Mortality Prediction with Bidirectional Coupled and Gumbel Subset Network on Irregularly Multivariate Time Series

Mortality prediction is a crucial challenge because of multivariate time series (MTS) complexity, which are sparse, irregularly, asynchronous and hold missing values for various reasons in a single acquisition. Various methods are proposed to deal with missing values for the final mortality prediction. However, existing models only capture the temporal dependencies within a time series and are inefficient to capture the dependencies between time series to rebuild missing values for mortality prediction. To address these challenges, in this paper, we present an end-to-end imputation and mortality prediction model, named bidirectional coupled and Gumbel subset network (BiCGSN), for mortality prediction with such irregularly multivariate time series. Our proposed model (BiCGSN) uses a recurrent network to learn the temporal dependencies (intra-time series couplings) and uses a Gumbel selector on multi-head attention to obtain the relationship between the variables (inter-time series couplings) in the forward and backward directions. Then the learned bidirectional inter-and intra-time series couplings are fused to impute missing values for further mortality prediction. We evaluate our model on PhysioNet2012 and COVID-19 datasets to imputation and predict mortality. Experiments show that BiCGSN obtains the AUC 0.869 and 0.911 on two real-world datasets respectively and outperforms all the baselines.