Disease Prediction on Wearables Activity Data Using Stacked Bidirectional LSTM

Abstract

Plenty of sleep and suitable activity behavior play an eventful role in preventing or even curing of a lot of chronic diseases situation. It is an important task to find the relations between chronic situations and sleep and activity. A considerable novel data source is provided by the speedy increase in the wearable health devices’ popularity. To solve the problem of the lack of clinical data from a large number of experimenters, we propose a novel unsupervised time series representation learning model which is called tim2vec. The vector representation of activity data is learned by tim2vec in the way of taking the activity level and the periodicity of human activity into account. Then, the learned vector is utilized to enhance the performance of the specific task semi-supervised learning models. A stacked bidirectional LSTM (SB-LSTM) for exploiting the disorder diagnosis with activity data captured by the wearable devices is used. Experimental assessment using activity data shows that the method mentioned above classifies and explains better than the traditional symbolic representation methods and some advanced prediction models.