Clinical text summarization

Clinical Text Summarization

Overview

Clinicians and analysts drown in text. This skill condenses long clinical documents into faithful, structured summaries. Because hallucination is unacceptable in healthcare, the skill pairs generation with an explicit faithfulness check that verifies summary claims against the source.

When to Use This Skill

• Turning a multi-day note bundle into a problem-oriented hospital course one-liner.
• Summarizing radiology/pathology reports into impressions.
• Producing a literature digest from many abstracts (pairs with pubmed-search).
• Compressing context before downstream extraction or QA.

Methods

1. Extractive (safe default) rank and select source sentences (TextRank, or embedding-centroid selection). Zero hallucination risk because every sentence is verbatim; best when faithfulness dominates.
2. Abstractive generate new phrasing with a transformer (BART/PEGASUS/T5; long-input variants like Longformer-Encoder-Decoder for >1k tokens; clinical-tuned models where licensing allows).
3. Hybrid extract salient sentences, then abstractively smooth them (extractive grounding reduces hallucination).
4. Long documents chunk → summarize chunks → summarize the summaries (map-reduce / refine), preserving section structure.

Example

from transformers import pipeline
summ = pipeline("summarization", model="facebook/bart-large-cnn")

def summarize_long(text, chunk=900):
    words = text.split()
    parts = [" ".join(words[i:i+chunk]) for i in range(0, len(words), chunk)]
    partials = [summ(p, max_length=130, min_length=30)[0]["summary_text"] for p in parts]
    return summ(" ".join(partials), max_length=160, min_length=40)[0]["summary_text"]

# Faithfulness guard: every summary entity must appear in the source
import medspacy
nlp = medspacy.load()
def unsupported_entities(summary, source):
    src = {e.text.lower() for e in nlp(source).ents}
    return [e.text for e in nlp(summary).ents if e.text.lower() not in src]
# non-empty list => possible hallucination => fall back to extractive

Evaluation

Report ROUGE and BERTScore for overlap with reference summaries, but treat them as necessary-not-sufficient they do not detect hallucination. Add a faithfulness/factual- consistency metric (entity-overlap above, or an NLI/QA-based check) and a clinician spot- review for high-stakes use. Prefer the extractive path whenever a hallucinated fact would be dangerous.

Outputs

• summaries.parquet doc_id, summary, method, ROUGE/BERTScore, faithfulness flag.
• flagged_summaries.csv summaries with unsupported claims for review.
• summary_report.md method comparison and quality metrics.

Healthcare Context

Designed for the long, repetitive, template-heavy nature of clinical notes. De-identify PHI first. The faithfulness guard reflects the clinical-safety bar: a fluent-but-wrong summary is worse than a plain extractive one. Complements clinical-nlp-entity-extraction and clinical-text-search-elk.

References

• Hugging Face summarization https://huggingface.co/docs/transformers/tasks/summarization
• BART (Lewis et al. 2020); PEGASUS (Zhang et al. 2020); BERTScore (Zhang et al. 2020).
• Survey: faithfulness/hallucination in clinical summarization.