Investigation of synonym expansion and self-alignment pretraining for enhancing Human Phenotype Ontology concept recognition.

Abstract

Accurate identification of Human Phenotype Ontology (HPO) terms from biomedical text is crucial for disease diagnosis and analysis. However, traditional named entity recognition (NER) methods often fall short by overlooking synonyms and entity variants, resulting in reduced accuracy and coverage. To address these limitations, we introduce three strategic enhancements to improve HPO concept recognition. The first strategy augments the training data by incorporating HPO synonyms at the instance level, which enhances the model's ability to recognize diverse phenotypic expressions. The second strategy introduces HPOBERT, a semantic approach that models HPO synonyms through self-aligned pretraining. This method closely aligns synonym representations while effectively distinguishing them from non-synonyms, thereby improving the model's ability to differentiate between concepts. The third strategy integrates both the instance level and semantic approach. We evaluated these enhancement strategies on four clinical text datasets annotated with HPO concepts. The results demonstrate significant improvements in classification accuracy, recall, and both micro and macro F1 scores. Additionally, our enhancement strategies showed strong performance in Named Entity Normalization (NEN) after NER, accurately linking recognized HPO concepts to standardized knowledge bases. Specifically, we observe improvements of 2.44% and 4.38% in NEN-F on the gold and silver standard datasets, respectively, highlighting the effectiveness of our approach. The source code is available at https://github.com/ZhuLab-Fudan/HPOTagger.