Medical Information Anonymization

MEDINA is a toolbox to de-identify texts, originally designed for clinical texts. This toolbox aims at de-identifying data using linear chain CRF (Wapiti tool, see https://github.com/Jekub) and producing delexicalized statistical models, i.e., without any form of surface (strictly no learning of tokens, but basic features based on tokens are used: upper/lower case, presence of digits, punctuation marks, etc.) in order to share models (since there is no nominative data in the models).

Files:

• lanceur.sh: all useful commands to process the data (assuming existing annotated data in corpus/appr/ and corpus/test/ files containing both *{ann,txt} files)

• scripts/pre_creeDictionnaire.bash: produces forme-lemme-pos.tab file in the data directory (list of forms, lemmas, and POS for French, from CNAM data); to be done only once

• scripts/pre_releveNgrammes.pl: from the previous dictionary, computes absolute frequency of trigrams of characters for French and produces a liste_ngrammes.txt file into the scripts/data/ folder

• scripts/zero_alignement.pl: converts BRAT annotations into embedded annotations (*.tag files are created); allows to manage both layered and discontinuous entities

• scripts/zero_tabulaire.pl: produces tabular files based on previous files, using a schema useful for CRF tools among following available ones: IO, BIO, BIO2, BIO2H, BWEMO, or BWEMO+

• scripts/pre_verification-tabulaire.pl: check if the produced tabular is safe for the Medina toolkit; unexpected annotations will be shown

• scripts/zero_supprimeO.pl: allows to remove unannotated lines in order to reduce the over-training of the O category; a context of 17 lines surrounding annotations seems to be useful; for mono-category models, this context must be extended (e.g., 40 lines) in order to ensure correct precision values

• config/config_zero.tpl: configuration template for Wapiti tool (for experiments based on semi-lexicalized, or fully lexicalized models, two other configuration files are given: config_lex.tpl and config_semi-lex.tpl)

• scripts/post_lexique.pl: applies the content of lexique.tab file (two columns: class and phrase) on the prediction files in order to complete the annotations (BWEMO annotation schema). Three arguments: -r (repository to prediction file: ./), -e (file extension of the input file: wap), -s (file extension of the output file: out)

• scripts/post_conversion.pl: converts the prediction file from BWEMO+ annotation schema to BIO2 annotation schema in order to perform an evaluation using the conneval.pl evaluation script

• scripts/post_differences.pl: highlights false positive and false negative from the prediction file produced by Wapiti

• scripts/crf-output-splitter.pl: allows to split the prediction file into single files, reproducing the original content (in terms of spaces and line breaks); predictions are represented with embedded tags; please indicate the output folder as 2nd argument

• scripts/post_antidatation.pl: random date shiffting based on previously identified dates

• scripts/post_pseudonymization.pl: pseudonymizes (1) person names based on lists of common first names and last names used in France and Québec, (2) city names, based on a list of cities from France, and (3) produces fake phone numbers

• scripts/post_combineColonnes.pl: merges several columns of predictions when mono-category models are used

The following commands allow:

• to train a CRF model based on existing BRAT annotations found in corpus/jorf/train/ (input training data)

• to apply this model on texts from corpus/jorf/test/ (input test data)

• to evaluate output predictions (assuming gold standard annotations exist for the test dataset)

• and to produce de-identified single files in the output folder named "output/", composed of fake phone numbers and pseudonymized person names and city names; other types of identified information are masked by a generic tag indicating the type of information.

These are end-to-end commands:

bash scripts/pre_creeDictionnaire.bash
perl scripts/pre_releveNgrammes.pl

perl scripts/zero_alignement.pl corpus/jorf/train/
perl scripts/zero_alignement.pl corpus/jorf/test/

perl scripts/zero_tabulaire.pl corpus/jorf/train/ tag tab_train.zero BWEMO+
perl scripts/zero_tabulaire.pl corpus/jorf/test/ tag tab_test.zero BWEMO+

perl scripts/pre_verification-tabulaire.pl tab_train.zero
perl scripts/pre_verification-tabulaire.pl tab_test.zero

wapiti train -t 2 -a rprop- -1 1 -p config_zero.tpl tab_train.zero modele-zero
wapiti label -p -m modele-zero tab_test.zero >sortie-zero.wap

perl scripts/post_lexique.pl -l lexique.tab -r ./ -e wap -s out
perl scripts/post_conversion.pl sortie-zero.out
perl scripts/conlleval.pl -d '\t' <sortie-zero.out

perl scripts/crf-output-splitter.pl sortie-zero.out output
perl scripts/post_antidatation.pl -r output/ -e sgml
perl scripts/post_pseudonymization.pl -r output/ -e dat

Simply add *.txt files (clinical texts written in French) into a folder (e.g., a "input" folder) and perform the following stages:

bash scripts/pre_creeDictionnaire.bash
perl scripts/pre_releveNgrammes.pl

tar -xvzf modele-deid.tar.gz
perl scripts/zero_tabulaire.pl input/ txt tab_test.zero BWEMO+
wapiti label -p -m modele-deid tab_test.zero >sortie-zero.wap
perl scripts/crf-output-splitter.pl sortie-zero.wap output
perl scripts/post_antidatation.pl -r output/ -e sgml
perl scripts/post_pseudonymization.pl -r output/ -e dat

Final de-identified files are *.pse files in the "output/" folder. Alternatively, you can check the *sgml files from the "output/" folder by adding or removing opening and closing tags for identifying information, and then to perform the two last stages.

This toolbox is licenced under the term of the two-clause BSD Licence:

Copyright (c) 2020, 2021, 2022 CNRS
All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
    * Redistributions of source code must retain the above
      copyright notice, this list of conditions and the following
      disclaimer.
    * Redistributions in binary form must reproduce the above
      copyright notice, this list of conditions and the following
      disclaimer in the documentation and/or other materials
      provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS
BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF
THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.

For help and feedback please contact the author below:

• Grouin Cyril <[email protected]>