Cuffless blood pressure estimation with dual acoustic sensors and a hybrid CNN-Bi-LSTM-attention network

Purpose Cardiovascular diseases require continuous, noninvasive blood pressure (BP) monitoring. This study aims to develop a cuffless BP estimation framework using dual acoustic sensors and deep learning. Design/methodology/approach Two acoustic sensors were positioned over the radial artery to acquire dual-channel pulse signals containing inter-sensor propagation delay information (i.e. local/segmental pulse transit time, PTT_local). Signals were preprocessed through downsampling, filtering, artifact removal and windowed segmentation. A hybrid model combining convolutional neural networks, bidirectional long short-term memory and dual attention mechanisms was used to extract spatiotemporal features for BP prediction. Findings Data from ten healthy subjects showed that, under the leave-one-subject-out evaluation, the proposed model achieved mean absolute errors of 3.21 mmHg for diastolic BP and 4.12 mmHg for systolic BP, with correlation coefficients of 0.946 and 0.918. Compared with representative deep-learning baselines, this approach provided improved accuracy and tighter agreement under subject-independent evaluation. Originality/value This work introduces a dual-acoustic-sensor acquisition system that explicitly leverages inter-sensor propagation information (PTT_local), together with a hybrid neural network enhanced by spatiotemporal attention. The framework demonstrates the feasibility of accurate, cuffless BP monitoring and offers potential for wearable health-care applications.