ThejanBW/medical-privacy-ner

Medical Privacy NER

Created: 2024
Creators: Thejan, Chinthani, Oshan
Repository: ThejanBW/medical-privacy-ner
Keywords: named entity recognition (NER) · clinical NER · entity extraction · de-identification · text redaction · patient data masking · patient privacy · protected health information (PHI) · personally identifiable information (PII) · sensitive information detection · clinical text · medical record text · HIPAA-oriented workflows · mobile and on-device models

Overview

This dataset has 1,491 English text records and 4,410 character-level entity spans across 13 labels. It combines medical report-style text with general-domain passages, including examples with Sri Lankan context.

It was created to train compact models (under 20 million parameters) for a mobile application that finds and masks identifying information in medical text. PII means personally identifiable information; NER means named entity recognition, the task of locating and classifying text spans such as names, dates, and addresses. The labels focus on identifying details; there are no dedicated disease, symptom, medication, or procedure labels.

Data format

all_data.json is a UTF-8 JSON array. Each record has:

Field	Description
sentence	Original text; it can contain multiple sentences or line breaks.
meta.note_id	Unique string record ID.
meta.sl_context	Optional Sri Lankan context flag, stored as the strings "true" or "false". Present on 270 records: 45 true, 225 false; the other 1,221 are untagged. This is not a strict location filter.
meta.report_type	Optional report category, present on 6 records.
spans	List of annotations; empty in 54 records.
spans[].id	Integer ID unique within the record.
spans[].start, spans[].end	Zero-based character offsets; start is inclusive, end is exclusive.
spans[].label, spans[].text	Entity label and the exact text at sentence[start:end].

For example, the first record labels Colombo General Hospital as HOSP at offsets [0, 24), 1864 as DATE, and Sri Lanka as LOC. Preserve the original text when using offsets; changing whitespace changes their positions.

Entity labels

Counts refer to annotated spans.

Label	Meaning in this dataset	Count
PERSON	Person name	931
DATE	Calendar date or year	606
HOSP	Hospital or healthcare facility; occasionally a facility ID	550
LOC	Location or address	434
STAFF	Healthcare staff name	370
ORG	Organization	302
ID	Patient, order, account, or other identifier	232
TIMESTAMPS	Clock time, duration, or relative time expression	208
AGE	Age	201
PHONE	Phone number	183
EMAIL	Email address	171
URL	Web address	117
BIRTHDATE	Birth date or birth year	105

Load the dataset

The README maps the entire JSON file to one train split for Hugging Face. The file has no predefined validation or test split.

from datasets import load_dataset

# After uploading README.md and all_data.json to the dataset repository:
dataset = load_dataset("ThejanBW/medical-privacy-ner", split="train")

record = dataset[0]
for span in record["spans"]:
    assert record["sentence"][span["start"]:span["end"]] == span["text"]

To load the file locally, use load_dataset("json", data_files="all_data.json", split="train") or Python's json.load. Create held-out splits before reporting model performance, and check for similar report templates across splits.

Quality and privacy

All 4,410 stored spans match their text and fall within their records; no spans overlap. The 1,491 record IDs and texts are unique. This verifies offset consistency, not that every identifying detail was labeled. For instance, the hospital identifier 54321 in record 12 has the label HOSP.

The text includes names, addresses, contact details, identifiers, and health-related narratives. The annotations do not mask the original text. The creators should confirm the source and sharing rights of the examples before public release, particularly if any were drawn from real patient records.

The mobile masking use case aims to support patient privacy in a HIPAA-oriented workflow. A model can miss identifiers, and these 13 labels are not a complete mapping to HIPAA's identifier list. Dataset use or model predictions alone do not establish HIPAA-compliant de-identification. See the HHS guidance on Safe Harbor and Expert Determination.

License and citation

License: MIT (mit in the Hugging Face metadata).

To cite this dataset, reference ThejanBW/medical-privacy-ner and the revision used. Created in 2024 by Thejan, Chinthani, and Oshan.