Modeling diagnostic code dropout of schizophrenia in electronic health records improves phenotypic data quality and cross-ancestry transferability of polygenic scores

Importance

Researchers commonly use counts of diagnostic codes from EHR-linked biobanks to infer phenotypic status. However, these approaches overlook temporal changes in EHR data, such as the discontinuation or “dropout” of diagnostic codes, which may exacerbate disparities in genomics research, as EHR data quality can be confounded with demographic attributes.

Objective

To address this, we propose modeling diagnostic code dropout in EHR data to inform phenotyping for schizophrenia in genomic analyses.

Design

We develop and test our diagnostic dropout model by analyzing EHR data from individuals with prior schizophrenia diagnoses. We further validate model performance on a subset of patients whose diagnoses were attained through chart review. Using PRS-CS and existing GWAS summary statistics, we first extrapolate polygenic weights. Then, we apply our dropout model’s outputs to construct a data-driven filter defining our target cohort for measuring polygenic score performance.

Setting

Our analysis utilizes EHR and genomic data from the Million Veteran Program.

Participants

To model diagnostic dropout in schizophrenia, we leverage data from 12,739 patients with a history of schizophrenia, after excluding outliers. For polygenic score analyses, we incorporate data from a potential pool of 8,385 European ancestry and 6,806 African ancestry patients with a history of schizophrenia.

Main outcomes and measures

We compare the performance of our diagnostic dropout model with alternative methodologies both in predicting diagnostic dropout on a holdout set, as well as on chart review labeled data. Using the top differential diagnosis predictors in our model, we select relevant cases by filtering out patients with a prior history of mood or anxiety disorders. We then test the impact of applying different filters for measuring polygenic score performance.

Results

When evaluated on chart review-labeled data, our model improves the area under the precision-recall curve (AUPRC) by 9.6% compared to competing methods. By applying our data-driven filter for schizophrenia, we achieve a 62% increase in the association effect size when transferring a European polygenic score to an African ancestry target cohort.

Conclusions and Relevance

These findings highlight the potential of modeling diagnostic code dropout to enhance the phenotypic quality of EHR-linked biobank data, advancing more equitable and accurate genomics research across diverse populations.

Key Points