CARES / README.md
chizhikchi's picture
Upload README.md with huggingface_hub
cd44191
metadata
annotations_creators:
  - expert-generated
language:
  - es
language_creators:
  - expert-generated
license:
  - afl-3.0
multilinguality:
  - monolingual
pretty_name: CARES
size_categories:
  - 1K<n<10K
source_datasets:
  - original
tags:
  - radiology
  - biomedicine
  - ICD-10
task_categories:
  - text-classification
dataset_info:
  features:
    - name: iddoc
      dtype: float64
    - name: id
      dtype: int64
    - name: full_text
      dtype: string
    - name: icd10
      sequence: string
    - name: general
      sequence: string
    - name: chapters
      sequence: int64
    - name: area
      sequence: string
  splits:
    - name: train
      num_bytes: 3377631
      num_examples: 2253
    - name: test
      num_bytes: 1426962
      num_examples: 966
  download_size: 2291080
  dataset_size: 4804593

CARES - A Corpus of Anonymised Radiological Evidences in Spanish 📑🏥

CARES is a high-quality text resource manually labeled with ICD-10 codes and reviewed by radiologists. These types of resources are essential for developing automatic text classification tools as they are necessary for training and fine-tuning our computational systems.

The CARES corpus has been manually annotated using the ICD-10 ontology, which stands for for the 10th version of the International Classification of Diseases. For each radiological report, a minimum of one code and a maximum of 9 codes were assigned, while the average number of codes per text is 2.15 with the standard deviation of 1.12.

The corpus was additionally preprocessed in order to make its format coherent with the automatic text classification task. Considering the hierarchical structure of the ICD-10 ontology, each sub-code was mapped to its respective code and chapter, obtaining two new sets of labels for each report. The entire CARES collection contains 6,907 sub-code annotations among the 3,219 radiologic reports. There are 223 unique ICD-10 sub-codes within the annotations, which were mapped to 156 unique ICD-10 codes and 16 unique chapters of the cited ontology.

As for the dataset train and test subsets, a stratified split was performed in order to guarantee that the number of labels in the test data is representative.