REAL-TIME MEDICAL ELECTRONIC DATA MINING BASED HIERARCHICAL ATTENTION MECHANISM

Data mining on clinical data has great potential to improve the treatment quality of hospital and increase the survival rate of the patients. Data-driven prediction technology strongly hinges on the data collection and analysis of patients’ vital signs. Deep neural networks are supported by a Recurrent Neural Network (RNN) architecture with Long Short-Term Memory (LSTM) units, and have achieved state-of-the-art results in a number of clinical prediction tasks. Recently, the architecture based on attention mechanism has achieved remarkable success in migration tasks, and has higher computing power in NLP (Natural Language Processing). In this paper, we recur to hierarchical attention and encoder-to-decoder based model to automatically learn features from medical records of time series of vital sign, categorical features which include demographics, hospitalization history, vital sign and laboratory tests. Moreover, instead of working as a black unexplainable box, we present the approach to extract potential informative risk factors, thereby helping doctors to make optimal decisions. Experiments show that our model is effective in extracting meaningful features, while the hierarchical attention mechanism can provide a better insight into relationships between different types of medical time series.