Comparison of Imputation Strategies for Incomplete Electronic Health Data

Background: Missing data are a common challenge in clinical tabular datasets and can substantially affect both statistical analyses and predictive modeling in healthcare. Methods: In this study, we conducted an empirical evaluation of 8 imputation strategies-median imputation, median imputation with missing indicator, multiple imputation by chained equations (MICE), MissForest, generative adversarial imputation network (GAIN), missing data imputation using autoencoders (MIWAE), self-attention-based imputation for time series, and gated recurrent unit with decay-across 3 heterogeneous real-world clinical datasets. The methods were assessed under controlled missingness simulations covering different mechanisms (missing completely at random, missing at random [MAR], and missing not at random [MNAR]) and missingness rates ranging from 10% to 90%. We evaluated imputation quality using multiple statistical metrics and examined its relationship with downstream classification performance across multiple classifiers. Results: Across the considered settings, MICE and MissForest showed consistently strong and stable performance, while deep generative approaches such as GAIN and MIWAE exhibited higher variability, particularly under the MAR and MNAR mechanisms and at higher missingness rates. Importantly, we observed that improvements in imputation accuracy did not necessarily translate into better downstream predictive performance. Conclusion: Rather than drawing general conclusions about method superiority, the results highlight the sensitivity of imputation outcomes to data characteristics, missingness mechanisms, and evaluation objectives. These findings underscore the importance of aligning imputation strategy selection with the intended downstream task and data regime when working with heterogeneous clinical tabular data.