WikiMed and PubMedDS: Two large-scale datasets for medical concept extraction and normalization research

Two large-scale, automatically-created datasets of medical concept mentions, linked to the Unified Medical Language System (UMLS).

WikiMed

Derived from Wikipedia data. Mappings of Wikipedia page identifiers to UMLS Concept Unique Identifiers (CUIs) was extracted by crosswalking Wikipedia, Wikidata, Freebase, and the NCBI Taxonomy to reach existing mappings to UMLS CUIs. This created a 1:1 mapping of approximately 60,500 Wikipedia pages to UMLS CUIs. Links to these pages were then extracted as mentions of the corresponding UMLS CUIs.

WikiMed contains:

• 393,618 Wikipedia page texts
• 1,067,083 mentions of medical concepts
• 57,739 unique UMLS CUIs

Manual evaluation of 100 random samples of WikiMed found 91% accuracy in the automatic annotations at the level of UMLS CUIs, and 95% accuracy in terms of semantic type.

PubMedDS

Derived from biomedical literature abstracts from PubMed. Mentions were automatically identified using distant supervision based on Medical Subject Heading (MeSH) headers assigned to the papers in PubMed, and recognition of medical concept mentions using the high-performance scispaCy model. MeSH header codes are included as well as their mappings to UMLS CUIs.

PubMedDS contains:

• 13,197,430 abstract texts
• 57,943,354 medical concept mentions
• 44,881 unique UMLS CUIs

Comparison with existing manually-annotated datasets (NCBI Disease Corpus, BioCDR, and MedMentions) found 75-90% precision in automatic annotations. Please note this dataset is not a comprehensive annotation of medical concept mentions in these abstracts (only mentions located through distant supervision from MeSH headers were included), but is intended as data for concept normalization research.

Due to its size, PubMedDS is distributed as 30 individual files of approximately 1.5 million mentions each.

Data format

Both datasets use JSON format with one document per line. Each document has the following structure:

{
    "_id":  "A unique identifier of each document",
    "text": "Contains text over which mentions are ",
    "title": "Title of Wikipedia/PubMed Article",
    "split": "[Not in PubMedDS] Dataset split: <train/test/valid>",
    "mentions": [
        {
            "mention": "Surface form of the mention",
            "start_offset": "Character offset indicating start of the mention",
            "end_offset": "Character offset indicating end of the mention",
            "link_id": "UMLS CUI. In case of multiple CUIs, they are concatenated using '|', i.e., CUI1|CUI2|..."
        },
        {}
    ]
}

Version history