CASIA OpenIR  > 毕业生  > 硕士学位论文
基于光电容积脉搏波的无创连续血压检测方法研究
赵锦
2024-05-14
Pages60
Subtype硕士
Abstract

心血管疾病作为当今世界主要死亡因素之一,对全球范围内的人类健康构成严峻威胁。人体血压水平与心血管疾病风险息息相关,高血压是心血管疾病发病最主要的诱导因素。血压是人体心血管系统中的一个重要生理参数,直接关系到血液在动脉系统中的流动状态和心脏的泵血功能。在医疗健康领域,血压测量是评估心血管健康的重要手段之一,实施连续血压波动监测和评估可以辅助预防心血管疾病的紧急发作。目前情况来看,传统血压测量方法无法在不影响用户正常生活状态下满足获取血压实时变动的实际需求,同时血压测量仪器的便携性和操作简易性需要进一步提升。现代医学、光电子学、深度神经网络和智能设备的快速发展为光电容积脉搏波(Photoplethysmography,PPG)技术铺平了道路,从而具备在日常生活场景中开展血压连续监测的广大前景。在先前的众多研究中,已经发现了PPG信号与血压数值之间变化的同一性。然而,PPG预测血压领域仍然存在通用模型血压特征提取能力不足、PPG信号个体差异性大以及模型域泛化困难的问题。因此,本文开展基于光电容积脉搏波的无创连续血压检测方法研究,旨在提升方法的实用价值,缩短科学研究与实际应用之间的距离。本文研究内容如下:

(1)本文设计了公开数据集一体化处理流程,提出了基于快速傅里叶变换和二维卷积网络相结合的的PPG血压特征提取器。本文以公开医疗数据集重症监护医疗信息系统(Medical Information Mart for Intensive Care III,MIMIC-III)为基础,提出了信号优化、信号分窗、标签绑定、信号评价、数据增强的标准化数据处理流程,其中信号优化采用巴特沃斯滤波、小波变换去基线漂移,而信号评价体系是一种自相关系数法和基于核密度估计的离群点检测法相结合的信号质量打分系统。依据PPG信号和动态血压波形短期内都具备一维时序周期特性,本文提出按照信号核心频率进行周期切分的维度变换策略,从而构建了一阶段1D-2D深度模型特征提取器。该模型能够同时获取PPG信号的周期内上下文信息和相邻周期同位置上下文信息,从周期相对位置角度提取血压特征。经过对比实验,本文提出的基于维度变换策略的二维特征提取方法,能够达到优于当前研究的血压预测效果。

(2)本文基于一阶段1D-2D血压特征提取器,设计完成二阶段血压校准模块,从而实现了能够跨受试者精准预测血压的2DBP\_Net完整架构。当前多数PPG信号预测血压的研究工作仅仅在一阶段就得出文章结论,并未深入探讨PPG信号的个体差异性这一核心问题。PPG信号个体差异性是无创血压测量技术走向成熟稳定的主要阻碍,目前仅依靠更新网络架构的思路无法避免过拟合现象。因此,本文提出一阶段1D-2D血压特征提取框架和二阶段ResNet-18网络相结合的比较特征方法。校准方法的核心在于抓住特征变化与血压变化的同一性,因为同一受试者的不同PPG信号都包含了相同的个体差异性,而不同的特征变动代表着不同的血压变动。通过实验结果证明,无需借助自身大量标签样本迁移学习,2DBP\_Net仅依靠一个自身标签样本就能够具备长期血压精准预测能力。

(3)本文针对国内居民设计了数据采集实验,并依靠迁移学习微调方法缓解模型域泛化问题。本文按照国民年龄分布征集受试者,使用Nova10 Pro手机摄像头收集PPG信号,使用OMRON HEM-7136血压仪收集血压标签,参考公开数据集的预处理方法获得自建数据集。本文使用迁移学习训练集先微调特征提取器第一线性层,完成后根据特征提取器生成的特征构建比较特征,然后微调校准模块所有卷积层和线性层。通过迁移学习验证集验证,2DBP\_Net迁移学习微调方法在自建数据集表现良好,完全满足国际医疗器械标准要求。实验结果表明,本研究所提出的双阶段血压预测方法能够克服不同采集设备、不同人群带来的困难,从而侧面反映出本文方法的实际应用价值。

Other Abstract

Cardiovascular diseases (CVDs) represent one of the leading causes of mortality worldwide, posing a significant threat to human health on a global scale. The level of human blood pressure is closely associated with the risk of cardiovascular diseases, with hypertension being the primary predisposing factor for the onset of CVDs. Blood pressure serves as a critical physiological parameter in the human cardiovascular system, directly impacting the flow state of blood in the arterial system and the pumping function of the heart. In the healthcare domain, blood pressure measurement stands as one of the key methods for evaluating cardiovascular health, and implementing continuous monitoring and assessment of blood pressure fluctuations can assist in preventing acute cardiovascular events. However, traditional methods of blood pressure measurement fail to meet the practical requirements for real-time blood pressure monitoring without affecting the normal lifestyle of users, exhibiting shortcomings in terms of instrument portability and simplicity.

The rapid development of modern medicine, optoelectronics, deep neural networks, and intelligent devices has paved the way for photoplethysmography (PPG) technology, opening up its widespread application prospects in various daily life scenarios for continuous monitoring. Previous studies have identified the consistency between PPG signals and blood pressure values. However, challenges persist in the field of blood pressure prediction using PPG, including insufficient capability of universal models in extracting blood pressure features, significant individual differences in PPG signals, and difficulties in model domain generalization.

Therefore, this study embarks on the research of non-invasive continuous blood pressure detection based on photoplethysmography, aiming to enhance the practical value of the method and bridge the gap between scientific research and practical application. The research objectives are outlined as follows:

(1) This study designs an integrated processing pipeline for public datasets and proposes a PPG blood pressure feature extractor based on the combination of fast Fourier transform and two-dimensional convolutional networks. Leveraging the publicly available critical care database Medical Information Mart for Intensive Care III (MIMIC-III), a standardized data processing pipeline is proposed, encompassing signal optimization, signal segmentation, label binding, signal evaluation, and data augmentation. The signal optimization employs Butterworth filtering and wavelet transform to remove baseline drift, and the signal evaluation system consists of a signal quality scoring system combining autocorrelation coefficient method and outlier detection based on kernel density estimation. Given that both PPG signals and dynamic blood pressure waveforms exhibit one-dimensional time-series periodic characteristics within a short term, this study proposes a dimension transformation strategy based on signal core frequency for periodic segmentation, thereby constructing a one-stage 1D-2D deep model feature extractor. This model can simultaneously capture the contextual information within PPG signal cycles and neighboring cycle positions, extracting blood pressure features from the perspective of cycle-relative positions. Through comparative experiments, the proposed two-dimensional feature extraction method based on dimension transformation strategy demonstrates superior blood pressure prediction performance compared to current research.

(2) Building upon the one-stage 1D-2D blood pressure feature extractor, this study designs a two-stage blood pressure calibration module to realize the complete architecture of 2DBP\_Net capable of accurately predicting blood pressure across subjects. Most existing studies on PPG signal-based blood pressure prediction only draw conclusions in the first stage without delving into the core issue of individual differences in PPG signals. Individual differences in PPG signals pose a major obstacle to the maturity and stability of non-invasive blood pressure measurement techniques, as overfitting phenomena cannot be avoided solely by updating network architectures. Therefore, this study proposes a comparison feature method combining the first-stage 1D-2D blood pressure feature extraction framework with the second-stage ResNet-18 network. The core of the calibration method lies in capturing the consistency between feature changes and blood pressure changes, as different PPG signals from the same subject all contain the same individual differences, while different feature changes represent different blood pressure changes. Experimental results demonstrate that 2DBP\_Net, without the need for extensive labeled samples or transfer learning, exhibits long-term accurate blood pressure prediction capability solely relying on its own labeled sample.

(3) This study designs data collection experiments for domestic residents and mitigates the problem of model domain generalization through transfer learning fine-tuning methods. Recruiting participants according to the national age distribution, PPG signals are collected using Nova10 Pro smartphone cameras, and blood pressure labels are collected using OMRON HEM-7136 sphygmomanometers. Referring to preprocessing methods of publicly available datasets, a self-built dataset is obtained. Transfer learning fine-tunes the first linear layer of the feature extractor based on the transfer learning training set, constructs comparison features based on the features generated by the feature extractor, and then fine-tunes all convolutional layers and linear layers of the calibration module. Validated through transfer learning validation set, the transfer learning fine-tuning method of 2DBP\_Net performs well on the self-built dataset, fully meeting the requirements of international medical device standards. Experimental results demonstrate that the proposed two-stage blood pressure prediction method can overcome difficulties arising from different collection devices and different populations, thereby reflecting the practical application value of the proposed method.

Keyword血压 光电容积脉搏波 快速傅里叶变换 深度学习 校准策略 迁移学习
Subject Area人工智能
MOST Discipline Catalogue工学
Language中文
Sub direction classification生物特征识别
planning direction of the national heavy laboratoryAI For Science
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56492
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
赵锦. 基于光电容积脉搏波的无创连续血压检测方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
基于光电容积脉搏波的无创连续血压检测方法(7584KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[赵锦]'s Articles
Baidu academic
Similar articles in Baidu academic
[赵锦]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[赵锦]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.