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

Electronic health record (EHR)-based phenotyping underpins genome-wide association studies, yet current ICD-code phenotypes rely heavily on manually curated lists such as Phecodes. These definitions are labour-intensive to maintain, inherently subjective, and may omit clinically relevant diagnostic codes, reducing study power. Advances in text embedding models offer an opportunity to automate and standardize ICD-based phenotype construction.

We developed Phecoder, an ensemble of pre-trained text embedding models that rank ICD codes by similarity from free-text phenotype descriptions. Nine embedding models and multiple unsupervised ensemble rank-fusion methods were evaluated against 1,125 PhecodeX phenotypes. Retrieval performance was assessed using recall and average precision at top-100 (R@100, AP@100). Expert clinical review of six neuropsychiatric phenotypes was undertaken to identify relevant ICD codes absent from PhecodeX. Cohort sizes under these new definitions were compared with PhecodeX across sex and ancestry strata in the Million Veteran Program (MVP).

Among individual models, Qwen3-Embedding-4B achieved the highest median recall (R@100 = 0.86). Ensemble rank-fusion further improved R@100 by 3%, and median AP@100 by 8%. Expert review confirmed that Phecoder retrieved additional clinically relevant ICD codes beyond PhecodeX across all six neuropsychiatric case studies. Median potential case expansion increased by 200%, with 700% increases for bipolar disorder and 2000% increase for eating disorders.

Manually defining ICD phenotypes has been critiqued as subjective, potentially yielding overly restrictive definitions that miss relevant codes. To address this issue, Phecoder algorithmically identifies relevant codes for ICD-based phenotyping. Phecoder extracts relevant ICD codes to expand the potential case pool across different demographic groups. Phecoder is easily applicable to future ICD-code releases and across different ICD coding versions that are used in different countries. Taken together, Phecoder has the potential to improve reproducibility in EHR data research.

This research was supported by the Department of Veterans Affairs MVP (MVP-000, MVP-076 and MVP-096). The MVP is supported by the Office of Research and Development, Department of Veterans Affairs. The authors thank the MVP staff, researchers, and volunteers, who have contributed to MVP, and especially who previously served their country in the military and now generously agreed to enroll in the study (see mvp.va.gov for more information). The contents do not represent the views of the U.S. Department of Veterans Affairs or the United States Government. This study was supported by the Veterans Affairs Merit grants: BX006500 (to D.B.) and BX004189 (to P.R.). This work was supported by the National Institutes of Health (NIH): R01MH125246 (to P.R.), R01AG078657 (to G.V.), R01AG067025 (to P.R.), and U24AG087563 (to P.R.).