Abstract

Abstract—Continuous blood pressure monitoring is of great
significance for the early diagnosis of cardiovascular diseases. To
address the limitations of current machine learning and deep
learning-based blood pressure (BP) prediction methods, which
rely on manual feature extraction and struggle to reconstruct
complete BP waveforms, this paper employs a continuous noninvasive
arterial BP detection model named Mamba-UNet based
on photoplethysmography (PPG) signals. The model deeply
integrates the selective state space model (Selective SSM) with the
U-Net architecture, achieving direct mapping from PPG signals
to arterial blood pressure (ABP) waveforms through end-to-end
modeling. In the encoder, the MambaConvBlock module captures
long-term temporal dependencies and individual vascular characteristics
of PPG signals by dynamically adjusting parameters
(Δ, B, C). The decoder employs a hybrid Mamba-convolution
structure, combining the global dynamic modeling capability of
SSM with the local feature extraction ability of convolution to
accurately reconstruct BP waveform details. The model design
balances the multi-scale feature integration advantages of U-Net
with the efficient long-sequence processing capability of Mamba.
Evaluated on the Sensors dataset (derived from MIMIC-III,
containing 1,131 ICU patient records), Mamba-UNet achieved
a mean absolute error (MAE) of 6.06 mmHg for diastolic blood
pressure (DBP) and 13.11 mmHg for systolic blood pressure
(SBP), outperforming models such as MLP, ResNet, and U-Net.
Index Terms—Arterial blood pressure (ABP), photoplethysmography
(PPG), Mamba, U-Net, non-invasive.