GSW-UNet: An Improved UNet Model Based on PPG Signal for Arterial Blood Pressure Waveform Estimation

Approximately one-third of global deaths are related to cardiovascular diseases (CVD), which are the leading cause of disability-adjusted life years, encompassing years of life lost due to disability and premature death[1]. For individuals with hypertension or severe cardiovascular diseases, continuous blood pressure monitoring is crucial. Current cuff-based blood pressure monitoring methods are non-invasive but discontinuous, whereas continuous blood pressure monitoring methods are mostly invasive and can only be used clinically by healthcare professionals. Therefore, we propose GSW-UNet, which utilizes deep learning to predict continuous blood pressure signals (ABP) non-invasively from photoplethysmography (PPG) signals. The algorithm is trained using the MIMIC-III dataset. Compared to other established BP measurement methods, our proposed approach achieves superior results, with an average absolute error (MAE) of approximately 4.58 mmHg for SBP, 2.83 mmHg for DBP, and 1.69 mmHg for MAP. Additionally, the measurements for DBP and MAP meet the A-grade standard in the British Hypertension Society (BHS) protocol, while DBP and MAP satisfy the standards set by the Association for the Advancement of Medical Instrumentation (AAMI).