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

Previous studies evaluating imputation methods on clinical time series do not jointly account for spatial (e.g., across features) and temporal (across time) structures in missingness. We present a simulation model that induces realistic spatio-temporal missingness in three real-world clinical datasets by sampling from Markov chains. A variety of imputation methods including last observation carried forward (LOCF), linear interpolation, and spatio-temporal autoencoder (STAE) are applied to fill these missingness patterns. Finally, we evaluate the influence of time series imputations on a downstream prediction task. Here we show that deep STAEs outperform simpler baseline methods when missingness patterns lack spatio-temporal structure. In contrast, linear interpolations perform best on spatio-temporal patterns, contributing to sufficient downstream performances. This study demonstrates that simple linear imputations attain robust performances for clinical real-world data with spatio-temporal missingness. We provide an open benchmark to evaluate the effect of missingness on clinical studies and prediction tasks.