Leveraging Transformers-based models and linked data for deep phenotyping in radiology.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Linked Data Technologies

Background:

• Electronic Health Records (EHRs) predominantly use free text, hindering secondary data use for clinical research.
• Current query mechanisms for cohort definition and patient matching rely heavily on structured data and clinical terminologies.
• There is a need for advanced methods to leverage the rich information contained within clinical free text.

Purpose of the Study:

• To develop a method for the secondary use of clinical text.
• To utilize Natural Language Processing (NLP) for tagging clinical notes with biomedical terminology.
• To design an ontology for mapping and classifying identified tags, enabling phenotyping queries.

Main Methods:

• Employed transformers-based NLP models (RoBERTa) to process radiology reports and identify UMLS Concept Unique Identifiers (CUIs).
• Mapped identified CUIs to multiple biomedical ontologies (e.g., SNOMED-CT, HPO, ICD-10) for phenotyping.
• Constructed a Linked Knowledge Base (LKB) using OWL constructs for automated reasoning and querying.

Main Results:

• Developed a Linked Knowledge Base (LKB) from annotated radiology reports.
• The LKB enables expressive, automated reasoning-based queries for specific clinical criteria.
• Successfully demonstrated a pipeline for transforming free text into a queryable knowledge base.

Conclusions:

• The integration of NLP and Linked Data creates scalable knowledge bases from standard ontologies, surpassing traditional relational databases for phenotyping.
• This approach facilitates the automatic mapping of free-text entities to a LKB, enabling efficient phenotyping queries.
• The method is extensible to other languages and valuable for large-scale research databases and distributed data source management.