ReBP: Short-Term Blood Pressure Estimation by Reconstructing PPG Signals Based on mmWave Radar

Non-contact blood pressure (BP) monitoring offers significant potential for healthcare applications due to its convenience and comfort. However, existing methods often require prolonged signal acquisition to ensure accuracy, limiting real-time monitoring capabilities. This paper proposes Reconstructed Blood Pressure (ReBP), a millimeter-wave radar-based system for photoplethysmogram (PPG) signal reconstruction and BP estimation, achieving high accuracy with short-duration inputs. ReBP integrates four modules: data acquisition, signal processing, waveform reconstruction, and BP estimation. First, pulse signals from wrist vasculature are captured via mmWave radar. Then, an adaptive filtering algorithm enhances periodic characteristics by selecting segments with maximal autocorrelation. Subsequently, a Cycle-consistent Generative Adversarial Network (CycleGAN), augmented with a PPG-AWARE loss function, reconstructs PPG waveforms enriched with vascular dynamics. Finally, for BP estimation, a multi-feature fusion network extracts temporal, spectral, and global features via ResNet, Gated Recurrent Units (GRU), and SpectrogramLSTM, fused through a cross-modal attention mechanism. Implemented on a TI IWR6843 radar platform, ReBP was validated with 26 participants, achieving mean absolute errors (MAEs) of 3.891 mmHg (SBP) and 3.646 mmHg (DBP) for 35-second inputs, 4.003 mmHg (SBP) and 3.873 mmHg (DBP) for 15-second inputs, and 4.383 mmHg (SBP) and 4.115 mmHg (DBP) for 5-second inputs, outperforming conventional methods in short-sequence scenarios.