Synthesizing Physiological and Motion Data for Stress and Meditation Detection

Abstract

In this work, we present the synthesis of physiological and motion data to classify, detect and estimate affective state ahead of time (i.e. predict). We use raw physiological and motion signals to predict the next values of the signal following a temporal modeling scheme. The physiological signals are synthesized using a one-dimensional convolutional neural network. We then use a random forest to predict the affective state from the newly synthesized data. In our experimental design, we synthesize and predict both stress and mediation states. We show the utility of our approach to data synthesis for prediction of stress and meditation states through two methods. First, we report the concordance correlation coefficient of the synthetic signals compared to the ground truth. Secondly, we report prediction results on synthetic data that are comparable to the original ground-truth signals.