Generative Models and Sentence Transformers for the Recognition and Normalization of Continuous and Discontinuous...

Area of Science:

• Bioinformatics and Computational Biology
• Medical Informatics
• Genetics and Genomics

Background:

• Accurate extraction and normalization of genetic phenotypes from clinical reports are crucial for understanding genetic conditions and advancing personalized healthcare.
• Challenges exist in recognizing discontinuous entity spans within clinical text, hindering precise data interpretation.
• Dysmorphology relies heavily on accurate phenotype representation for diagnosis and research.

Purpose of the Study:

• To develop a system (DiscHPO) for accurate extraction and normalization of genetic phenotypes from clinical examination reports.
• To specifically address the challenge of identifying and normalizing discontinuous entity spans in dysmorphology assessments.
• To improve the representation of genetic conditions in personalized medicine through enhanced phenotype recognition.

Main Methods:

• A two-phase pipeline approach was employed, starting with a sequence-to-sequence model for named entity recognition (NER) of spans.
• An entity normalization phase utilized a sentence transformer biencoder for candidate generation and a cross-encoder reranker for concept selection.
• The system was evaluated within the context of the BioCreative VIII shared task, Track 3.

Main Results:

• The top-performing normalization model achieved an F1-score of 0.723, and the best span extraction model reached an F1-score of 0.665 on the test set.
• Both models outperformed baseline methods, demonstrating effectiveness in handling continuous and discontinuous spans.
• On the validation set, the system achieved an F1-score of 0.631 for exact matches on discontinuous spans.

Conclusions:

• Exact extraction of entity spans is not always required for successful phenotype normalization.
• Partial mention matches that capture essential concept information are sufficient for clinical utility.
• The DiscHPO system shows promise for downstream clinical applications requiring accurate genetic phenotype data.