Mohammadi Hanieh, Tarvirdizadeh Bahram, Alipour Khalil, Ghamari Mohammad
Advanced Service Robots (ASR) Laboratory, Department of Mechatronics Engineering, School of Intelligent Systems Engineering, College of Interdisciplinary Science and Technology, University of Tehran, Tehran, Iran.
Department of Electrical Engineering, California Polytechnic State University, San Luis Obispo, CA, USA.
Sci Rep. 2025 Jul 1;15(1):22229. doi: 10.1038/s41598-025-07087-2.
Blood pressure (BP) serves as a fundamental indicator of cardiovascular health, measuring the force exerted by circulating blood against arterial walls during each heartbeat. This paper introduces an advanced deep learning framework for precise, non-invasive BP estimation via photoplethysmography (PPG) signals, addressing critical limitations in traditional, cuff-based BP measurement methods. Traditional methods, while reliable, are limited by their inability to provide continuous data, posing challenges for proactive health management. In contrast, PPG-based BP estimation facilitates continuous monitoring, crucial for wearable health technologies and real-time applications. Our proposed model leverages a hybrid architecture of convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM) layers, and an attention mechanism, enabling refined spatial and temporal feature extraction to enhance BP estimation accuracy. This approach is validated on an extensive dataset of 2064 patients from the MIMIC-II database, marking a significant increase in sample size over prior studies and thereby strengthening model robustness and generalizability. Through meticulous preprocessing steps, the model achieved an impressive mean absolute error (MAE) of 1.88 for systolic blood pressure (SBP) and 1.34 for diastolic blood pressure (DBP) across 5-fold cross-validation. These findings underscore the potential of integrating PPG and deep learning as a viable, scalable solution for wearable BP monitoring, providing a foundation for further advancement in accessible, non-invasive cardiovascular health monitoring technologies.
血压(BP)是心血管健康的一项基本指标,用于测量每次心跳期间循环血液对动脉壁施加的压力。本文介绍了一种先进的深度学习框架,用于通过光电容积脉搏波描记法(PPG)信号精确、无创地估计血压,解决了传统袖带式血压测量方法的关键局限性。传统方法虽然可靠,但由于无法提供连续数据而受到限制,这给主动健康管理带来了挑战。相比之下,基于PPG的血压估计有助于连续监测,这对可穿戴健康技术和实时应用至关重要。我们提出的模型利用了卷积神经网络(CNN)、双向长短期记忆(BiLSTM)层和注意力机制的混合架构,能够进行精细的时空特征提取,以提高血压估计的准确性。该方法在来自MIMIC-II数据库的2064名患者的广泛数据集上得到了验证,样本量比之前的研究有显著增加,从而增强了模型的稳健性和通用性。通过细致的预处理步骤,该模型在5折交叉验证中,收缩压(SBP)的平均绝对误差(MAE)达到了令人印象深刻的1.88,舒张压(DBP)的平均绝对误差为1.34。这些发现强调了将PPG和深度学习相结合作为可穿戴血压监测的一种可行、可扩展解决方案的潜力,为进一步推进可及的无创心血管健康监测技术奠定了基础。