Phecoder: semantic retrieval for auditing and expanding ICD-based phenotypes in EHR biobanks

Area of Science:

• Computational biology
• Biomedical informatics
• Genomics

Background:

• Electronic health record (EHR)-based phenotyping is crucial for genome-wide association studies (GWAS).
• Current methods rely on manually curated ICD code lists (e.g., Phecodes), which are labor-intensive, subjective, and may miss relevant codes, reducing study power.
• Advances in text embedding models offer a path to automate and standardize ICD-based phenotype construction.

Purpose of the Study:

• To develop and evaluate Phecoder, an ensemble of text embedding models for automated ICD-based phenotyping.
• To compare Phecoder's performance against existing PhecodeX phenotypes.
• To assess the impact of Phecoder on cohort size and diversity.

Main Methods:

• Developed Phecoder, an ensemble of pre-trained text embedding models to rank ICD codes by similarity to free-text descriptions.
• Evaluated nine embedding models and unsupervised ensemble rank-fusion methods against 1,125 PhecodeX phenotypes.
• Assessed retrieval performance using recall and average precision at top-100 (R@100, AP@100).
• Conducted expert clinical review for six neuropsychiatric phenotypes and compared cohort sizes in the Million Veteran Program (MVP).

Main Results:

• Phecoder, particularly with ensemble rank-fusion, significantly improved retrieval performance (e.g., +3% R@100, +8% AP@100) over individual models.
• Expert review confirmed Phecoder identified additional clinically relevant ICD codes not present in PhecodeX.
• Phecoder demonstrated substantial potential case expansion (median 200%, up to 2000% for specific disorders), increasing cohort completeness across demographic groups.

Conclusions:

• Phecoder offers an automated, objective, and efficient approach to ICD-based phenotyping, addressing limitations of manual curation.
• The framework enhances the identification of relevant diagnostic codes, leading to more comprehensive and reproducible EHR research.
• Phecoder's applicability to future ICD code versions and its potential to improve cohort completeness across diverse populations highlight its value in biomedical research.