Bias in De-identification

In the Name of Fairness: Assessing the Bias in Clinical Record De-identification (FAccT 2023).

In this project, we investigate the bias of de-identification systems on names in clinical notes via a large-scale empirical analysis.
We create 16 name sets that vary along four demographic dimensions: gender, race, name popularity, and the decade of popularity.
We insert these names into 100 manually curated clinical templates and evaluate the performance of nine public and private de-identification methods: spaCy, Stanza, flair, Amazon, Microsoft, Google, NeuroNER, Philter, and MIST.

Data

Our data can be downloaded from PhysioNet.

• General: the data used for our general analysis.
  • Input/names-first.csv and Input/names-last.csv store the first and last names in the 16 name sets we created. Please refer to Section 3.2 in our paper and Preparation/name.ipynb in this repository for the full detail.
  • Input/notes-base.csv stores the 100 manually curated clinical templates where we marked the occurrence of names and other categories of protected health information (PHI).
  • Input/notes-input.jsonl and Input/notes-label.jsonl store the 16,000 notes and the corresponding labels, respectively, for evaluating the performance on de-identifying names. Please refer to Section 3.4 in our paper and Preparation/note.ipynb in this repository for the full detail.
  • Output/notes-{method}.jsonl stores the prediction output of each of the nine evaluated de-identification baseline methods.
• Polysemy: the data used for assessing how polysemy names affect model performance.
  • Input/polysemies-input.jsonl and Input/polysemies-label.jsonl keep the notes and the corresponding labels, respectively, for evaluating the performance on de-identifying polysemy names. Please refer to Section 5.1 in our paper and Preparation/polysemy.ipynb in this repository for the full detail.
  • Output/polysemies-{method}.jsonl keeps the prediction output of each of the nine evaluated de-identification baseline methods on polysemy names.
• Finetune: the data used for fine-tuning spaCy and NeuroNER. Here, we consider two types of context (i.e., general and clinical) and two types of names (i.e., popular and diverse). Please refer to Section 6.1 in our paper and Preparation/finetune.ipynb in this repository for the full detail.
  • Input/context-general.jsonl and Input/context-clinical.jsonl contain the two types of context.
  • Input/names-popular.jsonl and Input/names-diverse.jsonl contain the two types of names. Input/names-test.jsonl contains the names used for populating the test notes.
  • Input/inputs-{context}+{name}.jsonl and Input/labels-{context}+{name}.jsonl contain the notes and the corresponding labels, respectively, for fine-tuning.
  • Input/inputs-test.jsonl and Input/labels-test.jsonl contain the notes and the corresponding labels, respectively, for evaluating fine-tuned methods.
  • Output/finetunes-{context}+{name}-{method}-{seed}.jsonl contains the prediction output of each of the two fine-tuned methods.

Service

In this folder, we examine each of the nine de-identification methods in the corresponding notebook.
We describe below the steps to set up the evaluation environment necessary for each method.

• spaCy
  • pip install -U 'spacy[cuda-autodetect]'
  • python -m spacy download en_core_web_trf
• Stanza
  • pip install stanza
• flair
  • pip install flair
• Amazon
  • Sign up for AWS and create an IAM user account
  • Set up the AWS Command Line Interface (CLI) and add a named profile in the AWS CLI config file
  • pip install boto3
• Microsoft
  • Sign up for Microsoft Azure
  • Create a language resource and keep the key and endpoint
  • pip install azure
• Google
  • Sign up for Google Cloud and create a project with billing and Data Loss Prevention (DLP) API enabled
  • Set up a service account as DLP user and get credential
  • Install Google Cloud CLI
• NeuroNER
  • Create and activate a new conda environment: conda create -n neuroner python=3.7 tensorflow-gpu=1.15 spacy=2.3 && conda activate neuroner
  • Install the gpu version of NeuroNER: pip install pyneuroner[gpu]
  • Download the English language module for Spacy: python -m spacy download en
  • Download word embeddings: wget -P External/NeuroNER/Data/Embedding http://neuroner.com/data/word_vectors/glove.6B.100d.zip && unzip External/NeuroNER/Data/Embedding/glove.6B.100d.zip -d External/NeuroNER/Data/Embedding/
  • Move and rename the fetched model to External/NeuroNER/Model/original+original
  • Create the input folder External/NeuroNER/Data/General/deploy/
  • Clean up all the intermediate outputs of NeuroNER during its prediction
• Philter
  • Download the source code to External/Philter by git clone https://github.com/BCHSI/philter-ucsf.git
  • Install the required packages: chardet, nltk, xmltodict
• MIST
  • Download and unzip MIST 2.0.4 to External/MIST/Model
  • Install MIST by running bash External/MIST/Model/install.sh (you may need Java and Python 2)
  • Download 2006 i2b2 De-identification Training Set to External/MIST/Data (you may need an account with DBMI Data Portal)

Analysis

In this folder, we analyze the prediction results of the examined de-identification methods and plot the figures in our paper.
Below is a snippet of these methods' overall performance (higher is better) and bias along dimensions (measured by recall equality difference, lower is better).

Method	Precision	Recall	F1	Gender	Race	Popularity	Decade
spaCy	0.917	0.629	0.746	0.002	0.013	0.028	0.007
Stanza	0.678	0.881	0.766	0.002	0.016	0.011	0.005
flair	0.920	0.974	0.946	0.003	0.006	0.008	0.002
Amazon	0.923	0.925	0.924	0.005	0.022	0.032	0.001
Microsoft	0.664	0.960	0.785	0.003	0.023	0.010	0.006
Google	0.609	0.869	0.716	0.009	0.025	0.014	0.010
NeuroNER	0.946	0.944	0.945	0.001	0.045	0.026	0.002
Philter	0.227	0.794	0.353	0.000	0.000	0.003	0.000
MIST	0.474	0.751	0.581	0.013	0.022	0.017	0.003

Citation

@inproceedings{xiao2023in,
  title={In the Name of Fairness: Assessing the Bias in Clinical Record De-identification},
  author={Xiao, Yuxin and Lim, Shulammite and Pollard, Tom Joseph and Ghassemi, Marzyeh},
  booktitle={FAccT},
  year={2023}
}

@article{lim2023annotated,
  title={Annotated MIMIC-IV discharge summaries for a study on deidentification of names (version 1.0)},
  author={Lim, Shulammite and Johnson, Alistair and Xiao, Yuxin and Moukheiber, Dana and Moukheiber, Lama and Moukheiber, Mira and Ghassemi, Marzyeh and Pollard, Tom},
  journal={PhysioNet},
  year={2023},
  note={"https://doi.org/10.13026/ngc0-0f54"}
}