medisim 💊

medisim is a collection of new large-scale medical term similarity datasets based on SNOMED-CT. The code in this repository creates the collection of datasets.

If you use one of these datasets in your research, please cite

@inproceedings{
SchulzJuric2020AAAI,
title={Can Embeddings Adequately Represent Medical Terminology? New Large-Scale Medical Term Similarity Datasets Have the Answer!},
author={Claudia Schulz and Damir Juric},
booktitle={34th AAAI Conference on Artificial Intelligence},
year={2020},
pages={to appear}
}

Contact: Claudia Schulz clauschulz1812@gmail.com or Damir Juric damir.juric@babylonhealth.com

Dependencies

This code is written in Python 3.7. The requirements are listed in requirements.txt.

pip3 install -r requirements.txt

Download the SNOMED-CT version 20190131 (https://utslogin.nlm.nih.gov/cas/login) and place the following files in the /SNOMED_files directory:

• sct2_Concept_Full_INT_20190131.txt
• sct2_Description_Full-en_INT_20190131.txt
• der2_cRefset_AttributeValueFull_INT_20190131.txt

Note: if you use a different SNOMED-CT version, this will result in different datasets!

Creating the Term Similarity Datasets

The /dataset_creation_from_SNOMED directory contains all code and data required to construct binary term similarity datasets from SNOMED CT. All datasets are tab-separated, containing instances of the form

term1 term2 1 (positive instance) or term1 term2 0 (negative instance).

The dataset creation is done in two steps:

1. creation of positive instances from SNOMED concept labels and from SNOMED concept deletions and
2. creation of negative instances from the positive ones using simple random sampling and/or advanced sampling based on Levenshtein distance.

To create all datasets run:

python3 create_datasets.py --snomed_path [location of SNOMED files if not ../SNOMED_files/] --dataset_path [location to save new datasets]

There are further arguments to control the dataset creation, however changing these will result in different datasets!

Detail on Positive Instances

positive_instances_from_labels.py and positive_instances_from_deletions.py create term pairs that form the positive instances in the datasets.

Positive Instances from Labels

positive_instances_from_labels() creates two datasets:

• FSN-SYN: for each active SNOMED concept c, create pairs FSN(c) - synonym(c) using the one fully specified name (FSN) and all the synonyms of c
• SYN-SYN: for each active SNOMED concept c, create pairs label1(c) - label2(c) using all labels (i.e. synonyms and FSN) of c

The sct2_Concept_Full_INT_20190131.txt file is used to check if concepts are active and that they are in the core module rather than the model component module, which consists of properties and descriptional concepts like 'Inactive Value'. Concept labels are then extracted from the sct2_Description_Full-en_INT_20190131.txt file.

In most cases FSNs end with a parenthesis indicating the concept's semantic type and there is exactly the same synonym without the paranthesis. Thus, parentheses are deleted from the FSN and the same synonym is disregarded to not create pairs of exactly the same labels.

Positive Instances from Deleted Concepts and their Replacement

positive_instances_from_substitutions() creates three datasets from concepts that got deleted in SNOMED and are replaced by another concept, resulting in similar concept pairs of the form deletedConceptFSN - replacementConceptFSN. SNOMED specifies various reasons for deletion and replacement. We extract concept pairs of the following reasons:

• same_as: a concept is deleted as it is a duplicate, the concept is specified to to be the same as its replacement
• possibly_equivalent_to: a concept is deleted as it is ambiguous, the concept is specified to be possibly equivalent to its replacement
• replaced_by: a concept is deleted as it is outdated or erroneous, the concept is specified to be replaced by its replacement

The type of replacement and the replacing concept are specified in der2_cRefset_AssociationFull_INT_20190131.txt.

For each deleted concept and its replacement concept, the FSNs of both concepts are used, which are obtained from the sct2_Description_Full-en_INT_20190131.txt file. Again, parentheses indicating the concept's semantic type are deleted from the FSN, as well as '[D]' in the label, which indicates that the concept is deprecated.

Detail on Negative Instances

negative_sampling_from_positive_instances() creates negative instances from the positive ones using two strategies:

1. random sampling: the first term term1 of each positive term pair is randomly matched with another term termX that does not form a positive instance with term1.
   This creates datasets with names ending _simple.
2. Levenshtein sampling: the first term term1 of each positive term pair is matched with a term termX that has smallest Levenshtein distance to term1, while not forming a positive instance with term1 or with any of its similar terms (as given by the positive instances containing term1).
   This creates datasets with names ending _advanced.

About Babylon

At Babylon we believe it’s possible to put an accessible and affordable health service in the hands of every person on earth. The technology we use is at the heart and soul of that mission. Babylon is a Healthcare Platform, currently providing online consultations via in-app video and phone calls, AI-assisted Triage and Predictive Health Assistance.

If you are interested in helping us build the future of healthcare, please find our open roles here.

Follow us on Twitter @Babylon_Eng or go to babylonhealth.com for more information