DescriptionCuff-less blood pressure (BP) estimation methods are the highly-desired replacement for conventional cuff-based methods, as they enable long-term and continuous monitoring of the BP with limited disturbance or need for manual operation. Cuff-less BP estimation methods can be generally classified into model-driven and data-driven methods. Both methods often utilize photoplethysmogram (PPG) or electrocardiogram (ECG), which can be continuously and non-invasively acquired.
Model-driven methods are based on the pulse wave transition theory, which relates the pulse transit time (PTT) to the blood pressure. An advantage that model-driven methods offer is that very few parameters have to be learned from the training set, thereby, making them efficient and computationally inexpensive. However, these methods require individual calibration, their accuracy decreases over time, and they generally require recordings of both ECG and PPG, which is hard to maintain over long periods. On the other hand, in data-driven methods, the need for subject-specific calibration and the requirement of using two or more physiological recordings are released. However, these methods typically require high computational budget and massive training datasets for learning a much larger quantity of parameters.
In this thesis, we first provide a comprehensive review of seven models that have been utilized in model-driven BP estimation methods, and discuss their advantages and limitations. We then present an overview of existing data-driven methods that have used ECG alone, PPG alone, or both signals for BP estimation.
Motivated by the notable performance of PPG-based data-driven BP estimation methods, we then present a novel transfer learning-based blood pressure estimation algorithm that utilizes visibility graph to form images from PPG recordings. The proposed method provides accurate BP estimation with only one PPG beat, while being computationally efficient requiring training of only one dense layer. Experimental results demonstrate that the proposed method offers comparable or better BP estimation accuracy compared to other data-driven methods with higher computational complexity, making it a suitable candidate for continuous BP monitoring applications.