Corpus

The 2014 De-identification dataset can be obtained at:

https://www.i2b2.org/NLP/DataSets/

Specifically, download the "2014 De-identification and Heart Disease Risk Factors Challenge. A data use agreement needs to be signed, so the dataset could not be included here.

Converting to CONLL format

To obtain the i2b2 2014 deidentification corpus in CONLL format, we used the tools bundled with NeuroNER, available at:

https://github.com/Franck-Dernoncourt/NeuroNER

1. First follow the instructions in NeuroNER/data/i2b2_2014_deid/readme.md Run the python script xml_to_brat.py (this requires Python 3).

2. Use the script brat_to_conll.py located in NeuroNER/src

   Specifically:

   We used 'spacy' (not 'stanford') for tokenizing. We downloaded the english language model en_core_web_sm-1.2.0.tar.gz from here: https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-1.2.0/en_core_web_sm-1.2.0.tar.gz

   Then:

   python3 -m spacy link en_core_web_sm en_default

   Then run the following in Python 3:

   import brat_to_conll brat_to_conll.brat_to_conll(input_folder, output_filepath, 'spacy', 'en_default')

   Run it three times (for the training, testing, and dev data) using the appropriate input_folder and output_filepath names.

Note on train/dev split

Note: according to the xml_to_brat.py script:

training-PHI-Gold-Set1 = training set training-PHI-Gold-Set2 = dev/validation

It appears this is what was used in the paper Lee et al (2017), "Transfer Learning for Named Entity Recognition with Neural Networks". They mention "60% [of train set] corresponds to the full official train set". The only way this makes sense is as the fraction train/(train+dev), which is closer to 66% (sentence level).

We could not find this training/dev split mentioned in any other documentation or papers related to the corpus.

Cleaning the data

The last few lines of file 180-03.xml are not formatted correctly in the final output; these were corrected manually.

Several entity types had too few (<20) instances and were removed. These were changed as follows:

HEALTHPLAN -> O # 1 mention URL -> O # 2 mentions FAX -> O # 10 mentions EMAIL -> O # 5 mentions DEVICE -> O # 7 mentions LOCATION_OTHER -> 0 # 17 mentions BIOID -> IDNUM # 1 mention

Remarks

There are still some sentence segmentation errors in the CONLL-formated files.

Cite as

If using the i2b2 2014 dataset, please cite as:

• Stubbs A, Uzuner O. (2015). Annotating longitudinal clinical narratives for de-identification: The 2014 i2b2/UTHealth corpus (http://www.ncbi.nlm.nih.gov/pubmed/26319540.). J Biomed Inform. 2015 Aug 28. PII: S1532-0464(15)00182-3. DOI: 10.1016/j.jbi.2015.07.020.
• Stubbs A, Kotfila C, Uzuner O. (2015). Automated systems for the de-identification of longitudinal clinical narratives: Overview of 2014 i2b2/UTHealth shared task Track 1 (http://www.ncbi.nlm.nih.gov/pubmed/26225918). J Biomed Inform. 2015 Jul 28. PII: S1532-0464(15)00117-3. DOI: 10.1016/j.jbi.2015.06.007.