Electronic Health Record Signatures Identify Undiagnosed Patients With Common Variable Immunodeficiency Disease

To develop a machine learning algorithm using phenotypic patterns from electronic health records (EHRs) to expedite the identification of common variable immunodeficiency (CVID) in undiagnosed patients.The study included a CVID cohort with 186 immunologist-confirmed cases and a matched control cohort with 1,106 patients identified from a university EHR with 880,000 patients.A model algorithm named PheNET was developed utilizing CVID phecodes or groupings of International Classification of Diseases codes that best represented the phenotype of CVID, along with IgG levels, data-balancing via upsampling, and inference frameworks such as logistic regression. Phecodes were selected from entries for CVID in the Online Mendelian Inheritance in Man catalog and selected from those that occurred at significantly higher rates in the CVID cohort compared with the matched control cohort. The PheNET model was assessed for accuracy, time to diagnosis, and external validity for identifying CVID from EHRs.PheNET was 17% better than existing rare disease detection algorithms for predicting CVID. Using longitudinal cohort data, PheNET was able to identify individuals with CVID before their formal diagnosis 64% of the time, on average 244 days before their recorded diagnosis. The model was applied to select 100 undiagnosed patients most likely to have CVID, and an immunologist found 74% of these patients to possibly, likely, or definitely have CVID. PheNET was applied to 5.6 million external EHRs to identify patients diagnosed with CVID, where it was 434 times better than random sampling and able to identify 49% to 77% of an EHRs CVID cases in the 10, 000 highest scoring patients.CVID presents a diagnostic challenge given its diverse infectious, inflammatory, and autoimmune phenotype. PheNET provides a unique machine learning algorithm for accurately and rapidly identifying individuals from EHRs with CVID.The heterogenous presentation of CVID often leads to diagnostic delays with higher morbidity, mortality, and healthcare-associated costs. This study presents a novel way to reduce the time to diagnosis to circumvent such issues. Further investigation is needed to understand the strengths and limitations of the model when practically applied in a clinical setting, and which patients identified by the model truly necessitate referral to an immunologist.