Unveiling the relationships between seismocardiogram signals, physical activity types and metabolic equivalent of task scores

Abstract

Objective: The diagnosis of metabolic syndrome and cardiovascular disorders can highly benefit from physical activity and energy expenditure assessment. In this study, we investigated the relationship between metabolic equivalent of task (MET) scores and seismocardiogram (SCG)-derived parameters. Methods: We worked with the PAMAP2 dataset and focused on the 3-axial chest acceleration data. We first segmented the 3-axial SCG signals into respiration (0-1 Hz), cardiac vibrations (1-20 Hz) and heart sounds (20-40 Hz) components. Additionally, we investigated their combinations: 0-20 Hz, 1-40 Hz and 0-40 Hz. We then windowed each signal, and extracted time and frequency domain features from each window. Using the MET scores and activity types, we trained linear regression and random forest classification models first using 80-20% split, then with leave-one-subject-out cross-validation (LOSO-CV). Additionally, we investigated the significance of each feature and axis. Results: For the 80-20% task, the best performing frequency bands were 0-1 Hz, 0-20 Hz, and 0-40 Hz, which yielded a (MET mean-squared-error, classification accuracy) pair of (0.354, 0.952), (0.367, 0.904), and (0.377, 0.914), respectively. When LOSO-CV was applied, we obtained (1.059, 0.865), (0.681, 0.868), and (0.804, 0.875) for each band, respectively. Additionally, our results revealed that the lateral axis provides the most critical information about cardiorespiratory effect of performed activities. Conclusion: Different SCG components can provide unique and substantial contributions to activity and energy expenditure assessment. Significance: This framework can be leveraged in the design of wearable systems for monitoring the activity and energy expenditure levels, and understanding their relationship with underlying cardiorespiratory parameters.