摘要
目的提出一种基于深度学习的回归模型,在自建数据集上,实现从消费级相机采集的人脸视频到血压值的估测。方法使用自建人脸视频数据集经数据筛选、预处理后按照8∶1∶1的比例随机分为训练集(111例)、验证集(14例)和测试集(14例);此外,分别使用绿、红、蓝三通道数据进行相同实验,比较通道选择对该任务的影响。结果在十折交叉验证条件下,所提出的模型在测试集上的平均绝对误差和标准差达到收缩压(5.98±5.22mmHg),舒张压(4.30±3.39mmHg),相关系数分别为0.70与0.64。绿色通道数据在该任务中表现明显优于其余两个通道。结论本文在一定程度上证明了利用深度学习实现基于人脸视频的远程血压估测的可行性,此外,绿色通道数据可能含有更多与血压相关的信息。
Objective:A regression model based on deep learning is proposed to estimate the blood pressure value from the facial video collected by the consumer camera on the self-built data set.Methods:After data filtering and preprocessing,the self-built face video data set is randomly divided into training set(111 cases),verification set(14 cases)and test set(14 cases)according to the ratio of 8:1:1:1.In addition,green,red and blue channel data are used for the same experiment to compare the impact of channel selection on the task.Results:Under the 10-fold cross-validation condition,the mean absolute error and standard deviation of the proposed model on the test set reached systolic blood pressure(5.98±5.22 mmHg)and diastolic blood pressure(4.30±3.39 mmHg).The Pearson correlation coefficient is 0.70 and 0.64 respectively.The performance of green channel data in this task is significantly better than the other two channels.Conclusion:This paper proves to some extent the feasibility of using deep learning to realize remote blood pressure estimation based on facial video.In addition,the green channel data may contain more information related to blood pressure.
作者
穆继文
唐晓英
MuJiwen;Tang Xiaoying(School of Life Science,Beijing Institute of Technology,Beijing,100081)
出处
《生命科学仪器》
2023年第5期65-70,共6页
Life Science Instruments
关键词
远程血压估测
人脸视频
深度学习
通道选择
回归模型
Remote blood pressure estimation
Facial video
Deep learning
Channel selection
Regression model
作者简介
穆继文,男,硕士,E-mail:943880379@qq.com;通讯作者:唐晓英,女,北京理工大生命学院学教授,博士生导师,中国电子学会生命电子学分会常务理事,E-mail:xiaoying@bit.edu.cn,专业方向:生物医学工程。
