Prediction of pediatric activity intensity with wearable sensors and bi-directional LSTM models

Abstract

Assessing activity intensity has its clinical importance to the treatment of diseases such as obesity. The metabolic equivalent of task (MET) is the objective numerical measure for assessing the intensity of general activities. Because daily activities vary, an activity cannot be easily mapped to a MET value, which makes intensity quantification of daily activities more challenging than monotonous activities. In this article, we use the data from wearable inertial measurement unit (IMU) sensors and a calorimetry machine to map the relationship between activity motions and intensities. In detail, we describe an end-to-end approach for predicting METs of preschoolers. Based on the collection of data from the two devices, we present a systematic approach to address the aforementioned challenges and to predict physical activity intensity of preschoolers. Specifically, a dynamic synchronization method is first proposed to deal with the displaced data series, which takes the dynamic time warping (DTW) as an evaluation criterion. Second, additional features are designed to reinforce the ability of intensity prediction. Third, proposed methods are tested on a two-layer bidirectional long short term memory (LSTM) network model to predict MET values. Our experimental results reveal the effectiveness of the end-to-end approach.