---
license: mit
size_categories:
- n<1K
task_categories:
- text-classification
- token-classification
pretty_name: FDA CDRH Device Recalls NER Dataset
dataset_info:
  features:
  - name: id
    dtype: string
  - name: ner_tags
    sequence: int64
  - name: tokens
    sequence: string
  - name: labels
    sequence: string
  splits:
  - name: train
    num_bytes: 3157376.00498008
    num_examples: 1606
  - name: test
    num_bytes: 790326.9950199203
    num_examples: 402
  download_size: 522272
  dataset_size: 3947703.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
tags:
- medical
---
# Dataset Card for FDA CDRH Device Recalls NER Dataset

This is a FDA Medical Device Recalls Dataset Created for Medical Device Named Entity Recognition (NER)

## Dataset Details

### Dataset Description

This dataset was created for the purpose of performing NER tasks. 
It utilizes the OpenFDA Device Recalls dataset, which has been processed and annotated for performing NER. 
The Device Recalls dataset has been further processed to extract the recall action element, which is utilized for annotation in this dataset.

- **Curated by:** Miriam Farrington for CS224N - Natural Language Processing with Deep Learning
- **Language(s) (NLP):** Python, TensorFlow
- **License:** MIT

### Dataset Sources [optional]

<!-- Provide the basic links for the dataset. -->


- **Source:** https://open.fda.gov/apis/device/recall/
- **Tools** https://doccano.github.io/doccano/
- **Paper [optional]:** Link TBD
- **Repository:** Link TBD

## Uses

### Direct Use

This dataset can be used to finetune a pretrained model such as BERT or BioBERT to identify and label medical device trade names, product codes and device components.

## Dataset Structure

This dataset contains the following fields: 

id - unique device recall identifier 
text - text of the device recall action 
label - NER label (B-DEVICE, I-DEVICE, O-DEVICE)

## Dataset Creation

#### Data Collection and Processing

The data was collected and preprocessed to extract the action element from the Recalls dataset. 
It was cleaned, de-duplicated and annotated in the B-I-O format for NER using the Doccano open-source annotation tool. 
The NER annotations (O-DEVICE, I-DEVICE, B-DEVICE) are mapped to corresponding tag values (0,1,2)
BioBERT tokenization is performed on the inputs and labels are realigned following tokenization 


### Annotations [optional]


#### Annotation process

A subset of the source data which represents this dataset was annotated utilizing the Doccano open-source annotation tool.
The annotation process follows this methodology when applying Device NER labels: 

1. Device Tradenames are annotated with the NER labels
2. Device names which represent more than 1 word are annotated using the B-DEVICE, I-DEVICE format.
3. Words which are part of the device name, but should be excluded from the dataset are labeled O-DEVICE
4. Remaining words are not labeled.

It should be noted that, while both "O" and unlabeled words might not be entities, 
the "O" label explicitly indicates that a word has been evaluated and deemed not to be an entity, whereas unlabeled words 
haven't been evaluated or are not relevant for NER.