Investigating the effect of intra-subject variability in seismocardiogram analysis

Abstract

Seismocardiogram (SCG) signal corresponds to the local chest vibrations originating from the contraction of the heart and has been leveraged to assess cardiovascular parameters and pathologies, especially in studies focusing on wearable system design. Two common challenges encountered in the analysis of SCG are the presence of inter-subject and intra-subject variability, affecting the generalizability and performance of the models. In this work, we focus on the intra-subject variability to show how the local body compositions and surrounding systems affect the morphology of the SCG signals. More specifically, we investigated the effects of sensor location and respiration phases on the SCG signals. Analysis of tri-axial SCG signals from 16 different locations on torso was treated as a 16-class classification problem, whereas the effect of inhalation and exhalation was formulated as a binary classification task. Using features from 5 different groups, our models achieved a distinction rate of %92 and %90 for the sensor location and respiration effect, respectively. Additionally, we showed that the effect of intra-subject variability is not the same for different SCG axes or feature groups. Overall, intra-subject variability has a great impact on the SCG characteristics, thus the inaccuracies stemming from sensor location or surrounding physiological systems should be minimized during analyses to have more generalizable models.