Imputation methods for serologic biomarkers in inflammatory bowel disease

Serologic biomarkers have emerged as a powerful tool for the diagnosis of Inflammatory Bowel Disease (IBD) and the differentiation between subgroups of IBD. However, missingness in serologic data can adversely affect the efficacy of any form of statistical or machine learning analysis, leading to biased predictions. This paper provides a thorough comparison of multiple imputation models that can be used for the imputation of serologic data under different missingness scenarios. All major forms of missingness, including Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR), were explored in relation to the serologic data. The imputation models used in this study encompass Multiple Imputation (MI) using Chained Equations (MICE), Iterative Imputer (II), and Autoencoders (AE). Across three real IBD cohorts and 2,400 simulated scenarios spanning MCAR/MAR/MNAR and 5-40% missingness, we evaluated imputers on direct accuracy, inferential signal, and predictive utility. No single method is universally optimal: iterative imputers (II-BR/KNN/RF) tend to lead at low-moderate missingness, whereas autoencoder-based (AE/VAE) approaches are more robust as missingness increases; all analyses are performed within-cohort to avoid information leakage.