Benchmarking imputation methods on real-world clinical time series with simulated spatio-temporal missingness.

Deep spatio-temporal autoencoders outperform simpler methods in imputing missing data when spatio-temporal structures are absent in clinical time series.

• Why it matters: Accurately handling missing data in clinical time series is crucial for reliable downstream predictions, yet current methods often overlook the combined spatial and temporal missingness patterns.
• What they did: A simulation model using Markov chains was developed to generate realistic spatio-temporal missingness in three real-world clinical datasets, and various imputation methods, including LOCF, linear interpolation, and STAE, were evaluated.
• The result: Linear interpolation proved most effective for spatio-temporal missingness, while deep STAE excelled when such structure was lacking, highlighting the robustness of simple methods and providing an open benchmark for future clinical data analysis.