电子测量与仪器学报

朱凌建,陈剑虹,王裕鑫,郑铱,王森,荀子涵.基于 GRU 神经网络的脉搏波波形预测方法研究[J].电子测量与仪器学报,2022,36(5):242-248

基于 GRU 神经网络的脉搏波波形预测方法研究

Research on prediction method of pulse wave waveformbased on GRU neural network

DOI：

英文关键词:GRU photoplethysmography pulse wave prediction physiological parameter monitoring data support

基金项目:陕西省重点研发计划（2020ZDLGY10 04)项目资助

作者	单位
朱凌建	1.西安理工大学机械与精密仪器工程学院
陈剑虹	1.西安理工大学机械与精密仪器工程学院
王裕鑫	1.西安理工大学机械与精密仪器工程学院
郑铱	1.西安理工大学机械与精密仪器工程学院
王森	1.西安理工大学机械与精密仪器工程学院
荀子涵	1.西安理工大学机械与精密仪器工程学院

Author	Institution
Zhu Lingjian	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology
Chen Jianhong	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology
Wang Yuxin	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology
Zheng Yi	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology
Wang Sen	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology
Xun Zihan	1.School of Mechanical and Precision Instrument Engineering, Xi′an University of Technology

摘要点击次数: 575

全文下载次数: 684

中文摘要:

随着生活水平的提高,人们对健康的关注度越来越高,尤其是适应快节奏生活的手环等便携式生理监测设备,备受人们青睐。光电容积脉搏波描记法(PPG)作为一种无创人体脉搏采集手段,被广泛应用于此类设备中。人体脉搏中包含很多生理信息,如血压、血糖、动脉硬化等,为了对这些信息进行提取和分析,目前主要采用机器学习的方法,通过提取脉搏波中的特征点计算特征参数进而建立生理参数模型。但此类方法需要大量且长期的脉搏数据,用于提高生理参数模型的精度,而长期的数据采集受环境限制较大且与便携式生理监测设备设计理念冲突,并且对脉搏波预测的研究存在空白。针对此问题,本文使用 Colaboratory 建立 GRU 神经网络模型与 LSTM 网络模型分别对脉搏波数据进行预测,并对影响模型性能的主要参数进行对比调参。而由自动化机器学习工具 AutoML_Alex 针对脉搏波数据分析并择优建立的 LightGBM 网络可以作为具有参考价值的基线模型。通过以上 3 个模型针对从不同个体采集到的大量脉搏波数据进行建模,对比其平均绝对百分比误差 MAPE,LSTM 为 0. 879%,单层 GRU 为 0. 852%,LightGBM 为 0. 842%,4 层 GRU 模型为 0. 828%,进而应用到不同个体上发现单层 GRU 模型的稳定性(MAPE)要优于其他模型。本文以 GRU 网络建立的脉搏波预测模型以不同个体短期脉搏数据为蓝本,对不同个体的长期脉搏波数据进行预测,进而对人体的动脉硬化生理情况进行监测达到早期发现早期预防的目的,同时为便携式生理监测设备提供技术和数据支持。

英文摘要:

With the improvement of living standards, people are paying more and more attention to health. In particular, the portable physiological monitoring device such as smart wristband that adapts to fast-paced life is favored by people. Photoplethysmography (PPG), as a non-invasive human pulse collection method, is widely used in such device. The human pulse contains a lot of physiological information. In order to extract and analyze this information, the method of machine learning is generally used to establish a mathematical model. However, such methods require a large amount of long-term pulse data to improve the accuracy of physiological parameter models. In response to the problem, this article uses Colaboratory to establish a GRU neural network model and together with LSTM to predict the pulse wave data and adjust the main parameters that affect the performance of the model. The automated machine learning tool AutoML_Alex analyzes the pulse wave data and establishes the LightGBM network based on the best ones, which can be used as a baseline model with reference value. Use large amounts of pulse wave data collected from different individuals to build three different models, compare with MAPE, LSTM is 0. 879%, single-layer GRU is 0. 852%, LightGBM is 0. 842%, and four-layer GRU model is 0. 828%, and apply different models to different individual predictions. It is found that the stability of the single layer in the GRU model is better in the application of different individuals. The results show that we can establish a GRU network model based on short-term pulse wave data of different individuals, predict long-term pulse wave data, and then monitor the human body's arteriosclerosis and other physiological conditions, while providing technical and data support for portable physiological monitoring equipment.

查看全文查看/发表评论下载PDF阅读器