Dual-Branch Fusion Neural Network Model: A Preliminary Demonstration of Absolute, Calibration-Free Blood Pressure Estimation from ECG and PPG Signals

This study presents a novel application of a dual-branch neural network to non-invasively estimate Systolic and Diastolic Blood Pressure (SBP and DBP) using fingertip electrocardiogram (ECG) and photoplethysmogram (PPG) signals, without the need for prior calibration. Addressing the gap in available databases, which often lack clear, unfiltered recordings suitable for non-clinical use, this research utilizes a newly preliminary developed single-subject dataset to enhance reliability and interpretability. Utilizing data collected from a single subject over two months, with 121 recordings of 30 s of synchronous ECG and PPG and relative blood pressure, the model incorporates both convolutional and LSTM branches to process and integrate spatial and temporal signal features effectively. The research demonstrates the neural network’s potential through a 10-fold cross-validation approach, achieving, in average, a mean absolute error of 3.82 ± 0.76 for SBP and 3.26 ± 0.76 for DBP across all folds. These results underscore the accuracy of the approach and highlight the need for further validation across a broader dataset to improve generalizability and reliability in non-clinical settings.