/Real-Time-Noninvasive-Continuous-Blood-Pressure-Estimation-Using-Machine-Learning

心血管疾病已成为全球范围内致人死亡的头号病因。为了能有效预防心血管疾病，血压的连续测量尤为重要。目前，连续血压测量分为无创测量和有创测量两种方式。有创测量虽然能达到较高的精度，但是操作复杂且存在感染风险。无创测量主要基于脉搏波。随着机器学习的发展，愈来愈多的人使用脉搏波特征参数法。该方法主要存在两点问题。其一，手动提取特征对波形的要求较高，特征选取受研究者先验知识影响，极有可能提取到非相关特征。其二，血压波形中包含的丰富生理信息未能被充分挖掘。大多数研究的预测目标为收缩压、舒张压等单一血压值，较少的研究关注血压整体波形的预测。 针对上述问题，本文创新性地将原本用于二维图像处理的U-Net模型引入一维血压预测中，提出了一种基于U-Net的PPG-ABP转换模型。该方法无需手动提取特征，仅使用光电血管容积脉搏波（PPG）信号便可预测出连续血压波形。相较于脉搏波特征法，本文方法在信号获取和处理上更为便捷，在结果输出上包含更丰富的血压波形信息。本文平均血压预测结果满足美国医疗仪器促进协会（AAMI）标准。在英国高血压协会（BHS） 标准下，舒张压与平均血压可达到等级B。此外，本模型针对高血压与正常人群的血压分类也能取得较好的效果。Cardiovascular disease has become the significant cause of death. To prevent such disease effectively, continuous measurement of blood pressure is important. Nowadays, there are two ways of blood pressure measurement: noninvasive measurement and invasive measurement. Although invasive measurement can achieve high precision, it is complex to operate and has infection risk. The noninvasive measurement uses pulse waves. With the development of machine learning, many studies make handcrafted features from pulse waves to predict blood pressure. There are two problems with this method. Firstly, feature extraction requires a high standard for waveform, which is not easily achieved in reality. Besides, feature selection is influenced by prior knowledge of researchers. It is very likely to extract non-related features. Secondly, the abundant physiological information of the blood pressure waveform is not extracted fully. Specifically, most of the research aims to predict systolic pressure (SBP) and diastolic pressure (DBP). Indeed, less research focuses on the prediction of the overall waveform of blood pressure. Given the above problems, the thesis introduces the U-Net model, originally used in two-dimensional image processing, into one-dimensional blood pressure prediction. A model based on U-Net was proposed, directly converting photoplethysmogram (PPG) to arterial blood pressure (ABP). The method does not need to extract the features manually. The continuous blood pressure waveform can be predicted only by using the PPG signal. In term of signal acquisition and processing, this method is more convenient. What’s more, it contains more information of blood pressure waveform in the output. The results of the mean arterial pressure (MAP) prediction meet the AAMI standard. DBP and MAP can reach level B under the BHS standards. In addition, the model can also achieve ideal results in the classification of hypertension and normal people.

Issues