LAUSR

Wearable continuous, cuff-less blood pressure (BP) estimation holds great significance for the prevention and treatment of cardiovascular diseases. Deep learning (DL) is widely used in BP estimation due to its powerful nonlinear modeling capabilities; however, the high computational demands and low parameter efficiency of DL exceed the computational limits of wearable devices. Optimizing DL-based BP estimation algorithms for lightweight performance is therefore clinically important. This study proposes a lightweight binary tree neural network (Bi-Tree-Net) based on the concept of distributed inference, using the tree algorithm as the framework and neural networks as the nodes. Bi-Tree-Net integrates the low computational requirements of tree algorithms with the multimodal feature extraction abilities of multiscale and multiresolution neural networks. By analyzing the morphology of low-amplitude waves produced by cardiac pulsation and the gradient of the ejection phase, it achieves continuous BP estimation. Bi-Tree-Net achieves 14.44 M FLOPs and 0.35 M parameters, with an inference time of 22.58 ms on the Jetson Nano. The mean absolute error (MAE) is as low as $3.21~\pm ~2.88$ mmHg, meeting the standards of Association for the Advancement of Medical Instrumentation (AAMI) and the British hypertensive Society’s grade A classification. Additionally, this study explores the effects of BP amplitude range and variability on estimation accuracy, examines the challenges in estimating systolic BP (SBP), and compares the roles of different electrocardiogram (ECG) waveform segments in continuous BP estimation. This research introduces a novel lightweight algorithm for continuous BP estimation, offering promising support for noninvasive, long-term BP monitoring with wearable devices.