A New Temporal Abstraction for Health Diagnosis Prediction using Deep Recurrent Networks

Temporal health data, either as electronic health record or from nursery home care units, usually include multivariate sparse temporal health data different from a regular time-series. Conventional neural network models cannot be used in such data; recurrent neural networks (RNN) (such as with long-term short memory (LSTM) cells) are used to model time-series. However, long-term variable-length sparse temporal data are not suitable for an efficient learning with RNN models. This research presents a novel pattern extraction technique for use in diagnosis prediction using deep learning techniques in recurrent neural networks. To predict diagnosis from such data, a window-based data abstraction technique called intensity temporal sequence (ITS) is proposed and tested. ITS enables presenting long-term sparse temporal data as a fixed-length sequence suitable for training by deep recurrent networks. To evaluate the method against other techniques, such as recent temporal patterns (RTP), a pattern simulator and anomaly injection method is developed to generate 100,000 patient records with 10 possible diseases over 10,000 units of time. The results indicate that ITS performs slightly better than RTP in terms of accuracy when using techniques other than LSTM. However, only ITS is suitable for learning LSTM; a model which performs better in terms of accuracy.