UMLS

The eds.umls pipeline component matches the UMLS (Unified Medical Language System from NIH) terminology.

Very low recall

When using the exact' matching mode, this component has a very poor recall performance. We can use thesimstring` mode to retrieve approximate matches, albeit at the cost of a significantly higher computation time.

Usage

eds.umls is an additional module that needs to be setup by:

1. pip install -U umls_downloader
2. Signing up for a UMLS Terminology Services Account. After filling a short form, you will receive your token API within a few days.
3. Set UMLS_API_KEY locally: export UMLS_API_KEY=your_api_key

import spacy

nlp = spacy.blank("fr")
nlp.add_pipe("eds.umls")

text = "Grosse toux: le malade a été mordu par des Amphibiens " "sous le genou"

doc = nlp(text)

doc.ents
# Out: (toux, a, par, Amphibiens, genou)

ent = doc.ents[0]

ent.label_
# Out: umls

ent._.umls
# Out: C0010200

You can easily change the default languages and sources with the pattern_config argument:

import spacy

# Enable the french and english languages, through the french MeSH and LOINC
pattern_config = dict(languages=["FRE", "ENG"], sources=["MSHFRE", "LNC"])

nlp = spacy.blank("fr")
nlp.add_pipe("eds.umls", config=dict(pattern_config=pattern_config))

See more options of languages and sources here.

Configuration

The pipeline can be configured using the following parameters :

Parameter	Description	Default
term_matcher	Which algorithm should we use : exact or simstring	"exact"
term_matcher_config	Config of the algorithm (SimstringMatcher's for simstring)	{}
pattern_config	Config of the terminology patterns loader	{"languages"=["FRE"], sources=None} (sources None means all available sources)
attr	spaCy attribute to match on (eg NORM, TEXT, LOWER)	"LOWER"
ignore_excluded	Whether to ignore excluded tokens for matching	False

Authors and citation

The eds.umls pipeline was developed by AP-HP's Data Science team and INRIA SODA's team.