Context-aware imputation for clinical time series

Missing data has been widely recognized as a key challenge of clinical time series analysis, which hinders the practical application of data-driven approaches to clinical data analytics [1], [2]. Various methods have been proposed to perform the time series imputation to alleviate this issue, yet most of them impose strong assumptions on the missing data, for instance, locality in Gaussian Process based models [3], lowrankness and temporal regularity in matrix/tensor factorization models [4], etc. More recently, researchers proposed to apply the Recurrent Neural Networks (RNNs) to tackle the missing data imputation problem for time series, where the RNNs try to capture and summarize the temporal dynamics using hidden state vectors [5]-[7].