Blood Pressure Prediction Using Deep Learning

Abstract

Although traditional cuff-based blood pressure (BP) monitoring is sporadic and laborious, BP is essential for cardiovascular health. We review deep learning approaches for cuff-less blood pressure estimation, such as CNNs, RNNs, Transformer models, and attention processes, and provide two new PPG-to-ABP waveform synthesis methods. The first (ASBP) maps one-dimensional PPG signals into arterial blood pressure waveforms using a VGG-16 encoder-decoder, while the second (SEANet) uses causal dilated convolutions in a calibration-free framework for continuous blood pressure estimation. Using correlation coefficient (CC), mean absolute error (MAE), and mean absolute percentage error (MAPE) measures, both models are trained and assessed on the UCI cuff-less BP dataset. The results have near-normal residual distributions and satisfy AAMI/BHS clinical criteria. An organized comparison of twenty cutting-edge studies demonstrates the variety of datasets, methodological advancements, and clinical usefulness. We present future work for reliable, generalized blood pressure monitoring with wearable PPG sensors and talk about challenges, including dataset heterogeneity and real-time deployment.