R2A-MAGNet: A Mamba-Based Attentive Gated Recurrent Network for Reconstructing Central Arterial Pressure from Radial Pressure Waveforms

Abstract Cardiovascular diseases require precise blood pressure monitoring, with central aortic pressure (CAP) being a stronger predictor of risk than peripheral arterial pressure (PAP). However, invasive methods for measuring CAP are accurate, they are costly and carry procedural risks, while non-invasive alternatives often lack accuracy and struggle to adapt to individual physiological differences. To address these limitations, this paper introduces R2A-MAGNet, a novel deep learning model that non-invasively reconstructs CAP waveforms from radial arterial pressure (RAP) waveforms. R2A-MAGNet combines CNNs for local feature extraction with GRUs, the Selective State Spaces Model (Mamba), Self-Attention, and Cross-Attention for global feature learning and enhanced information interaction. Tested on a real-world dataset, R2A-MAGNet outperforms existing models, achieving the lowest Mean Absolute Error for CAP (1.93 mmHg), RMSE for central systolic (2.77 mmHg), and RMSE for central diastolic pressure (1.48 mmHg). Visual and statistical analyses confirm its accuracy and reliability, underscoring its potential for advanced blood pressure monitoring.