Library for clinical NLP with spaCy.
MedSpaCy is currently in beta.
MedSpaCy is a library of tools for performing clinical NLP and text processing tasks with the popular spaCy
framework. The medspacy
package brings together a number of other packages, each of which implements specific
functionality for common clinical text processing specific to the clinical domain, such as sentence segmentation,
contextual analysis and attribute assertion, and section detection.
medspacy
is modularized so that each component can be used independently. All of medspacy
is designed to be used
as part of a spacy
processing pipeline. Each of the following modules is available as part of medspacy
:
medspacy.preprocess
: Destructive preprocessing for modifying clinical text before processingmedspacy.sentence_splitter
: Clinical sentence segmentationmedspacy.ner
: Utilities for extracting concepts from clinical textmedspacy.context
: Implementation of the ConText for detecting semantic modifiers and attributes of entities, including negation and uncertaintymedspacy.section_detection
: Clinical section detection and segmentationmedspacy.postprocess
: Flexible framework for modifying and removing extracted entitiesmedspacy.visualization
: Utilities for visualizing concepts and relationships extracted from text
Future work could include I/O, UMLS matching, relations extraction, and pre-trained clinical models.
You can install medspacy
using setup.py
:
python setup.py install
Or with pip:
pip install medspacy
After installing medSpaCy, you'll need to download a spaCy model to use as a base model.
The default in medSpaCy isen_core_web_sm
, which you can download as:
python -m spacy download en_core_web_sm
If you download other models, you can use them by providing the model name to medspacy.load(model_name)
.
The following packages are required and installed when medspacy
is installed:
- spaCy 2.3
- Other core medSpaCy packages:
Here is a simple example showing how to implement and visualize a simple rule-based pipeline using medspacy
:
import medspacy
from medspacy.ner import TargetRule
from medspacy.visualization import visualize_ent
# Load medspacy model
nlp = medspacy.load()
text = """
Past Medical History:
1. Atrial fibrillation
2. Type II Diabetes Mellitus
Assessment and Plan:
There is no evidence of pneumonia. Continue warfarin for Afib. Follow up for management of type 2 DM.
"""
# Add rules for target concept extraction
target_matcher = nlp.get_pipe("target_matcher")
target_rules = [
TargetRule("atrial fibrillation", "PROBLEM"),
TargetRule("atrial fibrillation", "PROBLEM", pattern=[{"LOWER": "afib"}]),
TargetRule("pneumonia", "PROBLEM"),
TargetRule("Type II Diabetes Mellitus", "PROBLEM",
pattern=[
{"LOWER": "type"},
{"LOWER": {"IN": ["2", "ii", "two"]}},
{"LOWER": {"IN": ["dm", "diabetes"]}},
{"LOWER": "mellitus", "OP": "?"}
]),
TargetRule("warfarin", "MEDICATION")
]
target_matcher.add(target_rules)
doc = nlp(text)
visualize_ent(doc)
For more detailed examples and explanations of each component, see the notebooks folder.
Here are some links to projects or tutorials which use medSpacy. If you have a project which uses medSpaCy which you'd like to use, let us know!
- VA_COVID-19_NLP_BSV: An NLP pipeline for identifying positive cases of COVID-19 from clinical text. Deployed as part of the Department of Veterans Affairs response to COVID-19
- clinspacy: An R wrapper for spaCy, sciSpaCy, and medSpaCy for performing clinical NLP and UMLS linking in R
- mimic34md2020_materials: A crash course in clinical data science from the University of Melbourne. For medSpaCy materials, see
notebooks/nlp-*.ipynb