Deep Padding and Alignment Strategies for Irregular Multivariate Clinical Time Series

To improve the accuracy of an RNN when processing sparse and irregular multivariate clinical time series, we introduce two stacked deep learning models built on top of it, namely Padd-GRU and Alignment-driven Neural Network (ALNN). The Padd-GRU performs data-driven padding and imputation to obtain equal-length univariate and fill-in missing values, respectively. Then, the ALNN component transforms the resulting padded irregular multivariate clinical time series into a pseudo-aligned (or pseudo-regular) latent multivariate time series. We use the MIMIC-3 and PhysioNet databases to evaluate and compare our model to the state-of-the-art models on the mortality prediction task.