clinical-sectionizer 0.1.1

Document section detector using spaCy for clinical NLP

Clinical Sectionizer

This package offers a component for tagging clinical section titles in docs. There are two different flavors of the sectionizer:

• Sectionizer: A spaCy component which is run on a Doc object and adds attributes to spaCy objects. This can be added to an NLP pipeline and be executed as part of nlp(text)
• TextSectionizer: A stand-alone object, independent of spaCy which takes a text and returns a list of tuples, where each tuple corresponds to a section in the text.

The sectionizer takes a list of patterns for section titles and searches for matches in a doc. When a section is found, it generates three outputs:

1. section_title: The normalized name of a section, a string
2. section_header: The span of the doc containing the header, a Span
3. section_span: The entire span of the doc containing the section, a Span

When using the spaCy Sectionizer, calling sectionizer(doc) adds the following extensions to spaCy objects:

• Doc.sections: A list of 3-tuples of (name, header, section)
• Token.section_span: The span of the entire section which the token occurs in
• Token.section_header: The span of the section header of the section a token occurs in
• Token.section_title: The name of the section header defined by a pattern
• Span attributes corresponding section_span, section_header, and section_title to the first token in a span

When using TextSectionizer, calling sectionizer(text) returns a list of 3-tuples which correspond to the outputs described above, but each as texts rather than spaCy objects: (section_title, section_header, section_text).

Sectionizng works best when you have robust and complete patterns which are designed specifically to your data. Each EHR uses different section titles, so you should adjust your patterns accordingly. A default list of patterns is loaded when using patterns="default". You can also set max_scope, which will limit the size of a section to a certain number of tokens. This can be useful to prevent sections for running on too far if following section headers aren't recognized.

Installation

You can install clinical_sectionizer via pip:

pip install clinical-sectionizer

Or by cloning this repository and running:

python setup.py install

Example

See notebooks/for more detailed examples.

>>> text = """Family History:
    Diabetes
    
    Past Medical History:
    Pneumonia
    
    Assessment and Plan:
    Atrial fibrillation. There is no evidence of pneumonia.
    """
>>> import spacy
>>> nlp = spacy.load(...) # Load a model which will match clinical concepts

>>> from clinical_sectionizer import Sectionizer
>>> sectionizer = nlp.add_pipe(Sectionizer(nlp))


>>> section_patterns = [
        {"section_title": "family_history", "pattern": "Family History:"},
        {"section_title": "past_medical_history", 
            "pattern": [
                {"LOWER": "past", "OP": "?"}, 
                {"LOWER": "medical"},
                {"LOWER": "history"}, 
                {"LOWER": ":"},
            ]
            
        },
        {"section_title": "assessment_and_plan", "pattern": "Assessment and Plan:"},
    ]
>>> sectionizer.add(section_patterns)

>>> nlp.add_pipe(sectionizer)
>>> doc = nlp(text)
>>> print(nlp.ents)
(Diabetes, Pneumonia, Atrial fibrillation, pneumonia)



>>> for (section_name, section_header, section) in doc._.sections:
        print(section_name, section_header, section, sep="\n")

family_history
Family History:
Family History:
Diabetes

past_medical_history
Past Medical History:
Past Medical History:
Pneumonia

assessment_and_plan
Assessment and Plan:
Assessment and Plan:
Atrial fibrillation. There is no evidence of pneumonia.

>>> for ent in doc.ents:
        print(ent, ent._.section_title)
    
Diabetes family_history
Pneumonia past_medical_history
Atrial fibrillation assessment_and_plan
pneumonia assessment_and_plan

Using cycontext, you can also use a visualizer which shows section headers, along with any extracted entities and optionally cycontext modifiers, in an NER-style visualization.

from cycontext.viz import visualize_ent
visualize_ent(doc, sections=True, context=False)