An Improved U-Net Architecture for Continuous Non-Invasive Arterial Blood Pressure Estimation

Cardiovascular diseases account for a significant portion of all fatalities, causing significant loss of life globally each year. Currently, continuous blood pressure monitoring is the most feasible option, given the limitations of invasive monitoring methods and the challenges in accuracy, complexity, and reliability of non-invasive techniques. We have delved into the advantages of deep learning and have eliminated the dependency on manual feature extraction. Consequently, we have introduced a network that incorporates information from various scales and an attention mechanism to tackle the deficiencies in existing methods, which might fail to consider the diverse scales inherent in PPG signals. This network accurately estimates continuous ABP waveforms from PPG signals, end-to-end, compensating for the phase shift and preserving waveform characteristics. Experiments demonstrate that our approach yields an mean absolute error of 3.20 ± 4.34 mmHg for DBP, 3.43 ± 3.91 mmHg for MAP, and 4.53 ± 6.22 mmHg for SBP estimation, respectively, outperforming existing works and meeting the stringent standards of both BHS and AAMI. These results demonstrate the effectiveness and potential application of this method in non-invasive blood pressure monitoring.