SafeImpute Offers Reliable Clinical Data Imputation with Error Control

Clinical data often suffers from pervasive missingness due to irregular patient visits and test ordering, posing a significant challenge for downstream analysis and decision-making. While many imputation methods exist, they typically focus on average accuracy and lack mechanisms to guarantee the reliability of individual imputed values for high-stakes clinical use. SafeImpute addresses this by offering a reliable imputation framework specifically designed for irregular and sparse clinical longitudinal records. It constructs an event graph to capture both intra-patient temporal trajectories and inter-patient clinical similarity, learning imputations via a two-relation Graph Neural Network (GNN) with adaptive fusion. Crucially, SafeImpute provides reliability guarantees by converting a proxy risk score into conformal p-values and applying the Benjamini-Hochberg procedure to control the False Discovery Rate (FDR) of unacceptable errors among released imputations at a user-specified tolerance. Experiments on real-world clinical datasets demonstrate strong imputation accuracy alongside robust error control.