Mapping layperson medical terminology into the Human Phenotype Ontology using neural machine translation models
In the medical domain there exists a terminological gap between patients and caregivers and the healthcare professionals. This gap may hinder the success of the communication between healthcare consumers and professionals in the field, with negative emotional and clinical consequences. In this work,...
| Autores: | , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | Fundació Sant Joan de Déu |
| Repositorio: | r-FSJD. Repositorio Institucional de Producción Científica de la Fundació Sant Joan de Déu |
| OAI Identifier: | oai:fsjd.fundanetsuite.com:p22233 |
| Acceso en línea: | https://fsjd.fundanetsuite.com/Publicaciones/ProdCientif/PublicacionFrw.aspx?id=22233 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine translation Word embedding Deep learning Medical informatics Deep phenotyping Human Phenotype Ontology |
| Sumario: | In the medical domain there exists a terminological gap between patients and caregivers and the healthcare professionals. This gap may hinder the success of the communication between healthcare consumers and professionals in the field, with negative emotional and clinical consequences. In this work, we build a machine learning-based tool for the automatic translation between the terminology used by laypeople and that of the Human Phenotype Ontology (HPO). HPO is a structured vocabulary of phenotypic abnormalities found in human disease. Our method uses a vector space to represent an HPO-specific embedding as the output space for a neural network model trained on vector representations of layperson versions and other textual descriptors of medical terms. We explored different output embeddings coupled to different neural network architectures for the machine translation stage. We compute a similarity measure to evaluate the ability of the model to assign an HPO term to a layperson input. The best-performing models resulted with a similarity higher than 0.7 for more than 80% of the terms, with a median between 0.98 and 1. |
|---|