UMS-Net: A Multi-Scale Difference Network for Cuffless Blood Pressure Estimation

Blood pressure monitoring is essential for preventing and managing cardiovascular diseases, which cannot be overstated. The advent of advanced sensor technology and machine learning has positioned wearable devices as potent tools in this domain. Yet, the precision required to track individual blood pressure variations remains challenging for current methods. Addressing this, we introduce UMS-Net, a novel deep-learning framework optimized for cuffless blood pressure monitoring through multi-scale differential correction. By fusing multi-scale feature extraction with multi-channel attention mechanisms, UMS-Net markedly improves the capture of critical information within PPG signals. Deploying deep separable convolutions refines feature extraction, bolstering the network’s capability to process multi-channel signals. This approach allows for a nuanced analysis of signal discrepancies and uncovers their deeper interconnections. Our validation on independent subjects revealed that UMS-Net attained a mean absolute error (MAE) of 5.88 mmHg for systolic (SBP) and 3.69 mmHg for diastolic (DBP) blood pressure on the VitalDB dataset, surpassing current technologies by 5.57% and 4.65%, respectively. Significantly, UMS-Net’s precision adheres to the stringent AAMI and BHS standards, underscoring its promise in non-invasive blood pressure monitoring and affirming its technological edge.