Ontology-driven and weakly supervised rare disease identification from clinical notes.

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Computational text phenotyping aims to identify patient disorders from clinical notes.
• Rare disease identification is challenging due to limited data and the need for expert annotation.

Purpose of the Study:

• To propose a novel method for rare disease identification using ontologies and weak supervision.
• To leverage pre-trained contextual representations (e.g., BERT) for improved text phenotyping.

Main Methods:

• An ontology-driven framework involving Text-to-UMLS concept linking using a Named Entity Recognition and Linking (NER+L) tool (SemEHR) and weak supervision.
• A UMLS-to-ORDO matching step to connect UMLS concepts to rare diseases in the Orphanet Rare Disease Ontology (ORDO).
• Development of a phenotype confirmation model using weak supervision to enhance Text-to-UMLS linking without expert-annotated data.

Main Results:

• Significant precision improvements (30-50% absolute score) in Text-to-UMLS linking with minimal recall loss compared to existing NER+L tools.
• Consistent performance across diverse clinical datasets, including MIMIC-III discharge summaries, MIMIC-III radiology reports, and NHS Tayside brain imaging reports.
• Successful extraction of rare disease cases often missed by traditional structured data methods (e.g., ICD codes).

Conclusions:

• The study validates a weakly supervised NLP pipeline for clinical notes, requiring minimal human annotation for validation and testing.
• The proposed deep learning approach effectively utilizes ontologies, NER+L tools, and contextual representations for rare disease phenotyping.
• Natural Language Processing (NLP) offers a complementary approach to traditional methods for improving rare disease estimation in clinical data.