Date of Award

5-2024

Document Type

Thesis - Closed Access

Degree Name

MS in Human Genetics

First Advisor

Laura Hercher, MS, CGC

Second Advisor

Dr. John Greally

Third Advisor

Monisha Sebastin, MS, CGC

Abstract

Human Phenotype Ontology (HPO) terms help identify and rank causative genes in exome/genome sequencing for patients with rare disease, yet diagnostic rates remain low. GenomeDiver reanalyzes phenotypes to prioritize features that distinguish variants and diseases. Manual extraction of phenotypic terms from Electronic Health Records is time-consuming, providing opportunities for natural language processing (NLP) to support the diagnostic process. We evaluated NLP system performance for Clinphen and Elastex in extracting HPO terms for use in GenomeDiver. 14 patients with various note types were randomly selected from the NYCKidSeq study. Two annotators independently extracted HPO terms from the 56 total notes. A third investigator adjudicated, creating the gold standard (GS) dataset. Pooled Kappa determined interannotator agreement. NLP’s were evaluated by comparing each system’s extracted HPO terms to the GS, obtaining precision, recall and F1. GS, Clinphen, and Elastex averaged identifying 6.96, 5.66, and 14.9 HPO terms per note, respectively, for a total of 239, 183, and 337 unique HPO terms across all notes. Interannotator agreement for GS = 0.67. Elastex’s recall was higher (0.69 vs. 0.44), while Clinphen’s precision was higher (0.64 vs 0.55). ClinPhen demonstrated higher precision, allowing more curated terms to be sent back to clinicians through GenomeDiver. Yet systems with higher recall are easier for providers to identify true positives and discard false positives from the list of phenotypic terms generated by NLP evaluation. Awareness of the limitations of NLP systems may optimize the utility of automated HPO extraction for the purposes of GenomeDiver.

Under author imposed embargo.
Available for download on Thursday, April 30, 2026

Share

COinS