Bias in administrative coding for mitral valve prolapse: Implications for electronic health record research

Electronic health records and administrative claims are increasingly utilized for conducting observational studies, pragmatic clinical trials, and training machine learning models. For accurate and reliable conclusions, input data requires validation. We sought to evaluate International Classification of Diseases, Tenth Revision (ICD-10) diagnosis coding for Mitral valve prolapse (MVP).

We compared ICD-10 diagnosis coding for MVP with echocardiographic evidence of MVP in a large academic center over the years 2018 - 2024.

Among 5,520 patients with an MVP ICD-10 code (I34.1), 1,883 had echocardiographic confirmation (positive predictive value 34.1%, 95% CI 32.9% - 35.4%). Among 5,667 patients with echocardiographic confirmed MVP, 3,784 lacked the corresponding ICD-10 code (sensitivity was 33.2%, 95% CI 32.0 % - 34.5%). In sub-analysis with granular covariates (N = 1,183), false-positive coding was associated with higher prevalence of atrial fibrillation, heart failure, and coronary artery disease than patients with confirmed MVP in univariate analysis. In multivariable analysis, heart failure was associated with increased odds of false-positive ICD coding (aOR 1.77, 95% CI 1.08 - 2.89; p = 0.02).

These findings suggest that administrative MVP codes do not accurately identify patients with the disease and appear to introduce non-random phenotype bias with implications for EHR research and machine learning.