Phenotypes Extraction from Text: Analysis and Perspective in the LLM Era

• Main Authors: Baddour, Moussa, Paquelet, Stephane, Rollier, Paul, De Tayrac, Marie, Dameron, Olivier, Labbe, Thomas
• Format: Conference Proceeding
• Language: English
• Subjects: Annotations; Data mining; Detectors; embed dings; entity linking; genetic; Intelligent systems; Large language models; LLM; Medical diagnostic imaging; Ontologies; phenoBERT; phenotype; Phenotypes; Semantics; Syntactics
• ISSN: 2767-9802
• DOI: 10.1109/IS61756.2024.10705235

Collecting the relevant list of patient phenotypes, known as deep phenotyping, can significantly improve the final diagnosis. As textual clinical reports are the richest source of phenotypes information, their automatic extraction is a critical task. The main challenges of this Information Extraction (IE) task are to identify precisely the text spans related to a phenotype and to link them unequivocally to referenced entities from a source such as the Human Phenotype Ontology (HPO). Recently, Language Models (LMs) have been the most suc-cessful approach for extracting phenotypes from clinical reports. Solutions such as PhenoBERT, relying on BERT or GPT, have shown promising results when applied to datasets built on the hypothesis that most phenotypes are explicitly mentioned in the text. However, this assumption is not always true in medical genetics. Hence, although the LMs carry powerful semantic abilities, their contributions are not clear compared to syntactic string-matching steps that are used within the current pipelines. The goal of this study is to improve phenotype extraction from clinical notes related to genetic diseases. Our contributions are threefold: First, we provide a clear definition of the phenotype extraction task from free text, along with a high-level overview of the involved functions. Second, we conduct an in-depth analysis of PhenoBERT, one of the best existing solutions, to evaluate the proportion of phenotypes predicted with simple string-matching. Third, we demonstrate how utilizing and incorporating large language models (LLMs) for span detection step can improve performance especially with implicit phenotypes. In addition, this experiment revealed that the annotations of existing dataset are not exhaustive, and that LLM can identify relevant spans missed by human labelers.