D3 visualizations, Underscore mixins, etc.

Documentation of Sigfried Gold's software on GitHub

Merki

Demo Source

Medication Extraction and Reconciliation Knowledge Instrument. A tool for extracting structured medication information from discharge summaries and other free-text narrative sources, described in this award-winning paper: Extracting Structured Medication Event Information from Discharge Summaries by Sigfried Gold, Noémie Elhadad, Xinxin Zhu, James J. Cimino, and George Hripcsak.

Citations where you may find more effective parsers built more recently. If you learn of any that are also open source, let me know!

MERKI Parser, Public Version, Documentation

Files:

ParseMeds.pm            parser code module
drugParseRules.yaml     parser rules.  can be edited if you understand regular expressions.
druglist.tsv            list of drug names, CUIs and TTY for ingredients and brand names from RxNorm
gpl.txt                 GPL3 license text
parseFromPerl.pl        example of how to call the parser from a Perl script
parseFromShell.pl       command line version.  run like:
                            echo "...tylenol 250mg po daily..." | perl parseFromShell.pl

MERKI was a sprawling, ambitious application I worked on during my time as a student of Biomedical Informatics at Columbia University. It’s purpose was to extract medication information from structured and free-text patient data, standardize and condense it, and produce a complete and concise listing of all medications mentioned in each patient’s electronic medical record. The larger project was never finished. The current files are a portable subset that allow the parsing of narrative clinical text for the extraction of structured medication information.

The following two lines from parseFromShell.pl show how to use the parser:

my $drugs = $parser->twoLevelParse($input, 
    ['drug', 'possibleDrug', 'context'], 
    ['dose', 'route', 'freq', 'prn', 'date']);  
print $parser->drugsToXML($drugs);

$parser->twoLevelParser goes over its input twice: once to extract drugs, possible drugs, and contexts; and a second time to find, within each drug or possible drug, the dose, route, frequence, prn and dates. twoLevelParser returns a Perl data structure which can then be passed to $parser->drugsToXML or $parser->drugsToHTMLTable in order to turn it into something more directly usable. Here is an example (taken from bits of random clinical text, and not meant to be clinically plausible):

unixshell$ echo "Discharge medications: Procardia XL 60 mg p.o. prn for severe wheezing, ferros sulfate 300 mg p.o. b.i.d., Cipro 250 mg p.o. q12hQ" | perl parseFromShell.pl
<drugs>
    <drug>
        <drugName>Procardia XL</drugName>
        <dose>60 mg</dose>
        <route>p.o.</route>
        <prn>prn for severe wheezing</prn>
        <startChar>23</startChar>
        <endChar>69</endChar>
        <textLength>47</textLength>
        <when>after discharge</when>
        <context>Discharge meds</context>
        <surroundingText>discharge medications: [Procardia XL 60 mg p.o. prn for severe wheezing], ferros sulfate 300 mg p.o. b</surroundingText>
    </drug>
    <possibleDrug>
        <drugName>ferros sulfate </drugName>
        <dose>300 mg</dose>
        <route>p.o.</route>
        <freq>b.i.d.</freq>
        <startChar>72</startChar>
        <endChar>104</endChar>
        <textLength>33</textLength>
        <when>after discharge</when>
        <context>Discharge meds</context>
        <surroundingText>p.o. prn for severe wheezing, [ferros sulfate 300 mg p.o. b.i.d.], D1DDD 250 mg p.o. q12hq</surroundingText>
    </possibleDrug>
    <drug>
        <drugName>Cipro</drugName>
        <dose>250 mg</dose>
        <route>p.o.</route>
        <freq>q12</freq>
        <startChar>107</startChar>
        <endChar>127</endChar>
        <textLength>21</textLength>
        <when>after discharge</when>
        <context>Discharge meds</context>
        <surroundingText>s sulfate 300 mg p.o. b.i.d., [Cipro 250 mg p.o. q12]hq</surroundingText>
    </drug>
</drugs>

To understand how the parser decides what counts as a drug, a possible drug, a context, a dose, route, etc., look at these tokens in drugParseRules.yaml. The parser itself (ParseMeds.pm) treats context tokens differently than drugs and possible drugs. Any context token found becomes the context attribute of all drugs and possible drugs following it, until another context is found.

Notice that “ferros sulfate” (“ferrous sulfate” misspelled) appears as a possible drug rather than as a drug. Since it is misspelled, it is not found in the drug lexicon, but it is still identified as a possible drug because it appears before a dose, route, and frequency. (Look at the definition of possibleDrug in drugParseRules.yaml.)

This application is far from perfect, and if you do find it worth using, there is a good chance you will want to modify it for your own uses.

Changing the drug lexicon should be fairly straightforward. You can add, delete, or change entries as you like, or use an entirely different lexicon. If you change the format of the lexicon, you may need to change aspects of the parser that load and look up drugs.

You may want to change the parsing rules to catch drug phrases that the current set of rules won’t catch, or, alternatively, to make the rules more conservative to prevent false positives. You’ll need to understand the basics of the YAML format (or just follow the example of the current drugParseRules.yaml file), and, more importantly, you’ll need to understand Perl regular expressions and the special way that the parsing rules are processed. I’ll explain how the parsing rules are processed now.

Tokens are divided into terminals and non-terminals. Tokens of either sort are transformed by the parser into single Perl regular expressions. The difference is that non-terminals can include terminals and other non-terminals in their definition. You’ll also notice that the way they are written is slightly different, but that is just to make them more readable. Also, terminals can include literal text, but non-terminals cannot (because they try to interpret literal text as a reference to another token.)

Terminals are made up of a name followed by a list of expressions or pieces of literal text. Take, for instance, the terminal cond:

cond:   [ud, ut dict, prm-breakthrough, '(were|was) held', discontinued, "dc'd"]

This will be converted into the (approximately) following regular expression:

/(ud|ut dict|prn-breakthrough|(were|was) held|discontinued|dc\'d)/

Actually, two other options will be added to the list of strings it will match: “u.d.” and “ut dict.” This is because the convenience rule dotsAfterLtrOk includes “ud” and the convenience rule dotsAtEndOk includes “ut dict”.

The terminal cond is used in the non-terminal prn:

- name: prn
  patterns:
    - 'asNeeded\s*qualifier'
    - '(cond|asNeeded)'

This translates into “prn” or “as needed” (that’s how “asNeeded” is defined) followed by a qualifier (something like “for severe pain”), OR something that matches the cond token, OR something that matches the asNeeded token. Generally you will see the parsing rules composed such that more specific, longer expressions appear before less specific, shorter expressions. This is because the first expression matched will be kept and subsequent expressions will not be tried.

Finally, you may also want to modify the parsing code itself, but that code is not documented and may be hard to understand.

If you do make changes, or even if you use the code at all, I would very much appreciate hearing from you. I may be able to offer assistance, and I may be able to make your improvements available to others.

Contact Sigfried Gold with questions.

Comments