Experiments aligning HPO and SNOMED medical ontologies to be able to map from SNOMED terms to the HPO term which best matches it.
HPO term namesHPO term names + synonyms + descriptionSNOMED term namesSNOMED term names + synonyms + description
SNOMED is a large medical ontology for general medical use that can verge into the sprawling with a very large number of terms and quite a lot of repetition. For genetic disease community the alternative HPO ontology has become popular, it is tightly maintained and tries to stay minimalistic and focused on
In NLP systems attempting to draw insights from medical records, such as extracting a set of symptoms, may be focused on a particular ontology. However, seeing as there are a variety of ontologies in common use systems can be made more powerful and general if they are able to include information from multiple ontologies.
Between ontologies there are some explicit mappings of terms but only in the minority of cases. Therefore, it is useful to build a system which is able to automatically account for cases without explicit term linking.
I will be using the recently released BERT sentence embedding model from Google to encode the names, synonyms and descriptions of terms to vectors of the same length.
For ease of use in getting vectors from new sentences I will be using BERT as a service which came out of Tencent AI Labs and makes it super easy to deploy BERT locally and remotely. In my case I was working on the train but my laptop does not have a GPU so I set up my home PC as a BERT server so I could hit it every time I wanted to get a sentence vectorised.
To fine tune the BERT model I am attempting to use a full text archive of PubMed from their open access portal as a training corpus.
Using the large, uncased english BERT model without fine tuning to create sentence embeddings from just the names of both SNOMED and and HPO terms.
First attempt at mapping SNOMED to HPO terms managed to get the correct HPO term in the top 10 most cosine similar terms 35.7% of the time.
Second experiment is to see if averaging sentence embeddings of the name, description and synonyms of terms creates stronger link.
In a first experiment in this strand only the HPO term vectors were created from averaged vectors of name, description and synonyms. When now compared to the name based SNOMED vectors the top 10 score rose to 50.7%.
After building rich SNOMED vectors and comparing them to the original name based HPO vectors a top 10 accuracy of 63.6% is achieved.
Now, combing rich embeddings of HPO and SNOMED terms a top 10 similarity score of 79.5% is achieved. When looking only at the top 3 most similar HPO terms to a SNOMED term the accuracy only falls to 71.27%. 58.4% of the time the most similar HPO term is correct.
Other ideas to try in this vein include averaging in vectors of parent categories or building multiple vectors for each HPO and SNOMED term including embeddings of the parent classes of the respective terms to include more information about the hierarchy of terms.
Seeing as SNOMED terms can often be degenerate it could be useful to average hierarchy embedding of all degenerate terms to create a wider view of the context of a term.
Third experiment is to see if fine tuning the BERT model on pubmed or another large medical text corpus improves the power of the sentence embeddings. Fine tuning will need to be done on a cloud TPU due to time and memory constraints.
Corpus to fine tune on acquired from the PMC open access archives