Esmaelpoor Jamal, Moradi Mohammad Hassan, Kadkhodamohammadi Abdolrahim
Amirkabir University of Technology, Tehran, Iran.
Amirkabir University of Technology, Tehran, Iran.
Comput Biol Med. 2020 May;120:103719. doi: 10.1016/j.compbiomed.2020.103719. Epub 2020 Apr 9.
Easy access bio-signals are useful to alleviate the shortcomings and difficulties of cuff-based and invasive blood pressure (BP) measuring techniques. This study proposes a multistage model based on deep neural networks to estimate systolic and diastolic blood pressures using the photoplethysmogram (PPG) signal.
The proposed model consists of two key ingredients, using two successive stages. The first stage includes two convolutional neural networks (CNN) to extract morphological features from each PPG segment and then to estimate systolic and diastolic BPs separately. The second stage relies on long short-term memory (LSTM) to capture temporal dependencies. Further, the method incorporates the dynamic relationship between systolic and diastolic BPs to improve accuracy.
The proposed multistage model was evaluated on 200 subjects using the standards of the British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation (AAMI). The results revealed that our model performance met the requirements of the AAMI standard. Also, according to the BHS standard, it achieved grade A in estimating both systolic and diastolic BPs. The mean and standard deviation of error for systolic and diastolic blood pressure estimations were +1.91±5.55mmHg and +0.67±2.84mmHg, respectively.
Our results highlight the benefits of the proposed model in terms of appropriate feature extraction as well as estimation consistency.
易于获取的生物信号有助于缓解基于袖带和侵入性血压测量技术的缺点和困难。本研究提出一种基于深度神经网络的多阶段模型，用于使用光电容积脉搏波描记图（PPG）信号估计收缩压和舒张压。
所提出的模型由两个关键部分组成，分两个连续阶段。第一阶段包括两个卷积神经网络（CNN），用于从每个PPG段提取形态特征，然后分别估计收缩压和舒张压。第二阶段依靠长短期记忆（LSTM）来捕捉时间依赖性。此外，该方法纳入了收缩压和舒张压之间的动态关系以提高准确性。
使用英国高血压学会（BHS）和医疗仪器促进协会（AAMI）的标准，在200名受试者上对所提出的多阶段模型进行了评估。结果表明，我们模型的性能符合AAMI标准的要求。此外，根据BHS标准，其在收缩压和舒张压估计方面均达到A级。收缩压和舒张压估计误差的平均值和标准差分别为+1.91±5.55mmHg和+0.67±2.84mmHg。
我们的结果突出了所提出模型在适当特征提取以及估计一致性方面的优势。
