NeuML/pubmed-h5n1

PubMed H5N1 Articles

Current as of January 5, 2025

This dataset is metadata (id, publication date, title, link) from PubMed articles related to H5N1. It was created using paperetl and the PubMed Baseline.

The 37 million articles were filtered to match either of the following criteria.

• MeSH code = D053124
• Keyword of H5N1 in either the title or abstract

Retrieve article abstracts

The full article abstracts can be retrieved via the PubMed API. This method accepts batches of PubMed IDs.

Alternatively, the dataset can be recreated using the following steps and loading the abstracts into the dataset (see step 5).

Download and build

The following steps recreate this dataset.

1. Create the following directories and files

   mkdir -p pubmed/config pubmed/data
   
   echo "D053124" > pubmed/config/codes
   echo "H5N1" > pubmed/config/keywords
   

2. Install paperetl and download PubMed Baseline + Updates into pubmed/data.

   pip install paperetl datasets
   
   # Install paperetl from GitHub until v2.4.0 is released
   pip install git+https://github.com/neuml/paperetl
   

3. Parse the PubMed dataset into article metadata

   python -m paperetl.file pubmed/data pubmed/articles pubmed/config
   

4. Export to dataset

   from datasets import Dataset
   
   ds = Dataset.from_sql(
       ("SELECT id id, published published, title title, reference reference FROM articles "
       "ORDER BY published DESC"),
       f"sqlite:///pubmed/articles/articles.sqlite"
   )
   ds.to_csv(f"pubmed-h5n1/articles.csv")
   

5. Optional Export to dataset with all fields

   paperetl parses all metadata and article abstracts. If you'd like to create a local dataset with the abstracts, run the following instead of step 4.

   import sqlite3
   import uuid
   
   from datasets import Dataset
   
   class Export:
       def __init__(self, dbfile):
           # Load database
           self.connection = sqlite3.connect(dbfile)
           self.connection.row_factory = sqlite3.Row
   
       def __call__(self):
           # Create cursors
           cursor1 = self.connection.cursor()
           cursor2 = self.connection.cursor()
   
           # Get article metadata
           cursor1.execute("SELECT * FROM articles ORDER BY id")
           for row in cursor1:
               # Get abstract text
               cursor2.execute(
                   "SELECT text FROM sections WHERE article = ? and name != 'TITLE' ORDER BY id",
                   [row[0]]
               )
               abstract = " ".join(r["text"] for r in cursor2)
   
               # Combine into single record and yield
               row = {**row, **{"abstract": abstract}}
               yield {k.lower(): v for k, v in row.items()}
   
       def __reduce__(self):
           return (pickle, (str(uuid.uuid4()),))
   
   def pickle(self, *args, **kwargs):
       raise AssertionError("Generator pickling workaround")
   
   # Path to database
   export = Export("pubmed/articles/articles.sqlite")
   ds = Dataset.from_generator(export)
   ds = ds.sort("published", reverse=True)
   ds.to_csv("pubmed-h5n1-full/articles.csv")