off-list on lab import, but I wanted  
these further thoughts (and a reference) to make it into the list  
archive.

The approach cited above yielded a low chance of incorrect matches  
(99.5% correct). I did not dig up the original article, to know the  
size of the patient populations being worked with, but as they were  
real-world, they may be guessed to be typical.

The requirement for a match on the complete last name (instead of  
just the first two characters) in a different study (URL captured  
below) was associated with similar match certainty of 97-99%. While  
the addition of a requirement to match the first name increased the  
certainty, it did so by <1% and lowered the successful match rate of  
88-93% by a further 10%.

Gaining some additional certainty by requiring *some* kind of match  
at the "first name(s)" level may be achievable by requiring only 2  
characters from the incoming message "firstnames" equivalent to match  
the first two characters of *any* name among a patient's possibly- 
multiple first names in GNUmed.

Vancouver being relatively tolerant of all sexualities, we may  
experience a higher rate of trans-gender whose coding variance  
between the laboratory and GNUmed could spoil matches. I think,  
though, that I will resist the urge to remove gender from the auto- 
match criteria because an independent criterion for a match exists  
anyway (see wiki LabImporters) and, where a gender miscode would have  
spoiled a match, it could be made apparent through a GNUmed wizard  
showing suggested matches for unmatched records.

Assessing record linkage between health care and Vital Statistics  
databases using deterministic methods.
Li B, Quan H, Fong A, Lu M.
Department of Community Health Sciences, University of Calgary
BMC Health Serv Res. 2006 Apr 5;6:48.
http://www.ncbi.nlm.nih.gov/pubmed/16597337? 
ordinalpos=1&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubm 
ed_DiscoveryPanel.Pubmed_Discovery_RA&linkpos=5&log 
$=relatedarticles&logdbfrom=pubmed