Skip to content
Branch: master
Go to file

Latest commit

soldni committed 8420573 Jun 28, 2020


Failed to load latest commit information.

NEW: v.1.4 supports starting multiple QuickUMLS matchers concurrently! I've finally added support for unqlite as an alternative to leveldb for storage of CUIs and Semantic Types (see here for more details). unqlite-backed QuickUMLS installation support multiple matchers running at the same time. Other than better multi-processing support, unqlite should have better support for unicode.


QuickUMLS (Soldaini and Goharian, 2016) is a tool for fast, unsupervised biomedical concept extraction from medical text. It takes advantage of Simstring (Okazaki and Tsujii, 2010) for approximate string matching. For more details on how QuickUMLS works, we remand to our paper.

This project should be compatible with Python 3 (Python 2 is no longer supported) and run on any UNIX system (support for Windows is experimental, please report bugs!). If you find any bugs, please file an issue on GitHub or email the author at


  1. Obtain a UMLS installation This tool requires you to have a valid UMLS installation on disk. To install UMLS, you must first obtain a license from the National Library of Medicine; then you should download all UMLS files from this page; finally, you can install UMLS using the MetamorphoSys tool as explained in this guide. The installation can be removed once the system has been initialized.
  2. Install QuickUMLS: You can do so by either running pip install quickumls or python install. On macOS, using anaconda is strongly recommended.
  3. Create a QuickUMLS installation Initialize the system by running python -m quickumls.install <umls_installation_path> <destination_path>, where <umls_installation_path> is where the installation files are (in particular, we need MRCONSO.RRF and MRSTY.RRF) and <destination_path> is the directory where the QuickUmls data files should be installed. This process will take between 5 and 30 minutes depending how fast the CPU and the drive where UMLS and QuickUMLS files are stored are (on a system with a Intel i7 6700K CPU and a 7200 RPM hard drive, initialization takes 8.5 minutes). python -m quickumls.install supports the following optional arguments:
    • -L / --lowercase: if used, all concept terms are folded to lowercase before being processed. This option typically increases recall, but it might reduce precision;
    • -U / --normalize-unicode: if used, expressions with non-ASCII characters are converted to the closest combination of ASCII characters.
    • -E / --language: Specify the language to consider for UMLS concepts; by default, English is used. For a complete list of languages, please see this table provided by NLM.
    • -d / --database-backend: Specify which database backend to use for QuickUMLS. The two options are leveldb and unqlite. The latter supports multi-process reading and has better unicode compatibility, and it used as default for all new 1.4 installations; the former is still used as default when instantiating a QuickUMLS client. More info about differences between the two databases and migration info are available here.

: If the installation fails on macOS when using Anaconda, install leveldb first by running conda install -c conda-forge python-leveldb.


A QuickUMLS object can be instantiated as follows:

from quickumls import QuickUMLS

matcher = QuickUMLS(quickumls_fp, overlapping_criteria, threshold,
                    similarity_name, window, accepted_semtypes)


  • quickumls_fp is the directory where the QuickUMLS data files are installed.
  • overlapping_criteria (optional, default: "score") is the criteria used to deal with overlapping concepts; choose "score" if the matching score of the concepts should be consider first, "length" if the longest should be considered first instead.
  • threshold (optional, default: 0.7) is the minimum similarity value between strings.
  • similarity_name (optional, default: "jaccard") is the name of similarity to use. Choose between "dice", "jaccard", "cosine", or "overlap".
  • window (optional, default: 5) is the maximum number of tokens to consider for matching.
  • accepted_semtypes (optional, default: see is the set of UMLS semantic types concepts should belong to. Semantic types are identified by the letter "T" followed by three numbers (e.g., "T131", which identifies the type "Hazardous or Poisonous Substance"). See here for the full list.

To use the matcher, simply call

text = "The ulna has dislocated posteriorly from the trochlea of the humerus."
matcher.match(text, best_match=True, ignore_syntax=False)

Set best_match to False if you want to return overlapping candidates, ignore_syntax to True to disable all heuristics introduced in (Soldaini and Goharian, 2016).

If the matcher throws a warning during initialization, read this page to learn why and how to stop it from doing so.

Server / Client Support

Starting with v.1.2, QuickUMLS includes a support for being used in a client-server configuration. That is, you can start one QuickUMLS server, and query it from multiple scripts using a client.

To start the server, run python -m quickumls.server:

python -m quickumls.server /path/to/quickumls/files {-P QuickUMLS port} {-H QuickUMLS host} {QuickUMLS options}

Host and port are optional; by default, QuickUMLS runs on localhost:4645. You can also pass any QuickUMLS option mentioned above to the server. To obtain a list of options for the server, run python -m quickumls.server -h.

To load the client, import get_quickumls_client from quickumls:

from quickumls import get_quickumls_client
matcher = get_quickumls_client()
text = "The ulna has dislocated posteriorly from the trochlea of the humerus."
matcher.match(text, best_match=True, ignore_syntax=False)

The API of the client is the same of a QuickUMLS object.

In case you wish to run the server in the background, you can do so as follows:

nohup python -m quickumls.server /path/to/QuickUMLS {server options} > /dev/null 2>&1 & echo $! >

When you are done, don't forget to stop the server by running.

kill -9 `cat`


  • Okazaki, Naoaki, and Jun'ichi Tsujii. "Simple and efficient algorithm for approximate dictionary matching." COLING 2010.
  • Luca Soldaini and Nazli Goharian. "QuickUMLS: a fast, unsupervised approach for medical concept extraction." MedIR Workshop, SIGIR 2016.
You can’t perform that action at this time.