awinml/pubmed_case_reports

PubMed Case Reports

A collection of 13,989 full-text case reports from the PubMed Central (PMC) Open Access subset, spanning 2005–2025. Each article includes structured metadata, abstract, full body text, and section-level annotations. This dataset is designed for medical NLP, clinical reasoning, and biomedical text mining.

Dataset Description

Summary

This dataset comprises case reports published in peer-reviewed medical journals, sourced from the PMC Open Access Subset. Case reports are detailed accounts of individual patient encounters — including presentation, diagnosis, treatment, and outcome — making them a rich resource for clinical reasoning research.

Each row corresponds to one article with:

• —Bibliographic metadata (title, journal, publication date, PMCID, license)
• —Structured abstract and full body text
• —Section-level breakdown of the body text with normalised section types (background, case presentation, investigations, management, outcome, discussion, etc.)

Supported Tasks

Languages

All articles are in English.

Dataset Structure

Data Fields

Each element in the sections list is a dictionary with:

In addition to the nested sections list, each section type is also available as a top-level string column (e.g. background, case_presentation, discussion) containing the concatenated text of all sections of that type, making it easy to access specific clinical sections without iterating over lists.

Data Splits

The dataset contains a single split:

Data Instance Example

{
  "pmcid": "PMC10000501",
  "title": "Multisystem Inflammatory Syndrome in Adults Associated with Recent Infection with COVID-19",
  "journal": "Diagnostics",
  "publication_date": "2023-03-04",
  "article_link": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10000501/",
  "license": "© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license",
  "abstract": "Multisystem inflammatory syndrome in adults (MIS-A) is a rare but severe complication of coronavirus disease 2019 (COVID-19)...",
  "body_text": "Background\\n\\nMultisystem inflammatory syndrome in adults (MIS-A) is a rare but severe complication...",
  "sections": [
    {"section_type": "background", "heading": "1. Introduction", "text": "Multisystem inflammatory syndrome in adults (MIS-A) is a rare..."},
    {"section_type": "case_presentation", "heading": "2. Case Report", "text": "A 34-year-old male with no significant past medical history..."},
    {"section_type": "discussion", "heading": "3. Discussion", "text": "MIS-A is a newly recognized entity that presents with..."},
    {"section_type": "conclusion", "heading": "4. Conclusions", "text": "This case highlights the importance of considering MIS-A..."}
  ],
  "background": "Multisystem inflammatory syndrome in adults (MIS-A) is a rare...",
  "case_presentation": "A 34-year-old male with no significant past medical history...",
  "discussion": "MIS-A is a newly recognized entity that presents with...",
  "conclusion": "This case highlights the importance of considering MIS-A...",
  "investigations": "",
  "management": "",
  "outcome": "",
  "other": "",
  "references": ""
}

Dataset Statistics

Overview

Per-Column Statistics

Note: Section-type columns (background, case_presentation, etc.) contain empty strings when no section of that type is present in an article. The "Non-empty" column shows how many rows have content for each field.

Publication Years

The collection skews toward recent publications, with the majority (51%) from 2021–2025.

Top Journals

License Distribution (Top 10)

Section Types

Sections are normalised into the following categories:

¹ abstract section types from the body are not duplicated as a separate column since the article's structured abstract is already available as the top-level abstract field.

Dataset Creation

Source Data

The raw XML files were downloaded from the PMC Open Access Subset via the PMC Cloud Service S3 bucket (pmc-oa-opendata).

Curation Process

1. 1.Discovery: Case report PMCIDs were discovered by querying PubMed with the Case Reports[pt] filter, segmented by year to stay within the 10K-result cap.
2. 2.Download: Full-text XML was downloaded in parallel using async HTTP (aiohttp) from the PMC Cloud Service S3 bucket.
3. 3.Parsing: Each XML file was parsed to extract:
4. 4.Article metadata (title, journal, publication date, license, abstract)
5. 5.Section-aware body text with normalised section type labels
6. 6.Text cleaning (citation markers, figure references, whitespace normalisation)
7. 7.Format: The parsed data was saved as a Parquet file.

Curation Rationale

Clinical case reports are a uniquely valuable genre of medical literature. Unlike randomised trials or systematic reviews, case reports provide fine-grained, narrative descriptions of individual patient journeys — from initial presentation through diagnosis, treatment, and follow-up. This makes them especially suited for:

• —Evaluating clinical reasoning capabilities of language models
• —Building retrieval-augmented generation systems for rare conditions
• —Training models to understand the structure of clinical narratives

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("awinml/pubmed_case_reports", split="train")

# Access a single example
example = ds[0]
print(example["title"])
print(example["abstract"])

# Iterate in streaming mode (low memory)
ds_stream = load_dataset("awinml/pubmed_case_reports", split="train", streaming=True)
for i, row in enumerate(ds_stream):
    if i >= 10:
        break
    print(row["pmcid"], row["title"][:80])

Working with Sections

The dataset provides two ways to access section text:

1. Via the `sections` list (full detail with original headings):

from datasets import load_dataset

ds = load_dataset("awinml/pubmed_case_reports", split="train")

for row in ds:
    sections = row["sections"]
    presentation = [s for s in sections if s["section_type"] == "case_presentation"]
    if presentation:
        print(f"PMCID: {row['pmcid']}")
        print(f"Heading: {presentation[0]['heading']}")
        print(f"Text: {presentation[0]['text'][:200]}...")
        break

2. Via direct section columns (simpler, concatenated across all sections of that type):

from datasets import load_dataset

ds = load_dataset("awinml/pubmed_case_reports", split="train")

for row in ds:
    if row["case_presentation"]:
        print(f"PMCID: {row['pmcid']}")
        print(f"Case presentation: {row['case_presentation'][:200]}...")
        break

3. Section length analysis with pandas:

from datasets import load_dataset
import pandas as pd

ds = load_dataset("awinml/pubmed_case_reports", split="train")
df = ds.to_pandas()

# Which sections are most common?
for col in ["background", "case_presentation", "discussion", "conclusion"]:
    pct = (df[col].str.len() > 0).mean() * 100
    print(f"{col}: {pct:.1f}% of articles have this section")

# Average discussion length
df[df["discussion"] != ""]["discussion"].str.len().describe()

Converting to Pandas

import pandas as pd
from datasets import load_dataset

ds = load_dataset("awinml/pubmed_case_reports", split="train")
df = ds.to_pandas()

# Journal distribution
print(df["journal"].value_counts().head(10))

# Average body text length by year
df["year"] = pd.to_datetime(df["publication_date"]).dt.year
print(df.groupby("year")["body_text"].apply(lambda x: x.str.len().mean()))

Retrieval-Augmented Generation

from datasets import load_dataset
from sentence_transformers import SentenceTransformer

# Load and chunk for RAG
ds = load_dataset("awinml/pubmed_case_reports", split="train", streaming=True)

# Build a simple in-memory index from body texts
corpus = []
pmcids = []
for i, row in enumerate(ds):
    if i >= 1000:
        break
    corpus.append(row["body_text"][:2000])  # first 2000 chars
    pmcids.append(row["pmcid"])

model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = model.encode(corpus, show_progress_bar=True)

Limitations & Considerations

• —Case reports are inherently anecdotal: They describe individual patient experiences and should not be treated as population-level evidence.
• —Publication bias: Journals are more likely to publish rare or novel cases, so the dataset may overrepresent unusual presentations.
• —Temporal skew: The collection is weighted toward recent publications (51% from 2021–2025).
• —No structured outcome labels: The dataset does not include standardised diagnostic or treatment outcome labels — this is a raw text corpus.
• —License variability: Articles carry different Creative Commons licenses. Users should verify license compatibility for their specific use case (see the license field per row).
• —No PHI redaction guarantee: While the source articles are published in open-access journals, individual case reports may contain identifiable patient information. Users should exercise appropriate caution.

Citation

If you use this dataset, please cite it as:

@misc{awinml_pubmed_case_reports_2025,
  author       = {Ashwin Mathur},
  title        = {PubMed Case Reports: A Dataset of Full-Text Clinical Case Reports from PMC},
  year         = {2025},
  publisher    = {Hugging Face},
  journal      = {Hugging Face Datasets},
  howpublished = {\\url{https://huggingface.co/datasets/awinml/pubmed_case_reports}}
}

License

The dataset itself is released under CC BY 4.0. Individual articles carry their own licenses as specified in the license field and may have additional restrictions.

Contact

For questions or feedback, open an issue on the dataset repository.