Abstract
Goal: This study presents an enhanced stacked U-Net deep learning model for cuffless blood pressure estimation using only photoplethysmogram signals, aiming to improve the accuracy of non-invasive measurements. Methods: To address the challenges of systolic blood pressure estimation, the model incorporates velocity plethysmogram input and employs additive spatial and channel attention mechanisms. These enhancements improve feature extraction and mitigate decoder mismatches in the U-Net architecture. Results: The model satisfies the Grade A criteria established by the British Hypertension Society and meets the accuracy standards of the Association for the Advancement of Medical Instrumentation, achieving mean absolute errors of 3.921 mmHg for systolic and 2.441 mmHg for diastolic blood pressure. It outperforms PPG-only spectro-temporal methods and achieves comparable performance to the joint photoplethysmogram and electrocardiogram one-dimensional squeeze-and-excitation network with long short-term memory architecture. Conclusions: The proposed model shows strong potential as a practical, low-cost, and non-invasive solution for continuous, cuffless blood pressure monitoring.