Abstract
Cardiovascular diseases (CVDs) remain a leading cause of global mortality, accounting for approximately 32% of all deaths worldwide. The increasing prevalence of CVDs, coupled with an aging population and sedentary lifestyles, necessitates innovative approaches for continuous monitoring of cardiovascular (CV) parameters during daily activities. While mobile healthcare technologies have advanced significantly, particularly in wearable devices, the independent utilization of photoplethysmogram (PPG) and pressure signals has limited the potential for comprehensive cardiovascular monitoring. This thesis investigates the synergistic coupling of PPG and pressure signals through both analytical algorithms and data-driven machine learning approaches to enhance the monitoring of critical CV parameters, specifically blood pressure (BP) and peripheral oxygen saturation (SpO2).This study begins with a comprehensive review of wearable technologies for monitoring BP and SpO2. It examines key physiological principles, traditional measurement techniques, and biosignal modalities (e.g., optical and mechanical methods). Challenges such as motion artifacts, contact pressure variability, calibration, and noise management in wearable cardiovascular monitoring are highlighted. Traditional approaches often overlook the impact of contact pressure, leading to inconsistent measurements and reduced accuracy.
Building on this foundation, a theoretical analysis of the coupling effect between photoplethysmogram (PPG) and tonoarteriogram (TAG) signals is presented, resulting in a fusion model for cuffless, continuous BP estimation. Derived from the Radiative Transfer Equation, the model explicitly leverages mechanical-optical coupling effects. Validation with 27 participants demonstrated mean absolute errors (MAE) of 3.24/3.59 mmHg and root mean square errors (RMSE) of 4.06/4.48 mmHg for diastolic and systolic BP, respectively. Cross-week tracking confirmed the model’s robustness for long-term monitoring under varied conditions.
The research then extends to SpO2 monitoring, which has gained unprecedented importance during the global COVID-19 pandemic. Traditionally reliant on multiwavelength PPG signals and crucial for both critical care and heart failure diagnosis, SpO2 monitoring faces unique challenges in wearable applications. While contact pressure has emerged as a significant factor in PPG sensor performance, its potential contribution to measurement accuracy has been largely overlooked in conventional approaches. This study presents a multiscale data-driven encoder-decoder model for end-to-end SpO2 monitoring, incorporating both PPG and contact pressure signals. The model employs advanced deep learning architectures to process multi-modal inputs and extract complex feature interactions. Clinical trials involving 30 subjects demonstrated superior performance compared to traditional ratio-of-ratios methods, with reduced RMSE (0.75% vs 1.01%) and standard deviation (0.73% vs 0.90%). The improved accuracy and reliability were particularly notable during motion and under varying environmental conditions, addressing key limitations in current wearable SpO2 monitoring systems.
In conclusion, this thesis demonstrates the advantages of biophysics-guided, machine learning-based fusion models for cardiovascular monitoring. By constructively utilizing contact pressure variations, the proposed methods enhance accuracy and reliability across diverse real-world scenarios. These findings have significant implications for clinical applications and personal health monitoring, advancing continuous, reliable solutions for CVD management. Future work will extend the fusion framework to additional parameters, adaptive calibration, and real-time signal quality assessment, promoting progress in mobile health technologies.
| Date of Award | 29 Apr 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Yuanting ZHANG (Supervisor), Xinge YU (Supervisor) & Ni Zhao (External Co-Supervisor) |