Automatic Prediction of Multiple Associated Diseases Using a Dual-Attention Neural Network Model

In recent years, disease prediction based on electronic health records (EHR) has attracted extensive attention in the field of biomedical text mining. However, the existing work has two issues. First, most of the existing methods focus on the prediction of a single disease and little attention is paid to the prediction of multiple associated diseases. Second, these methods usually use simple feature modeling, and fail to fully capture and mine the information from EHR, which usually contains two main information: the textual description and physical indicators. To address these issues, we design a dual-attention neural network model to predict the probability of coronary heart disease and kidney disease in hypertension patients. Specifically, the proposed model consists of three main parts: a textual module, a numerical module and a global BiLSTM. Given one piece of EHR, the textual module is utilized for encoding the textual information, such as diagnosis texts. The numerical module handles the numerical indicators, such as physical indicators. The dualattention mechanism enables the model to better capture the intrinsic and implicit semantic features behind the clinic texts and numerical indicators, respectively. The experimental results on the datasets show the effectiveness of the proposed model, and our model outperforms previous methods and strong neural baselines by a large margin. Meanwhile, the attention mechanism can capture the risk factors between the associated diseases.