DeID version 1.1 Change Log
============================
DeID version 1.1 is now compatible with Perl 5.10. Additionally, DeID
version 1.1 fixed several bugs in DeID version 1.0. For more details,
please see the description below.
Bug Fixes and Changes
=====================
* Modified code to be compatible with Perl 5.10.
* Fixed bugs that caused the scrubbed notes from DeID to give
incorrect/garbled output in the .res file (scrubbed notes from
DeID). Under DeID version 1.0, two nursing notes (out of the 2,434
notes) in the gold standard corpus were affected by this bug. If you
had previously downloaded version 1.0, and would like to know which
notes in the gold standard were affected by this bug, the two notes
were note 7 for patient 54 (with header line
"START_OF_RECORD=54||||7||||") and note 18 for patient 144 (with
header line "START_OF_RECORD=144||||18||||").
* Fixed bugs that caused DeID to mistakenly report certain PHIs as a
date PHI in the scrubbed notes. More specifically, PHI locations are
encoded as a key that consists of a number sequence that corresponds
to the start and end locations of the PHI within the text. In the
rare occasion that a PHI's location key matches that of a Date PHI
tag's location key (as a subsequence), DeID will mistakenly report the
PHI as a date PHI.
* Fixed date shift bugs for certain date formats. The following types
of date patterns should now be date shifted correctly in the scrubbed
output when running in the output mode:
- April 1st, 2002
- April 1st
- April 4th, 2002.
- April 4th 2002
- April 5th, 2002
- 2nd of April
- 2nd of Apr.
- 20th Oct., 1989.
- 20th Oct., 89.
- 20th of Oct. 1989.
- 20th Oct, 1989.
- 20th Oct, 89.
- 20th of Oct 1989.
* Fixed the problem that prevented the phone filters to identify phone
numbers when they run into text. Code modification from phone filter
does not change the code's performance on the Gold Standard corpus.
* More robust pattern recognition on text that are split into multiple
lines. More specifically, lines that begin with blank spaces are no
longer considered paragraph separators. Instead, they are considered
part of the same paragraph as previous lines. This allows for
better pattern matching for phone numbers and dates that are split
into two lines and the second line was padded with blank spaces in
the beginning.
Changes
========
* Updated the name indicators to include phrases such as "name will
be" to better locate newborn names (does not change code's
performance on the Gold Standard Corpus).
* Allow the users to set the parameter "Two Digit Year Threshold" in
the configuration files. The parameter is used to determine whether
to interpret the year as a year in the 1900's or 2000's.
The threshold must be a 1- or 2-digit number.
Two digit years > Threshold are interpreted as in the 1900's
Two digit years <= Threshold are interpreted as in the 2000's
In this release, the "Two Digit Year Threshold" parameter is
set to 30 in the configuration files (deid-output.config and
deid.config) as follows:
Two Digit Year Threshold = 30
The threshold is set according to the re-identified date range
that appear in our gold standard corpus.
Files Modified (from version 1.0)
=================================
deid.pl
deid.config
deid-output.config
GSoutput/id.info
GSoutput/id.res
GSstats/id.info
GSstats/id.phi
doc/DeidUserManual.doc
doc/DeidUserManual.pdf
Performance of DeID code
==========================
This version changes the performance of the software on Gold Standard
only slightly: the number of false positives increased by one (from
545 in version 1.0 to 546 in current version). This additional false
positive is due to the fact that lines that begin with blank spaces
are no longer considered paragraph separators in the current version.
See below for the current performance. When running the code in
performance comparison mode, you should see the following output on
your screen.
*******************************************************************************************************************
De-Identification Algorithm: Identifies Protected Health Information (PHI) in Discharge Summaries and Nursing Notes
*******************************************************************************************************************
Starting de-identification (version 1.1)...
Running deid in performance comparison mode.
Using PHI locations in id.deid as comparison. Output files will be:
id.phi: the PHI locations found by the code.
id.info: debug info about the PHI locations.
==========================
Num of true positives = 1720
Num of false positives = 546
Num of false negatives = 59
Sensitivity/Recall = 0.967
PPV/Specificity = 0.748
==========================
This document was last updated 6/10/09.