Initial version

README.md

@@ -0,0 +1,111 @@
+---
+language:
+- en
+license: apache-2.0
+---
+
+# 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](https://github.com/neuml/paperetl) and the [PubMed Baseline](https://pubmed.ncbi.nlm.nih.gov/download/).
+
+The 37 million articles were filtered to match either of the following criteria.
+
+- MeSH code = [D053124](https://meshb-prev.nlm.nih.gov/record/ui?ui=D053124)
+- Keyword of `H5N1` in either the `title` or `abstract`
+
+## Retrieve article abstracts
+
+The full article abstracts can be retrieved via the [PubMed API](https://www.nlm.nih.gov/dataguide/eutilities/utilities.html#efetch). 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
+
+    ```bash
+    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`.
+
+    ```bash
+    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
+
+    ```bash
+    python -m paperetl.file pubmed/data pubmed/articles pubmed/config
+    ```
+
+4. Export to dataset
+
+    ```python
+    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.
+
+    ```python
+    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")
+    ```