Liability threshold model-based disease risk prediction based on electronic health record phenotypes.

Area of Science:

• Genomics
• Biomedical Informatics
• Computational Biology

Background:

• Electronic health records (EHRs) are valuable for genomic research but face challenges in accurate case-control labeling.
• Current methods using phenotype codes often lead to suboptimal analyses in downstream research.
• Developing robust methods for EHR data utilization is crucial for advancing genetic studies.

Purpose of the Study:

• To introduce the liability threshold phenotypic integration (LTPI) method for deriving continuous phenotypes from EHR data.
• To enhance disease risk prediction and genome-wide association study (GWAS) power.
• To provide insights into nontarget traits associated with specific diseases.

Main Methods:

• The LTPI method combines genetic relatedness with diverse phenotypic data (diagnosis codes, family history, lab results, biomarkers).
• An automatic trait selection algorithm is employed to optimize model performance.
• The method was validated using simulations and applied to the eMERGE network and UK Biobank datasets.

Main Results:

• LTPI demonstrated consistent performance gains in disease risk prediction compared to conventional methods.
• The method significantly improved genome-wide association study (GWAS) power.
• LTPI maintained similar false-positive rate control compared to existing approaches.

Conclusions:

• The liability threshold phenotypic integration method offers a superior approach for phenotype definition in EHR-based genomic research.
• This method enhances the utility of EHRs for disease risk prediction and genetic discovery.
• LTPI provides a powerful tool for leveraging complex phenotypic data in large-scale genetic studies.