Non-Contact Blood Pressure Prediction Using Radar with a Lightweight Temporal Multi-Scale Feature Fusion Network.

Hypertension is a major global health issue, and continuous, convenient blood pressure monitoring is of great significance for early screening and intervention. To address the insufficient exploitation of multi-scale temporal features and the high model complexity in existing radar-based non-contact blood pressure prediction methods, we propose a lightweight temporal multi-scale feature fusion network (LULMNet), for blood pressure prediction and waveform reconstruction. LULMNet adopts a two-stage training strategy. In the first stage, a lightweight one-dimensional U-Net (1D U-Net) is employed for blood pressure waveform reconstruction and intermediate-to-deep temporal feature extraction. In the second stage, systolic and diastolic blood pressure are estimated via multi-scale fusion of intermediate and deep features from the encoder of the 1D U-Net, followed by LSTM-based temporal modeling and regression through a global average pooling (GAP)and a two-layer fully connected prediction head. Experimental results show that the proposed model achieves an error of 3.21 ± 4.94 mmHg for systolic blood pressure (SBP) and 2.25 ± 3.39 mmHg for diastolic blood pressure (DBP), satisfying the Grade A standard of the British Hypertension Society (BHS). In addition, the normalized mean absolute error (NMAE) for waveform reconstruction is as low as 0.044. These results indicate that the proposed method maintains low model complexity while ensuring prediction accuracy, with only 3.0 M parameters and 0.37 G floating-point operations (FLOPs), demonstrating strong potential for non-contact continuous blood pressure monitoring.