Benchmarking Imputation Strategies for Missing Time-Series Data in Critical Care Using Real-World-Inspired Scenarios

Handling missing data remains a central issue in ICU time-series analysis, where data gaps often stem from non-random factors such as sensor disconnections or clinical workflows. In this study, we systematically benchmarked several imputation strategies using monitoring data from MIMIC-IV and designed masking scenarios that reflect common ICU patterns, including random dropouts, temporary monitoring interruptions, and sensor-specific failures. We compared simple statistical approaches (mean, LOCF, interpolation), classical machine learning techniques (MICE, MissForest), and deep learning models (SAITS, BRITS, US-GAN, GP-VAE). SAITS, based on Transformer architecture, achieved the best performance in most settings. However, linear interpolation—despite its simplicity—yielded robust estimates in short univariate gaps and occasionally performed comparably to neural models. Our findings suggest that while deep learning methods improve overall imputation accuracy, simpler and more interpretable approaches may be sufficient for many ICU applications. This work introduces a practical framework for evaluating time-series imputation strategies under realistic constraints, with a focus on clinical relevance and operational deployability.