Attention-based residual improved U-Net model for continuous blood pressure monitoring by using photoplethysmography signal

Blood pressure (BP) is an important clinical indicator for cardiovascular health assessment, and accurate monitoring of continuous BP is still a challenging task. In this paper, an attention-based residual improved U-Net (ARIU) model is proposed to improve the accuracy of continuous BP monitoring by using the photo-plethysmography (PPG) signal. This model consists of an improved U-Net employed to learn the high dimen-sional features from PPG signal, an attention module embedded in the skip connections to reduce redundancy of learning features, and a residual module replaced common convolution to prevent degradation problems and enhance generalization performance. The raw PPG signals and arterial BP download from the MIMIC-III data-base, the first and second derivatives of PPG signal are utilized as additional inputs to increase the multiform of input information, and a data input way of parallel-based fusion are adopted to improve the effectiveness of information mining. After data preprocessing, the dataset used in this study contains 150,000 samples, belonging to 100 subjects. The reliability of the proposed model is verified by the ablation experiments, and the advancement of the model is demonstrated by the comparison experiments with other state-of-art methods. The mean absolute error (MAE) and standard deviation (STD) of systolic blood pressure (SBP) predicted by the proposed model are 4.75 mmHg and 6.72 mmHg respectively, and that of diastolic blood pressure is 2.81 mmHg and 4.59 mmHg. The results meet the requirements of the Advancement of Medical Instrumentation (AAMI) and reach the Grade A of the British Hypertension Society (BHS) protocol.

Automated summary

In this paper, a attention-based residual improved U-Net (ARIU) model is proposed for accurate continuous blood pressure monitoring using photo-plethysmography (PPG) signal. The model effectively learns high dimensional features, reduces redundancy, and enhances performance through the use of attention module and residual module. Experimental results demonstrate that the model meets medical standards and performs better than other methods.