Title:Bayesian Recurrent Framework for Missing Data Imputation and Prediction with Clinical Time Series

Abstract:Real-world clinical time series data sets exhibit a high prevalence of missing values. Hence, there is an increasing interest in missing data imputation. Traditional statistical approaches impose constraints on the data-generating process and decouple imputation from prediction. Recent works propose recurrent neural network based approaches for missing data imputation and prediction with time series data. However, they generate deterministic outputs and neglect the inherent uncertainty. In this work, we introduce a unified Bayesian recurrent framework for simultaneous imputation and prediction on time series data sets. We evaluate our approach on two real-world mortality prediction tasks using the MIMIC-III and PhysioNet benchmark datasets. We demonstrate strong performance gains over state-of-the-art (SOTA) methods, and provide strategies to use the resulting probability distributions to better assess reliability of the imputations and predictions.