Multi-Task Deep Neural Networks for Irregularly Sampled Multivariate Clinical Time Series

Abstract

Multivariate clinical time series data, such as those contained in Electronic Health Records (EHR), often exhibit high levels of irregularity, notably, many missing values and varying time intervals. Existing methods usually construct deep neural network architectures that combine recurrent neural networks and time decay mechanisms to model variable correlations, impute missing values, and capture the impact of varying time intervals. The complete data matrices thus obtained from the imputation task are used for downstream risk prediction tasks. This study aims to achieve more desirable imputation and prediction accuracy by performing both tasks simultaneously. We present a new multi-task deep neural network that incorporates the imputation task as an auxiliary task while performing risk prediction tasks. We validate the method on clinical time series imputation and in-hospital mortality prediction tasks using two publicly available EHR databases. The experimental results show that our method outperforms state-of-the-art imputation-prediction methods by significant margins. The results also empirically demonstrate that the incorporation of time decay mechanisms is a critical factor for superior imputation and prediction performance. The novel deep imputation-prediction network proposed in this study provides more accurate imputation and prediction results with EHR data. Future work should focus on developing more effective time decay mechanisms for simultaneously enhancing the imputation and prediction performance of multi-task learning models.

Keywords: Electronic Health Record; Irregularly Sampled Multivariate Time Series; Multi-Task Learning; Temporal Representation Learning.

Fig. 1.

Schematic description of the proposed approach.