Data-driven methods for analyzing ballistocardiograms in longitudinal cardiovascular monitoring
Cardiovascular disease (CVD) is the leading cause of death in the US; about 48% of American adults have one or more types of CVD. The importance of continuous monitoring of the older population, for early detection of changes in health conditions, has been shown in the literature, as the key to a successful clinical intervention. We have been investigating environmentally-embedded in-home networks of non-invasive sensing modalities. This dissertation concentrates on the signal processing techniques required for the robust extraction of morphological features from the ballistocardiographs (BCG), and machine learning approaches to utilize these features in non-invasive monitoring of cardiovascular conditions. At first, enhancements in the time domain detection of the cardiac cycle are addressed due to its importance in the estimation of heart rate variability (HRV) and sleep stages. The proposed enhancements in the energy-based algorithm for BCG beat detection have shown at least 50% improvement in the root mean square error (RMSE) of the beat to beat heart rate estimations compared to the reference estimations from the electrocardiogram (ECG) R to R intervals. These results are still subject to some errors, primarily due to the contamination of noise and motion artifacts caused by floor vibration, unconstrained subject movements, or even the respiratory activities. Aging, diseases, breathing, and sleep disorders can also affect the quality of estimation as they slightly modify the morphology of the BCG waveform.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 License.