README.md

metadata

annotations_creators:
  - train: programmatically-generated
  - val: programmatically-generated
  - test: programmatically-generated, expert-validated
language_creators:
  - found
languages:
  - en
licenses:
  - unknown
multilinguality:
  - monolingual
size_categories:
  - 10K<n<100K **FIX**
task_categories:
  - structure-prediction
task_ids:
  - named-entity-recognition

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Repository: Github
• Paper: Pending

Dataset Summary

EpiSet4NER is a bronze-standard dataset for epidemiological entity recognition of location, epidemiologic types, and created using weakly-supervised teaching methods

locations, epidemiological identifiers (e.g. "prevalence", "annual incidence", "estimated occurrence") and epidemiological rates (e.g. "1.7 per 1,000,000 live births", "2.1:1.000.000", "one in five million", "0.03%")

These are the V3 training (456 abstracts), validation (114 abstracts), and programmatically generated test (50 abstracts) set. The training set was copied to datasets/EpiCustomV3 and renamed train.tsv. The validation set was copied to datasets/EpiCustomV3 and datasets/Large_DatasetV3 and renamed val.tsv. The V3 test set (uncorrected) is important as it is used by Find efficacy of test predictions.ipynb to find the efficacy of the programmatic labeling, but was otherwise not used with the model. NIH NCATS GARD

An example of 'train' looks as follows.

{
    "id": "333",
    "tokens": ['Conclusions', 'The', 'birth', 'prevalence', 'of', 'CLD', 'in', 'the', 'northern', 'Netherlands', 'was', '21.1/10,000', 'births', '.'],
    "ner_tags":  [0, 0, 0, 3, 0, 0, 0, 0, 0, 1, 0, 5, 6, 0],
}

Data Fields

The data fields are the same among all splits.

• id: a string feature.
• tokens: a list of string features.
• ner_tags: a list of classification labels, with possible values including O (0), B-LOC (1), I-LOC (2), B-EPI (3), I-EPI (4),B-STAT (5),I-STAT (6).

Data Splits by number of tokens

name	train	validation	test
EpiSet	117888	31262	13910

Dataset Creation

This bronze-standard dataset was created from 620 rare disease abstracts Programmatic Labeling using statistical and rule-based methods (Weakly Supervised Teaching)

Evaluation Level	Entity	Precision	Recall	F1
Entity-Level	Overall	0.559	0.662	0.606
	Location	0.597	0.661	0.627
	Epidemiologic Identifier	0.854	0.911	0.882
	Epidemiologic Rate	0.175	0.255	0.207
------------------	--------------------------	-----------	--------	-------
Token-Level	Overall	0.805	0.710	0.755
	Location	0.868	0.713	0.783
	Epidemiologic Type	0.908	0.908	0.908
	Epidemiologic Rate	0.739	0.645	0.689

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

A sample of 500 disease names were gathered from ~6061 rare diseases tracked by GARD.

Annotations

Annotation process

See here and then here

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

Generates whole_abstract_set.csv and positive_abstract_set.csv. whole_abstract_set.csv is a dataset created by sampling 500 rare disease names and their synonyms from GARD.csv until ≥50 abstracts had been returned or the search results were exhausted. Although ~25,000 abstracts were expected, 7699 unique abstracts were returned due to the limited research on rare diseases. After running each of these through the LSTM RNN classifier, the positive_abstract_set.csv was created from the abstracts which had an epidemiological probability >50%. positive_abstract_set.csv will be passed to create_labeled_dataset_V2.ipynb

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

[More Information Needed]

Citation Information

[More Information Needed]

Contributions

Thanks to @William Kariampuzha at NCATS/Axle Informatics for adding this dataset.