Multi-scale Temporal Memory for Clinical Event Time-Series Prediction

The objective of this work is to develop and study dynamic patient-state models and patient-state representations that are predictive of a wide range of future events in the electronic health records (EHRs). One challenge to overcome when building predictive EHRs representations is the complexity of multivariate clinical event time-series and their short and long-term dependencies. We address this challenge by proposing a new neural memory module called Multi-scale Temporal Memory (MTM) linking events in a distant past with the current prediction time. Through a novel mechanism implemented in MTM, information about previous events on different time-scales is compiled and read on-the-fly for prediction through memory contents. We demonstrate the efficacy of MTM by combining it with different patient state summarization methods that cover different temporal aspects of patient states. We show that the combined approach is 4.6% more accurate than the best result among the baseline approaches and it is 16% more accurate than prediction solely through hidden states of LSTM.