Multi-disease prediction using LSTM recurrent neural networks

Prediction of future clinical events (e.g., disease diagnoses) is an important machine learning task in healthcare informatics research. In this work, we propose a deep learning approach to perform multi-disease prediction for intelligent clinical decision support. The proposed approach utilizes a long short-term memory network and extends it with two mechanisms (i.e., time-aware and attention-based) to conduct multi-label classification based on patients' clinical visit records. The former mechanism (time-aware) is used to handle the temporal irregularity across clinical visits whereas the latter mechanism (attention-based) assists in determining the importance of each visit for the prediction task. Using a large clinical record data set (over 5 million records) collected from a hospital in Southeast China, we show that our proposed approach outperforms a variety of traditional and deep learning methods in predicting future disease diagnoses. We further study the impacts of different time interval choices for the time-aware mechanism and compare the performances of existing attention-based mechanisms with the one proposed in our study. Our work has implications for supporting physician diagnoses via the use of intelligent systems and more broadly for improving the quality of healthcare service.