A Concise Temporal Data Representation Model for Prediction in Biomedical Wearable Devices

Abstract

Predictive analytics and event forecasting using deep learning techniques require processing long-term historical data which is infeasible in low-power wearable devices. Such devices are constrained in memory and computational power, and are pushed to their limits by resource hungry deep neural networks. Current techniques either ignore historical data, or convert temporal sequences to pattern sequences, eliminating valuable properties such as time and/or recency. The proposed model maps arbitrary-length multivariate discrete time series to a concise sequence, called mapped interval sequence (MIS). MIS retains original data properties such as time, recency, and scale, without being susceptible to missing values. Life model for time series (LMts) mapping, is capable of mapping billions of data elements with sampling rate of several kHz or higher into a sequence of 32 elements or fewer. Furthermore, a new loss function called as tolerance error is introduced to improve long-term forecasting events using LMts. In a smart health Internet of Things environment, LMts enables real-time health predictions depending on both recent and historical data. In addition, the LMts model can predict the approximate time of events, with granularity of seconds and up to years. Experimental results show that, compared to previous studies in fall prediction, LMts achieves the same 100% accuracy with a single long short-term memory layer, while covering $16{\boldsymbol \times }$ longer time period and using $80{\boldsymbol \times }$ less weight parameters. LMts is also used to forecast human fall up to 14 s in advance even with 50% missing values.