Dataset Card for HIPA-AI Dataset
The dataset includes 100 Reddit posts scraped from healthcare-related subreddits split into 80 labeled training posts, 10 labeled validation posts, and 10 unlabeled test posts. They are classified either as HIPAA violations (yes) or not HIPAA violations (no).
Dataset Details
Dataset Description
A full unlabeled dataset was scraped from nursing, medicine, doctor, physician assistant, CounselingPsychology, and nursepractictioner subreddits. Posts were filtered by the keywords "patient", "physician", "doctor", and "case". Only text-only posts were collected. Both the original post and a limited number of comments were collected. These posts were randomized. From this set of 2083 randomized posts/comments, 100 were kept for annotation. These posts were annotated and Zoiya Morell acted as the tiebreaker to decide on disputed labels. This dataset is version 2, which has extra characters and paragraph spacing removed for ease of use.
- Curated by: Brayden Cloutier, Zoiya Morell; University of Michigan-Flint
- Language(s) (NLP): English
- License: Apache 2.0 -Codalab Results Website: https://codalab.lisn.upsaclay.fr/competitions/20568#results -Codalab Results csv: https://1drv.ms/x/s!AjKvun2Ij8a8jcV9ehd2kN1M5bCxdw?e=oYMqRO
Dataset Sources
- Repository: https://github.com/BraydenAC/510-HIPA-AI
- Demo: Baseline models included in above repository.
Uses
The HIPA-AI dataset may be used to train a machine learning model to predict whether text features a HIPAA violation. It is also suitable for other healthcare and text-related model training. This dataset is for research purposes only.
Direct Use
The HIPA-AI dataset may be used for further research in healthcare or technology fields with proper citation and responsible conduct surrounding sensitive topics or possible patient data insecurity.
Out-of-Scope Use
The HIPA-AI dataset may not be used to denounce or divulge protected information regarding healthcare systems, patients, or medical practitioners. All posts are anonymous and may not be used to bring legal action against posters. These posts are best for text-only binary classifiers.
Dataset Structure
The HIPA-AI dataset is sorted into two columns. The "Features" column contains the entire text of the post, while the "Label" column contains its label (yes for HIPAA violation, no for not a violation). Each row is one data point. The data features a random 80-10-10 split of the hundred posts with no extra balancing measures. The dataset is fairly balanced, with 51 "no" labels and 49 "yes" labels. Comment posts may contain fewer words or relevant information, while original posts are likely to be lengthier.
Dataset Creation
Curation Rationale
This dataset was created in the interest of protecting patient privacy by identifying possible patient data leaks online. It is our goal to help posters understand the dangers of posting healthcare information and prevent HIPAA violations before they occur.
Source Data
The data was collected from Reddit's open-source public subreddits, including nursing, medicine, doctor, physician assistant, CounselingPsychology, and nursepractictioner subreddits.
Data Collection and Processing
Posts and comments were collected by date and shuffled after collection. They were filtered by text-only posts and the keywords "patient", "physician", "doctor", and "case". IntelliJ Idea was originally used to create the scraper script, which uses praw and Reddit account authorization. Other imported libraries include re (regular expressions), csv, os, and random.
Who are the source data producers?
The source of this data is intentionally anonymous as it includes publicly posted Reddit posts from users. Their usernames have not been provided to protect their identities, and their age and demographic information are also unknown. It can be assumed users should fit a demographic within Reddit's user policies.
Annotations
Annotation process
100 posts were annotated by two annotators, and then Zoiya Morell acted as the tiebreaker for disagreements. Annotations were completed through potato using the annotation guidelines provided in the above repository. Annotator disagreement was calculated through Cohen's Kappa, resulting in k = 0.4592.
Who are the annotators?
Annotators include Asma Arrak, Souha Ben Hassine, and Zoiya Morell.
Personal and Sensitive Information
The HIPA-AI dataset contains some strong language, sensitive health data, biases, private information, and may contain information regarding healthcare systems and policies. All timestamps and usernames have been removed, but posts may still contain some identifying information.
Bias, Risks, and Limitations
Reddit posters may be a unique demographic, and their views and writing style should not be considered to be a majority. Additionally, these healthcare subreddits reflect certain biases and perspectives.
Recommendations
Introducing this dataset to a machine learning model may cause it to adopt these biases--using this dataset for purposes other than the original goal should be done with caution. Additionally, comment posts may not contain all the information necessary to decide whether the post is a HIPAA violation. As a small dataset, supplemental posts will likely be needed, along with a means of converting text-based data into usable features.
Glossary
HIPAA: Health Insurance Portability and Accountability Act
Dataset Card Authors
Zoiya Morell