PhD defence of James Skoric – Wearable health monitoring with seismocardiography and generative modeling

Abstract

Cardiovascular disease (CVD) remains a leading cause of death and disability worldwide, underscoring the critical need for early detection and continuous monitoring. Wearable technologies have emerged as a promising solution, offering non-invasive, real-time assessment of physiological signals in daily life. Among these, seismocardiography (SCG)—a technique that captures chest wall vibrations from cardiac activity using low-cost accelerometers—has the potential to enable affordable, comfortable, and continuous monitoring. However, several limitations hinder its widespread adoption. SCG signals are highly susceptible to motion artifacts during ambulatory activity, complicating their use in real-world settings. Furthermore, the scarcity of large, annotated SCG datasets limits the development and generalization of machine learning models. Compared to conventional modalities, SCG remains underutilized for extracting detailed cardiac features or supporting clinical use. This thesis aims to address these limitations and advance the utility of SCG for wearable cardiac monitoring by improving signal robustness, mitigating data scarcity, and demonstrating its functional relevance across diverse applications.

First, we develop a novel motion artifact reduction algorithm to improve SCG signal quality in ambulatory conditions. We show that our algorithm improved heart rate estimation accuracy during walking, up to a −19 dB signal-to-noise ratio without electrocardiography (ECG). Our solution is directly applicable to SCG monitoring in daily life. Second, we tackle data scarcity with a generative adversarial network to create synthetic, individualized SCG heartbeats. We show that we can successfully replicate SCG signal morphology with tunable features and demonstrate its utility in a lung volume classification task, with synthetic data matching the accuracy of real data, and an augmentation approach increased accuracy by 3%. Third, we develop a generative adversarial network that uses 6-axis vibrational cardiography (VCG) to reconstruct multiple cardiac waveforms – electrocardiography (ECG), impedance cardiography (ICG), photoplethysmography (PPG), and non-invasive blood pressure (NIBP). We demonstrate that the approach achieves strong morphological and temporal alignment between estimated and reference signals, with median Pearson correlation coefficients of 0.808, 0.907, 0.833, and 0.929 for ECG, NIBP, ICG, and PPG, respectively. This result demonstrates the feasibility of using a single motion sensor to estimate rich, multimodal cardiac information, offering a simplified and scalable alternative to traditional multi-sensor systems. Finally, we explore the use of synthetic SCG data to improve cross-domain pulmonary hypertension (PH) detection. We leverage generative models and introduce a dataset selection method to optimize the composition of synthetic and real training data. Our approach improves the out-of-distribution PH detection performance, increasing the area under the ROC curve from 0.51 to 0.86. These results highlight the potential of generative SCG modelling in a clinically relevant scenario with limited training data.

Collectively, these contributions advance seismocardiography as a viable and scalable modality for wearable cardiac monitoring. By improving signal robustness, addressing data scarcity, enabling multimodal estimation, and demonstrating clinical relevance, this work expands the role of SCG in both research and real-world applications.