Multi-task deep representation learning method for electronic health records

Abstract

Electronic health records (EHRs) data plays an important role in the development of healthcare undertaking. There are many challenges in mining EHRs, such as temporality, irregularity, sparsity, bias, etc. Thus effective feature extraction and representation are key steps before any further applications. In this paper, we propose a multi-task deep representation learning method (MTDRL) with the objective of extracting the valuable clinical information from raw data and learning an effective and interpretable patient representation. Firstly, MTDRL utilizes Bidirectional Gated Recurrent Unit (BiGRU) as an encoder to learn the hidden state vectors which are used as inputs for the following networks. This encoding part can be seen as a shared network for all tasks. Secondly, patient's in-hospital mortality prediction and sequence reconstruction are simultaneously conducted based on the encoding network. Specifically, an attention mechanism and a fully-connected layer are incorporated in prediction task and BiGRU is implemented in the other task to reconstruct the visit sequences. Finally, we apply MTDRL to real EHRs data and the experimental results demonstrate that MTDRL is capable of learning more effective patient representation and has a significant improvement in the performance of patient's in-hospital mortality prediction. Meanwhile, the prediction results can be effectively interpreted with the attention mechanism and provide a clinically meaningful references.