Introduction
According to the World Health Organization statistics, hypertension is the number one attributable cause of death. Pulse wave velocity (PWV) is a well-known measure of long-term blood pressure burden. Hypertension guidelines worldwide recommend to measure PWV regularly in all hypertensive patients to assess the adequacy of blood pressure control. PWV can only be obtained by measurement device in medical facilities. With the advance of technology, wearable devices have been recognized as a promising tool to facilitate the assessment and management of cardiovascular risks.
In this innovation, we for the first time develop an accurate PWV estimation algorithm based on wrist signals collected by the smart watch, the SENSIO. Besides, blood pressure (BP) estimation from the personal model is also developed. Both photoplethysmography (PPG) from wrist and ECG were collected and used to generate the estimation algorithm. PPG contains information regarding the left ventricular ejection and the properties of arterial tree. It has been widely adopted for estimations of PWV and vascular age. Compared to finger PPG, the morphological features of wrist PPG are less recognizable. We developed the novel mechanisms for missing-feature imputation and ambiguous-feature resolution to enhance the feature extraction from wrist PPG signals. The availability of wrist PPG signals for feature extraction increased from ~60% to 99.1%. The weighted pulse decomposition is also designed to decompose the PPG waveform into five component waves and thus the position, width, and amplitude of component waves were analyzed. Combined features were generated from wrist PPG and ECG signals. We employed the machine learning XGBoost algorithm to construct the hierarchical regression models for PWV estimation and deep convolutional neural network (CNN) model for BP estimation. With this model, we successfully estimate PWV from wrist PPG and ECG, with the root mean square error reduced from 230 cm/s (from finger PPG and ECG for PWV estimation in the literature) to 155 cm/s. The average mean absolute error and average root mean square error of the personal SBP model achieves 6.99 mmHg and 7.58 mmHg while the average mean absolute error and average root mean square error of the personal DBP model achieves 6.24 mmHg and 6.65 mmHg.
The breakthroughs achieved in this innovation include the novel feature imputation and ambiguity resolution strategy of PPG and the algorithm for PWV estimation from wrist PPG and ECG and for personal BP estimation from wrist PPG. The single-precision and double-precision algorithms for real-time processing have been realized. We have applied for patent protection. Hence, the applicability of wearable device, especially smart watch, for long-term and pervasive monitoring of personal health could be achieved in the outpatient setting in National Taiwan University Hospital. This technique is also essential for refining hypertension control on an individual basis. This system just won the 2022 Future Tech Award.