Imputation methods for serologic biomarkers in inflammatory bowel disease

Area of Science:

• Biomedical Informatics
• Statistical Modeling
• Inflammatory Bowel Disease Research

Background:

• Serologic biomarkers are crucial for Inflammatory Bowel Disease (IBD) diagnosis and subtyping.
• Missing data in serologic datasets can compromise statistical and machine learning analyses, leading to biased predictions.

Purpose of the Study:

• To comprehensively compare various multiple imputation models for serologic data in IBD.
• To evaluate imputation performance under different missing data mechanisms (MCAR, MAR, MNAR) and varying percentages of missingness.

Main Methods:

• Utilized three distinct imputation techniques: Multiple Imputation by Chained Equations (MICE), Iterative Imputer (II), and Autoencoders (AE).
• Assessed imputation models across three real-world IBD cohorts and 2,400 simulated scenarios.
• Evaluated performance based on direct accuracy, inferential signal preservation, and predictive utility.

Main Results:

• No single imputation method demonstrated universal superiority across all scenarios.
• Iterative imputation methods (II-BR, KNN, RF) generally performed better at low to moderate missingness levels.
• Autoencoder-based methods (AE, VAE) showed greater robustness with increasing percentages of missing data.

Conclusions:

• The choice of imputation method for IBD serologic data should be guided by the extent of missingness.
• Iterative imputers are suitable for less incomplete datasets, while autoencoders offer a more robust solution for highly incomplete data.
• Within-cohort analysis is essential to prevent information leakage and ensure reliable imputation results.