Research on a non-invasive blood pressure prediction method based on deep learning and multifeature fusion of PPG signal

To address the limitations of traditional Blood Pressure (BP) measurement methods, particularly the issues of significant inter-individual variability and low prediction accuracy in existing BP prediction models, this paper proposes a deep learning framework that extracts and fuses multi-features from a single photoplethysmography (PPG) signal for noninvasive continuous BP prediction. The proposed framework began by extracting temporal, frequency, and morphological features from the PPG signal in parallel. These features were then fused via a cross-attention mechanism and fed into a Convolutional Recurrent Neural Network (CRNN) to capture temporal dependencies. The performance evaluation yielded a mean error (ME) of 0.46 mmHg for systolic blood pressure (SBP) and 1.26 mmHg for diastolic blood pressure (DBP), which underscores its effectiveness in capturing long-term trends in physiological parameters. This work presents a novel approach for non-invasive continuous BP monitoring, paving the way for its clinical application in long-term care.