Text-­‐based phenotyping with Semantic Deep Learning: setting up experiments for women's health Mercedes Arguello Casteleiro1,2, Niamh Joyce2, Nava Maroto3, Maria Jesus Fernandez Prieto4, Tim Furmston1, Diego Maseda Fernandez 5, Phillip Lord6, Chris Wroe7, John Keane1, Ying Cheong2 and Robert Stevens1

Abstract

    Semantic Deep Learning (SemDeep) aims to combine Semantic Web and Deep Learning research. Vector arithmetic formulas can be applied to neural language models from Deep Learning. We set up experiments to investigate if incorporating prior knowledge (what is known about the disease) into the vector arithmetic formulas may bring a better performance. This paper investigates a SemDeep approach for text-based phenotyping of four health issues affecting women worldwide: menopause, endometriosis, miscarriage, and infertility. The candidates for the disease phenotype are n-grams that can be mapped to SNOMED CT.

Keywords

•    1   Embeddings
• SNOMED CT
• phenotype

-

Female-specific health issues such as menopause, endometriosis, miscarriage, infertility are well known. However, there are no dedicated ontologies for them in well-known repositories, such as BioPortal [2] or the Open Biological and Biomedical Ontology Library [3]. We looked into the UK SNOMED CT human-readable subset [4] and the USA Value Set Authority Center (VSAC) [5]. Only menopause has two value sets in the VSAC with more than 10 SNOMED CT concepts [5].

Considering the lack of dedicated ontologies for women’s health issues, we decided to use terms appearing in clinical evidence summaries, such as British Medical Journal (BMJ) Best Practice documents (BMJ-BP for short) [6], for our SemDeep experiments.

Figure 1 has an overview of the experiments set up. A brief overview is the following: • Creation of embeddings. We employed word2phrase from word2vec [7] to compute n-grams from 300K PubMed articles (PMSB dataset). There are 423,462 n-grams appearing 5 times or more in the PMSB dataset. The vector representations were learnt with word2vec [7]. • EXP1. This is the baseline experiment, taking the disease name as input for the cosine formula. The phenotype candidates are the top-40 ranked n-grams (highest cosine value). • EXP2 and EXP3. These experiments investigate the incorporation of prior knowledge, i.e. terms appearing in BMJ-BP and having with vector representations.

• Performance metrics. We calculate the ROC curve, precision and the median of the rank. 3. References  

References

[1] SNOMED CT , 2023 . URL: http://snomed.org/refsetpg. [2] BioPortal , 2023 . URL: https://bioportal.bioontology.org. [3] Open Biological and Biomedical Ontology Foundry, 2023 . URL: https://obofoundry.org. [4] UK SNOMED CT subsets , 2023 . URL: https://isd.digital.nhs.uk/trud. [5] USA Value Set Authority Center (VSAC) , 2023 . URL: https://vsac.nlm.nih.gov. [6] BMJ Best Practice , 2023 . URL: https://bestpractice.bmj. com. [7] word2vec , 2023 . URL: http://code.google.com/p/word2vec.