PPG-based continuous arterial blood pressure estimation via multi-scale cross attention fusion

Abstract

Continuous monitoring of arterial blood pressure (ABP) plays a vital role in the early detection and management of cardiovascular diseases. Most existing ABP estimation methods based on photoplethysmography (PPG) signals typically use convolutional neural network (CNN) to extract local time-domain features, while neglecting the vascular elasticity and hemodynamic characteristics in the frequency-domain. To tackle the issue, we propose a multi-scale cross attention fusion network (MCAFNet). It utilizes the time-domain and frequency-domain information of PPG. Specifically, the network leverages ConvNeXt and Transformer to extract local time-domain and global frequency-domain features, respectively. Transform with sparse attention reduces computational complexity and effectively focuses on the most relevant information. Meanwhile, the feature fusion with cross attention effectively integrates complementary information from both time-domain and frequency-domain, improving the representation ability of features. To verify the effectiveness of our approach, we perform evaluations using the Medical Information Mart for Intensive Care (MIMIC) database. For diastolic blood pressure, mean arterial pressure, and systolic blood pressure, the mean absolute error ± standard deviation are of 1.29 ± 2.11, 1.02 ± 1.47, and 2.48 ± 3.89 mmHg, respectively. This performance meets the standards of Association for the Advancement of Medical Devices (AAMI) and British Hypertension Society (BHS), outperforming current state-of-the-art approaches.