CN114569096B - Non-contact continuous blood pressure measurement method and system based on video stream - Google Patents
Non-contact continuous blood pressure measurement method and system based on video stream Download PDFInfo
- Publication number
- CN114569096B CN114569096B CN202210193095.8A CN202210193095A CN114569096B CN 114569096 B CN114569096 B CN 114569096B CN 202210193095 A CN202210193095 A CN 202210193095A CN 114569096 B CN114569096 B CN 114569096B
- Authority
- CN
- China
- Prior art keywords
- blood pressure
- pulse wave
- feature
- video
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 73
- 238000009530 blood pressure measurement Methods 0.000 title claims abstract description 34
- 230000036772 blood pressure Effects 0.000 claims abstract description 180
- 238000012545 processing Methods 0.000 claims abstract description 39
- 238000012216 screening Methods 0.000 claims abstract description 15
- 230000035488 systolic blood pressure Effects 0.000 claims abstract description 14
- 230000035487 diastolic blood pressure Effects 0.000 claims abstract description 12
- 230000006870 function Effects 0.000 claims description 29
- 238000000605 extraction Methods 0.000 claims description 28
- 238000012549 training Methods 0.000 claims description 24
- 238000010187 selection method Methods 0.000 claims description 20
- 238000013528 artificial neural network Methods 0.000 claims description 19
- 238000010276 construction Methods 0.000 claims description 17
- 238000000354 decomposition reaction Methods 0.000 claims description 17
- 238000001914 filtration Methods 0.000 claims description 17
- 238000005516 engineering process Methods 0.000 claims description 16
- 239000002245 particle Substances 0.000 claims description 16
- 238000012163 sequencing technique Methods 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 16
- 239000000284 extract Substances 0.000 claims description 14
- 238000004458 analytical method Methods 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 13
- 230000002776 aggregation Effects 0.000 claims description 11
- 238000004220 aggregation Methods 0.000 claims description 11
- 238000012880 independent component analysis Methods 0.000 claims description 11
- 230000009466 transformation Effects 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 8
- 230000001815 facial effect Effects 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 230000014759 maintenance of location Effects 0.000 claims description 6
- 230000005540 biological transmission Effects 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000012706 support-vector machine Methods 0.000 claims description 4
- 239000008186 active pharmaceutical agent Substances 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 230000008030 elimination Effects 0.000 claims description 3
- 238000003379 elimination reaction Methods 0.000 claims description 3
- 230000035772 mutation Effects 0.000 claims description 3
- 238000010183 spectrum analysis Methods 0.000 claims description 3
- 230000003068 static effect Effects 0.000 claims description 3
- 238000000658 coextraction Methods 0.000 claims description 2
- 230000004069 differentiation Effects 0.000 claims description 2
- 230000000630 rising effect Effects 0.000 claims description 2
- 238000012544 monitoring process Methods 0.000 abstract description 7
- 238000005259 measurement Methods 0.000 abstract description 3
- 238000011160 research Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 230000002159 abnormal effect Effects 0.000 description 3
- 239000008280 blood Substances 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000000926 separation method Methods 0.000 description 2
- 238000010989 Bland-Altman Methods 0.000 description 1
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 1
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 210000001367 artery Anatomy 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 230000002526 effect on cardiovascular system Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 210000000744 eyelid Anatomy 0.000 description 1
- 238000009532 heart rate measurement Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000004393 prognosis Methods 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/021—Measuring pressure in heart or blood vessels
- A61B5/02108—Measuring pressure in heart or blood vessels from analysis of pulse wave characteristics
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/0059—Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence
- A61B5/0075—Measuring for diagnostic purposes; Identification of persons using light, e.g. diagnosis by transillumination, diascopy, fluorescence by spectroscopy, i.e. measuring spectra, e.g. Raman spectroscopy, infrared absorption spectroscopy
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/021—Measuring pressure in heart or blood vessels
- A61B5/02108—Measuring pressure in heart or blood vessels from analysis of pulse wave characteristics
- A61B5/02116—Measuring pressure in heart or blood vessels from analysis of pulse wave characteristics of pulse wave amplitude
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
- A61B5/024—Detecting, measuring or recording pulse rate or heart rate
- A61B5/02416—Detecting, measuring or recording pulse rate or heart rate using photoplethysmograph signals, e.g. generated by infrared radiation
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/725—Details of waveform analysis using specific filters therefor, e.g. Kalman or adaptive filters
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7253—Details of waveform analysis characterised by using transforms
- A61B5/7257—Details of waveform analysis characterised by using transforms using Fourier transforms
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Veterinary Medicine (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Physiology (AREA)
- Artificial Intelligence (AREA)
- Cardiology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Signal Processing (AREA)
- Mathematical Physics (AREA)
- Vascular Medicine (AREA)
- Evolutionary Computation (AREA)
- Fuzzy Systems (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
The invention discloses a non-contact continuous blood pressure measurement method and system based on video stream. Belonging to the field of blood pressure measurement, the method comprises the following steps: a pulse wave signal processing and feature extracting module; a single person model building and screening module; a PPG signal feature selection module; a SVR general blood pressure model building module; the video signal acquisition and processing module; and a video blood pressure prediction module. The video blood pressure prediction result shows that the prediction results of the systolic pressure and the diastolic pressure meet the A-level standard of AAMI and BHS. The invention can achieve the purpose of accurately measuring continuous blood pressure by only collecting pulse wave signals in video, has simple measurement mode, does not need to be contacted with human body, and has wide application value and prospect for clinical and daily health monitoring.
Description
Technical Field
The invention belongs to the field of blood pressure measurement, and particularly relates to a non-contact continuous blood pressure measurement method and system based on video stream.
Background
Blood Pressure (BP) directly reflects the functional state of human cardiovascular and cerebrovascular vessels and is an important basis for disease diagnosis, efficacy evaluation and prognosis analysis. Compared with the traditional blood pressure detection method, the continuous noninvasive non-contact blood pressure detection method has the advantages of portability, comfort, simplicity and the like.
In recent years, some researches estimate blood pressure according to pulse wave characteristics, for example, literature 1(Wang,Y.,et al.Continuous blood pressure prediction using pulse features and Elman neural networks.in 2017 IEEE 17th International Conference on Communication Technology(ICCT).2017.) utilizes pulse wave single-channel signals to extract few pulse wave waveform characteristics and construct a BP prediction model. However, the prediction accuracy of the method needs to be improved, and no related extraction method of pulse wave signals exists.
Regarding the acquisition of pulse waves, the principle of acquiring the pulse waves by utilizing the photoelectric volume is mature. Research on acquiring pulse waves based on video is continuously advancing. The pulse wave signal is successfully extracted from the face video signal using the principle of Independent Component Analysis (ICA) as in document 2 (Xu, f., et al, HEART RATE measurement based on face video sequence 2015.941435).
With the development of pulse wave extraction technology and prediction algorithm, research on a continuous non-invasive non-contact blood pressure measurement method based on video is considered. Blood pressure is predicted using an artificial intelligence model, as in literature 3(Y.Fang,P.Huang,M.Chung and B.Wu,″A Feature Selection Method for Vision-Based Blood Pressure Measurement,″2018IEEE International Conference on Systems,Man,and Cybernetics(SMC),Miyazaki,Japan,2018,pp.2158-2163,doi:10.1109/SMC.2018.00371.), by video recording of facial and palm signals. The signals required to be acquired by the method are more, only the differences among the models are compared, and the prediction accuracy is not high.
The patent CN108272449 discloses a non-contact blood pressure monitoring method and system. The video signal streams of the first sampling position and the second sampling position on the body of the monitored person are obtained, the transmission speed of pulse waves is extracted, and the transmission speed is converted into a blood pressure value, so that the dynamic monitoring of blood pressure is realized. The system and the method can realize non-contact dynamic continuous blood pressure monitoring, improve the comfort experience of the user and feed back the blood pressure condition of the user in real time. The method needs to collect two video signals of the face and the fingertip at the same time, and the complexity of signal collection is greatly increased.
The patent number CN110706826 discloses a non-contact real-time multi-person heart rate and blood pressure measuring method based on video images. Acquiring a video image of a face tissue area of a detected person in a non-contact mode, and extracting the face image according to the video image; performing image processing on skin tissues meeting preset conditions in all the face areas and generating image data; and calculating heart rate and blood pressure values of the detected crowd according to the image data. The method only utilizes pulse wave conduction time to establish the relationship with blood pressure, the calculation error is relatively large in the process of carrying out real-time measurement of blood pressure, and the accuracy of a blood pressure prediction result is not described.
The patent of China patent No. CN106821356 discloses a cloud continuous blood pressure measurement method and system based on an Elman neural network. And taking the characteristic points of the extracted pulse wave signals as input of an Elman neural network, predicting the blood pressure value by adopting a trained Elman neural network model, and taking the obtained predicted value as a continuous blood pressure measured value. The method utilizes a neural network method to predict blood pressure, but a non-contact pulse wave signal acquisition and processing method is not described.
The patent number CN110090010 discloses a non-contact blood pressure measuring method and a non-contact blood pressure measuring system. According to the method, after the facial video of a user is obtained, image frames of a first region of interest and a second region of interest are screened out, a heart rate estimated value is obtained by blind source separation, a photoelectric volume pulse wave of the region of interest is extracted and determined according to the heart rate estimated value, and a blood pressure value is obtained by using a relation equation of time lag input time lag and blood pressure of a PPG signal, so that direct contact and compression on a human body can be avoided. The method cannot realize real-time blood pressure monitoring.
Chinese patent No. CN105011921 discloses a method for measuring blood pressure by video analysis. According to the method, an arterial video of one side of a superficial artery of a human body is shot, the blood pressure of the human body is measured before and after the shot video by a conventional method, video signals are analyzed according to the measured high-pressure value and low-pressure value, the correlation coefficient of the blood pressure and the arterial scale is calculated, and the steps are repeated during actual measurement to obtain a specifically measured high-pressure value and low-pressure value. The blood pressure measuring method can be obtained by shooting video and analyzing the video, and can finish blood pressure measurement without touching a human body. The method also needs to measure the high-voltage value and the low-voltage value of the human body by using a conventional method before and after shooting the video, and is complex in operation.
The patent of China patent No. CN108523867 discloses a self-calibration PPG noninvasive blood pressure measurement method. According to the method, a PPG blood pressure measured value is measured according to a linear relation between pulse wave conduction time and blood pressure, and the PPG blood pressure measured value is calibrated according to a blood pressure calibration model and personal characteristic information. According to the method, the PPG blood pressure measured value is calibrated according to the personal characteristic information to improve the measuring efficiency and result, so that the blood pressure measured result is less disturbed. The method is used for contact blood pressure measurement, and a bracelet is required to be worn to collect PPG signals.
Disclosure of Invention
In order to solve the problems of the blood pressure measurement method and meet the requirements of people on blood pressure measurement under the rapid development of intelligent medical treatment, the invention provides a non-contact continuous blood pressure measurement method and system based on video streams, which realize real-time monitoring of blood pressure information and have great application prospects in daily life and clinical monitoring. The prediction from the facial video to the blood pressure is carried out, and the AAMI and BHS standards are verified to be met, so that the accuracy of acquiring the blood pressure based on the video signal is improved, and the blood pressure condition of the user is fed back in real time.
The invention is realized by adopting the following technical scheme:
a video stream based non-contact continuous blood pressure measurement system comprising:
The pulse wave signal processing and feature extracting module is used for preprocessing a pulse wave signal by adopting a discrete wavelet decomposition technology and a wavelet denoising method, analyzing the waveform features of the PPG and the expression of the physiological information of a reference human body in the PPG waveform, and extracting the time-frequency domain features of a pulse wave sequence;
The single-person model construction and screening module is used for establishing three single-person blood pressure prediction models according to the time-frequency domain characteristics of the pulse wave sequence extracted by the pulse wave signal processing and characteristic extraction module, wherein the three single-person blood pressure prediction models are respectively single-person blood pressure prediction models of an Elman neural network based on particle swarm optimization, and are based on the single-person blood pressure prediction model of a deep belief network and the single-person blood pressure prediction model of SVR, and the three single-person blood pressure prediction models are evaluated and screened to obtain an optimal model which is the single-person blood pressure prediction model of SVR;
The PPG signal feature selection module is used for carrying out feature sequencing according to the time-frequency domain features of the pulse wave sequence extracted by the pulse wave signal processing and feature extraction module to obtain a feature weight sequencing result;
The SVR general blood pressure model construction module is used for selecting time-frequency domain features of pulse wave sequences with different numbers according to the feature weight sequencing result to reconstruct an SVR single blood pressure prediction model to obtain an SVR general blood pressure prediction model, and carrying out consistency analysis on the prediction result and actual blood pressure to obtain and verify an optimal feature subset and stability of the optimal feature subset;
The video signal acquisition and processing module firstly acquires the face video data to be tested and the fingertip PPG signal by using a camera, extracts RGB sequences from the region of interest by using a frame selection technology for the acquired face video data, filters the RGB sequences by using a color distortion filtering algorithm, extracts video pulse wave signals from the filtered RGB sequences by using an independent component analysis algorithm, and carries out consistency estimation with the fingertip PPG signal;
the video blood pressure prediction module is used for extracting time-frequency domain features of a video pulse wave sequence by adopting a method in the pulse wave signal processing and feature extraction module, extracting a video pulse wave feature subset according to a pulse wave optimal feature subset, carrying out consistency estimation with the optimal pulse wave feature subset of a fingertip PPG signal, obtaining the optimal video pulse wave feature subset, taking the optimal video pulse wave feature subset as input, and carrying out blood pressure prediction by utilizing a SVR general blood pressure prediction model obtained through training.
The invention further improves that the pulse wave signal processing and characteristic extracting module comprises:
Extracting time domain features, namely adopting a findpeaks function of Matlab, setting the minimum amplitude and the minimum distance between adjacent points to realize the positioning of pulse wave peaks, further searching the minimum value between the adjacent wave peaks to realize the positioning of the pulse wave peaks, adopting a trapz function of Matlab to realize the calculation of a rising area SS and a falling area DS, and adopting a diff function to realize the calculation of first-order differentiation of pulse wave signals to obtain the time domain features of the pulse waves;
Carrying out frequency domain feature extraction, namely carrying out Fourier transformation on each of left and right 10 pulse wave signals taking a current pulse wave as a center through a fft function of Matlab, adopting a max function of Matlab to realize the determination of the frequency and amplitude between 0.3-1.6Hz and 1.6-3Hz in fundamental waves and second harmonics of the pulse wave signals, adopting 'db6' wavelet as a mother wavelet to carry out 6-layer decomposition extraction on the pulse wave signals to obtain wavelet domain features, adopting an empirical mode decomposition tool box of Maltab to realize the decomposition of the pulse wave signals, and then carrying out Hilbert transformation on IMF components obtained by decomposition to obtain Hilbert transform domain features;
The co-extraction includes 78 time-frequency domain features, including time features, amplitude features, and area features.
The invention further improves in that the single model constructing and screening module comprises:
The method comprises the steps of establishing a single blood pressure prediction model of an Elman neural network with particle swarm optimization, adopting a single hidden layer, adopting a particle swarm optimization algorithm to find an optimal network weight and a threshold value, adopting mutation operation, and re-initializing particles with random probability after each particle update;
The SVR single blood pressure prediction model is built, a Libsvm library is adopted for training, an epsilon-SVR model of a radial basis function kernel is adopted, and the super parameters in the model are determined through a grid parameter optimizing function SVMcgForRegress for regression problems, namely: penalty coefficient C, tolerance E of kernel parameter g and termination criterion in kernel function;
The method comprises the steps of establishing a single blood pressure prediction model based on a deep belief network, pre-training by adopting a limited Boltzmann machine, keeping as much characteristic information as possible when characteristic parameters are reduced in dimension, initializing a neural network weight by using trained limited Boltzmann machine network parameters, and adjusting the weight by using training data in combination with an error back transmission algorithm;
When three single blood pressure prediction models are trained, 60% of data are used for model training by adopting a three-fold cross validation method, and 40% of data are used as test data; according to AAMI standard of blood pressure measurement, average absolute error of 3 times of obtained predicted blood pressure and actual blood pressure and average of standard deviation are used as evaluation indexes of model accuracy, and three modeling methods are compared and evaluated to obtain an optimal SVR single blood pressure prediction model.
The invention further improves that the PPG signal characteristic selection module also comprises a characteristic selection method designed in the specification, and the method specifically comprises the following steps:
using 65 tested data to participate in feature selection, wherein each tested data is about 3 data in different time periods, 190 groups of data are used, and each data length is ten minutes;
Selecting RRelieff, a filtering type feature selection method for neighbor component analysis and two embedded feature selection methods for supporting vector machine recursion feature elimination by a support vector machine and average influence value by a neural network, and adopting a stability aggregation technology based on ranking to perform integrated feature selection;
The four feature selection methods are adopted to respectively perform feature selection on each group of data, each group of data can obtain four feature weight sequences, firstly, 190 groups of feature sequences obtained by one feature selection method are subjected to stability aggregation to obtain feature weight sequences of the feature selection method, and then the feature weight sequences obtained by four different feature selection methods are subjected to stability aggregation to obtain a final feature weight sequence result.
The invention further improves that the SVR general blood pressure model building module comprises:
The SVR general blood pressure model construction data is random one group of data in 65 tested 3 groups of data in the feature selection module, test data is other two groups of data in 65 tested 3 groups of data selected for participation features and 37 tested data selected for participation features, 213 groups of data are used, the number of feature parameters are respectively valued for 3-20 and 30, 40, 50 and 60 for test, and random data in the 65 tested data are used for SVR general blood pressure model construction;
Taking the average absolute error, the mean value of standard deviation and the standard deviation of 213 groups of predicted data as evaluation standards of selected feature numbers, primarily screening out an optimal feature subset, obtaining a systolic pressure optimal feature subset of 6 and a diastolic pressure optimal feature subset of 7, adding 263 groups of data of 5 tested 50 groups acquired by an experimental platform, and testing and verifying a SVR general blood pressure model and the optimal feature subset number;
Selecting 15, 25, 35, 45 and 55 tested data to participate in feature selection, according to the determined optimal feature subset, comparing the first 6 features of the systolic pressure prediction retention feature weight result with the first 7 features of the diastolic pressure prediction retention feature weight result, and comparing the optimal feature subset obtained by selecting 65 tested data to participate in feature selection with the number of feature parameters, and verifying the stability of the optimal feature subset.
A further improvement of the present invention resides in a video signal processing module comprising:
Recording facial dynamic characteristics by using a camera, wherein the frame rate is 50fps, the distance is 1m, the fingertip PPG signals are synchronously collected, the sampling rate is 1000Hz, and the tested person keeps a static state in the collecting process;
Extracting the region of interest by adopting a frame selection technology, reading a video into frames by using VideoReader functions of matlab, selecting the region of interest by taking pictures of intermediate frames, and respectively calculating the gray average value of the region of interest of each frame of three RGB channels to obtain an RGB sequence;
Performing filtering processing on the video signal by using a color distortion filtering algorithm, performing time normalization processing on the RGB sequence, performing characteristic conversion after Fourier transformation, performing inverse Fourier transformation after energy calculation and weight calculation on the converted RGB sequence, and performing inverse time normalization to obtain the filtered RGB sequence;
Extracting components 1, 2 and 3 from the filtered RGB sequence by using an independent component analysis algorithm, automatically identifying three components by using spectrum analysis, selecting the component 1 with obvious PPG characteristics, and denoising the component 1 by using a Chebyshev I type IIR filter for band-pass filtering to obtain a video pulse wave signal;
And (3) carrying out accuracy estimation on the video pulse wave signal by utilizing the fingertip PPG signal, comparing and evaluating the similarity of the output signals by waveform, and comparing and evaluating the accuracy of the output signals containing physiological information by instantaneous heart rate.
The invention further improves that the video blood pressure prediction module comprises:
Extracting time-frequency domain features of a video pulse wave sequence by adopting a method in a pulse wave signal processing and feature extraction module, extracting a video pulse wave feature subset according to a pulse wave optimal feature subset, evaluating the correctness of the video feature subset for blood pressure prediction by comparing the optimal feature subset of a fingertip PPG signal with the video pulse wave feature subset, and eliminating features with average absolute percentage error higher than 20% again to obtain the optimal video pulse wave feature subset as test data;
the optimal video pulse wave characteristic subset is 6 characteristics of P5, F1, F2, delta T, RP and RP6, based on experimental data of 5 tested video PPG signals collected by an experimental platform, each tested video PPG signal collects 15-30 groups of data in different time periods, and each group of data is about 2 minutes, and the total number of data is 100; and taking the optimal video pulse wave characteristic subset of 100 groups of data as input, and carrying out blood pressure prediction by using the SVR universal blood pressure prediction model obtained through training.
A video stream-based non-contact continuous blood pressure measurement method, comprising:
The pulse wave signal processing and feature extraction module pre-processes the pulse wave signal by adopting a discrete wavelet decomposition technology and a wavelet denoising method, analyzes the waveform features of the PPG and the expression of the physiological information of a reference human body in the PPG waveform, and extracts the time-frequency domain features of the pulse wave sequence;
the single-person model construction and screening module establishes three single-person blood pressure prediction models according to the time-frequency domain characteristics of the pulse wave sequence extracted by the pulse wave signal processing and characteristic extraction module, wherein the three single-person blood pressure prediction models are respectively single-person blood pressure prediction models of an Elman neural network based on particle swarm optimization, and are based on the single-person blood pressure prediction model of a deep belief network and the single-person SVR blood pressure prediction model, and the three single-person blood pressure prediction models are evaluated and screened to obtain an optimal model which is the single-person SVR blood pressure prediction model;
The PPG signal feature selection module performs feature sequencing according to the time-frequency domain features of the pulse wave sequence extracted by the pulse wave signal processing and feature extraction module to obtain a feature weight sequencing result;
The SVR general blood pressure model construction module is used for selecting time-frequency domain features of pulse wave sequences with different numbers according to the feature weight sequencing result to reconstruct an SVR single blood pressure prediction model to obtain an SVR general blood pressure prediction model, and carrying out consistency analysis on the prediction result and actual blood pressure to obtain and verify an optimal feature subset and stability of the optimal feature subset;
The video signal acquisition and processing module firstly acquires face video data to be tested and a fingertip PPG signal by using a camera, extracts RGB sequences from an interested region by using a frame selection technology for the acquired face video data, filters the RGB sequences by using a color distortion filtering algorithm, extracts video pulse wave signals from the filtered RGB sequences by using an independent component analysis algorithm, and carries out consistency estimation with the fingertip PPG signal;
the video blood pressure prediction module adopts a method in the pulse wave signal processing and feature extraction module to extract time-frequency domain features of a video pulse wave sequence, extracts a video pulse wave feature subset according to a pulse wave optimal feature subset, carries out consistency estimation with the optimal pulse wave feature subset of a fingertip PPG signal, obtains the optimal video pulse wave feature subset, takes the optimal video pulse wave feature subset as input, and carries out blood pressure prediction by utilizing a SVR general blood pressure prediction model obtained through training.
The invention has at least the following beneficial technical effects:
1. contactless portability
After the face signals acquired by the video are processed, pulse wave signals are extracted to predict blood pressure, and compared with the traditional method, the system has the non-contact characteristic, can be applied to a mobile phone, acquires the video signals by using a camera of the mobile phone, and realizes convenient and quick blood pressure measurement.
2. Realizing continuous blood pressure prediction
Based on continuous pulse wave and blood pressure data, three single blood pressure prediction models and SVR universal blood pressure models are constructed, continuous blood pressure value blood pressure estimation is carried out by adopting continuous pulse wave data extracted from video, the prediction result meets AAMI and BHS standards, and continuous blood pressure prediction is realized.
3. High running speed
By adopting a feature selection method based on a ranking stability aggregation technology combining data diversity and functional diversity, 6/7 features are extracted from 78 pulse wave feature parameters to predict diastolic pressure and systolic pressure, a SVR universal blood pressure model is constructed, and model training speed and prediction speed are improved.
4. High stability
For the screened single blood pressure prediction model, the prediction result of the SVR general blood pressure model is subjected to consistency analysis with the existing blood pressure measurement method, so that the single blood pressure prediction model has good consistency, and the two methods can be used instead of each other; for the screened optimal feature subset, selecting the feature numbers of 3-20, 30, 40, 50 and 60 for testing, selecting 15, 25, 35, 45 and 55 tested participation feature choices respectively, and verifying the stability and universality of the optimal feature subset; and carrying out consistency estimation on the pulse wave signals extracted from the video and the fingertip PPG signals acquired synchronously, and determining an optimal video pulse wave feature subset to carry out blood pressure prediction.
5. The blood pressure prediction accuracy is improved, and the prediction result meets the A-level standard of AAMI and BSH.
The data diversity, the experimental data come from pulse wave, blood pressure and video signals collected by the experimental platform and continuous blood pressure signals based on a blood pressure database MIMIMIIC; the extraction method of the time-frequency domain features of the pulse wave sequence provides an extraction method of 78 time-frequency domain features of the pulse wave, establishes a pulse wave feature parameter library as completely as possible, and is used for single person model construction and screening; screening blood pressure prediction models, namely establishing three single blood pressure prediction models according to the time-frequency domain characteristics of the pulse wave sequence extracted by the pulse wave signal processing and characteristic extraction module, and screening out an optimal model for constructing a general blood pressure model; sorting the features, namely removing related features interfering with blood pressure prediction in pulse waves by adopting four feature selection methods and an integrated stability aggregation method; the video signal preprocessing, the accuracy of extracting pulse wave signals is improved through a signal separation technology based on the victory and optical characteristics of skin, the RGB sequence is filtered through a color distortion filtering algorithm, and then the video pulse wave signals are extracted from the filtered RGB sequence through an independent component analysis algorithm.
Drawings
Fig. 1 is a design flow chart of a non-contact non-invasive continuous blood pressure measurement method based on video stream.
FIG. 2 is a flow chart of a design for sorting pulse wave features according to the present invention.
Fig. 3 is a flow chart of video signal acquisition and processing.
Fig. 4 shows a graph of the results of video blood pressure prediction, where fig. 4 (a) is the SBP prediction result of the same SVR prediction model for 100 sets of experimental data, and fig. 4 (b) is the DBP prediction result of the general SVR prediction model for 100 sets of experimental data.
FIG. 5 is a graph showing the results of a consistency analysis of video estimated blood pressure and experimentally measured standard blood pressure, wherein FIG. 5 (a) is a graph showing predicted DBP correlation with actual DBP, FIG. 5 (b) is a graph showing DBP predicted Bland Altman, FIG. 5 (c) is a graph showing predicted SBP correlation with actual SBP, and FIG. 5 (d) is a graph showing SBP predicted Bland Altman.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other. The invention will be described in detail below with reference to the drawings in connection with embodiments.
Fig. 1 is a design flow chart of a non-contact non-invasive continuous blood pressure measurement method based on video stream. The system completely establishes a measuring flow based on video streaming non-contact noninvasive continuous blood pressure, extracts the time-frequency domain characteristics of 78 pulse wave sequences according to MIMINIC database data and experimental acquisition data, is used for constructing and screening a single model, evaluates and screens three single blood pressure prediction models, and establishes an optimal model as an SVR single blood pressure prediction model; performing feature screening on the time-frequency domain features of the extracted pulse wave sequence, determining an optimal feature subset, and establishing an SVR universal blood pressure model by adopting an optimal model screened by single blood pressure prediction, so as to further verify the stability of the optimal feature subset and the model; according to the facial videos acquired by the experimental platform, after a region of interest is selected, pulse wave signals are extracted and are subjected to accuracy evaluation with fingertip PPG signals, an optimal video pulse wave feature subset is obtained, and the SVR universal blood pressure prediction model obtained through previous training is utilized to conduct blood pressure prediction, so that video-based blood pressure measurement is achieved.
The invention analyzes the waveform characteristics of PPG and the expression of the physiological information of a reference human body in the PPG waveform, extracts the characteristics of pulse wave sequences as completely as possible, and finally researches the time-frequency characteristic parameters of the pulse waves used in the research to obtain 78 characteristics in total, namely time characteristics, amplitude characteristics, area characteristics, first-order differential characteristics, physiological parameter characteristics, fourier change domain characteristics, wavelet domain characteristics and Hilbert transform domain characteristics. The rest time characteristics and the amplitude characteristics except the pulse wave waveform characteristic quantity K are obtained according to mathematical operation after the wave crest and wave trough characteristic points of the pulse wave are obtained. By adopting a findpeaks function of Matlab, the pulse wave crest is positioned by setting the minimum amplitude and the minimum distance between adjacent points, so that the pulse wave crest is positioned by further searching the minimum value between the adjacent wave crests. In the area characteristics, the trapz function of Matlab is adopted to calculate the ascending area SS and the descending area DS, and then the other characteristics in the area characteristics can be obtained through mathematical operation of a formula by combining corresponding time parameters. When calculating the pulse wave waveform characteristic quantity K based on the calculation of the area characteristic, the Pm value is the area variation quantity of unit time, and the Ps and Pd values are the wave crest and wave trough values of the pulse wave signal in the period. In the first-order differential feature, diff functions are adopted to realize the calculation of the first-order differential of the pulse wave signals, and the positioning of the wave crest and the wave trough point of the first-order differential signals is also realized according to findpeaks functions, so that corresponding first-order differential feature parameters can be calculated. The remaining physiological parameter characteristics can be obtained by mathematical operations of the calculated characteristics. The feature extraction of the Fourier transform domain is that we do Fourier transform on the unit of about 10 pulse wave signals (21 pulse wave signals in total) taking the current pulse wave as the center through the fft function of Matlab. The characteristic extraction of the wavelet domain adopts 'db6' wavelet as a mother wavelet to carry out 6-layer decomposition on the pulse wave signal, and a low-frequency approximation signal Aj and a high-frequency detail signal Dj are obtained. EDM decomposition of pulse wave signals is achieved by adopting Maltab empirical mode decomposition toolboxes, a group of pulse wave signals are taken as an example, 10 IMF components can be obtained through EDM decomposition, a residual sequence is obtained, and IMF energy moment characteristics Ti can be obtained through normalization processing.
The single person model is built, the tested pulse wave-blood pressure pair data in 102 MIMIMIIC databases and the tested pulse wave-blood pressure pair data collected based on an experimental platform are utilized, and the length of each tested pulse wave-blood pressure pair data is 5-100 minutes. Before a single model is built, preprocessing training data, marking the corresponding position of a characteristic parameter mean value which is more than 1.7 times or less than 0.3 times in the training data as 1, taking the systolic pressure as a characteristic, and marking the corresponding position of two adjacent systolic pressure changes which are more than 5 as 1; counting the number of abnormal characteristic parameters in each group of training data, and if the abnormal data quantity exceeds a threshold value 15, eliminating the group of data as abnormal data. And then carrying out normalization operation on the characteristic parameters and the blood pressure, and respectively applying the same mapping rule to the characteristic parameters and the blood pressure data in the test data. And establishing a single blood pressure prediction model of the particle swarm optimized Elman neural network, wherein the Elman neural network is set to be input into 78 characteristic parameters, the numerical value of the hidden node is determined to be between 6 and 19, and the predicted systolic pressure and diastolic pressure are output. Adopting a single hidden layer, adopting a particle swarm optimization algorithm to find an optimal network weight and a threshold value, adopting mutation operation, and re-initializing particles with random probability after each particle update; the SVR single blood pressure prediction model is built, a Libsvm library is adopted for training, an epsilon-SVR model of a radial basis function kernel is adopted, and the super parameters in the model are determined through a grid parameter optimizing function SVMcgForRegress for regression problems, namely: penalty coefficient C, tolerance E of kernel parameter g and termination criterion in kernel function; the method comprises the steps of establishing a single blood pressure prediction model based on a deep belief network, pre-training by adopting a limited Boltzmann machine, keeping as much characteristic information as possible when characteristic parameters are reduced in dimension, initializing a neural network weight by using trained limited Boltzmann machine network parameters, and adjusting the weight by using training data in combination with an error back transmission algorithm; when three single blood pressure prediction models are trained, 60% of data are used for model training by adopting a three-fold cross validation method, and 40% of data are used as test data; according to AAMI standard of blood pressure measurement, average absolute error of 3 times of obtained predicted blood pressure and actual blood pressure and average of standard deviation are used as evaluation indexes of model accuracy, and three modeling methods are compared and evaluated to obtain an optimal SVR single blood pressure prediction model. For 102 tested, SBP prediction results obtained by the SVR construction model all meet the condition that the average absolute error is less than or equal to 3.1mmHg, the error standard deviation is less than or equal to 3.9mmHg, and DBP prediction results all meet the condition that the average absolute error is less than or equal to 1.75mmHg and the error standard deviation is less than or equal to 2.7mmHg.
Fig. 2 is a flow chart of a design of PPG feature selection in accordance with the present invention. A total of 65 data tested are utilized to participate in feature ordering, each tested is about 3 data in different time periods, 190 groups are used, and the length of each data period is about 10 minutes. And selecting RRelieff, 2 filtering type feature selection methods for neighbor component analysis, and 2 embedded feature selection methods for average influence value based on a neural network by means of support vector machine recursion feature elimination of a support vector machine, wherein the total of 4 feature sorting methods are used for respectively sorting the features of each tested data. Feature selection of RRelieff is accomplished by a relieff function of matlab; performing neighbor component analysis by selecting fsrnca functions in matlab, and learning feature weights by using regularization; in the SVM-RFE algorithm, after a feature sequence is input, calculating the weight corresponding to each feature, sequencing by taking the weight as a sequencing standard, removing the feature with the minimum weight, and outputting a feature sequence after updating the sequencing sequence; the MIV algorithm calculates the relative contribution rate of the input variable to the output by comparing the difference between the data obtained after the neural network is input, namely the influence value of the input change on the output, and ranks according to the relative contribution rate to obtain the ordered feature sequence. And (3) adopting a method based on rank stability aggregation, aggregating the feature sequences obtained by each method based on the diversity of the focusing data to obtain 4 total feature sequences, and adopting a method based on rank stability aggregation to obtain a final feature sequence, namely a feature weight sequencing result according to the diversity of the focusing feature selection method. Each feature is scored by a feature selection technique based on a ranked stability aggregation approach, ordered according to score, from feature 1 through feature n. This process is repeated in each feature selection iteration to obtain a set of ranked lists, combining the ranked lists according to their rankings, and creating a final ranked result using the aggregated rankings, the integrated stability aggregate measuring whether the feature is ranked within the number of determined feature parameters, further distinguishing the selected feature rankings.
The SVR general blood pressure model construction data is random one group of data in 65 tested 3 groups of data in the feature selection module, test data is other two groups of data in 65 tested 3 groups of data selected for participation features and 37 tested data selected for participation features, 213 groups of data are used, the number of feature parameters are respectively valued for 3-20 and 30, 40, 50 and 60 for test, and random data in the 65 tested data are used for SVR general blood pressure model construction; taking the average absolute error, the mean value of standard deviation and the standard deviation of 213 groups of predicted data as evaluation standards of selected feature numbers, primarily screening out an optimal feature subset, obtaining a systolic pressure optimal feature subset of 6, a diastolic pressure optimal feature subset of 7, pulse wave features for SBP prediction of F2, P5, W1, RP6, RP5 and delta T, and pulse wave features for DBP prediction of P5, F2, delta T, W2, RP6, F1 and RP5. Adding 263 groups of data of 5 tested 50 groups of data acquired by an experimental platform to test and verify the SVR general blood pressure model and the optimal feature subset number; selecting 15, 25, 35, 45 and 55 tested data to participate in feature selection, according to the determined optimal feature subset, comparing the first 6 features of the systolic pressure prediction retention feature weight result with the first 7 features of the diastolic pressure prediction retention feature weight result, and comparing the optimal feature subset obtained by selecting 65 tested data to participate in feature selection with the number of feature parameters, and verifying the stability of the optimal feature subset.
Fig. 3 is a flow chart of video signal acquisition and processing. The dynamic facial features are recorded by using a video camera, the frame rate is 50fps, the distance is 1m, the fingertip PPG signals are synchronously collected, the sampling rate is 1000Hz, and the tested person keeps a static state in the collecting process. Selecting a cheek middle area below the lower eyelid of the eye by adopting a frame selection technology, extracting the cheek middle area as an interested area, reading a video into frames by using a VideoReader function of matlab, selecting the interested area by taking a picture of an intermediate frame, and respectively calculating the gray average value of the interested area of each frame of three RGB channels to obtain an RGB sequence; and filtering the video signal by using a color distortion filtering algorithm, performing time normalization on the RGB sequence, performing characteristic conversion after Fourier transformation, performing inverse Fourier transformation after energy calculation and weight calculation on the converted RGB sequence, and performing inverse time normalization to obtain the filtered RGB sequence. And processing the filtered RGB sequence by using an independent component analysis algorithm, performing zero-mean processing on RGB signals, solving a unmixed matrix by using a Newton iteration method, separating to obtain source signals, and extracting a component 1, a component 2 and a component 3. In the ICA decomposition result, the maximum value of the ratio of the maximum value to the average value in the spectrogram is obtained in a component 1 with obvious PPG characteristics, three components are automatically identified by utilizing spectrum analysis, the component 1 with obvious PPG characteristics is selected, a Chebyshev I type IIR filter is adopted for denoising the component 1 by band-pass filtering, the passband of the band-pass filter is set to be 0.8Hz-3Hz, the blocking frequencies are 0.5Hz and 4Hz, and the maximum passband attenuation of 1dB and the minimum stopband attenuation of 15dB are finally obtained, and 100 groups of video pulse wave signals are finally obtained.
Video blood pressure prediction, which is to estimate the accuracy of a video pulse wave signal by utilizing a fingertip PPG signal collected synchronously, and evaluate the accuracy of an output signal for blood pressure prediction by comparing the features extracted by the fingertip PPG and the video PPG and combining an optimal feature subset, wherein the results are shown in a table 1, two features with larger average absolute percentages of W1 and W2 are deleted, and P5, F1, F2, delta T, RP and RP6 are used for predicting systolic and diastolic pressures of the video signal to obtain the optimal video pulse wave feature subset. And taking the optimal video pulse wave characteristic subset as input, and carrying out blood pressure prediction by taking the 6 characteristics tested in the previous 65 steps as input training to obtain an SVR general blood pressure prediction model.
TABLE 1 mean of MAPE of pulse wave extracted from 100 groups of tested data video and fingertip PPG extracted features
FIG. 4 is a graph showing the results of video blood pressure prediction, and as can be seen from FIGS. 4 (a) and (b), the SBP/DBP prediction results of 100 experimental data groups all meet the MAE of 5mmHg or less and the standard deviation STD of 8mmHg or less in AAMI standard. As shown in Table 2, the SBP/DBP predictions for the 100 experimental data set also met and were far superior to the BHS A grade criteria: the error is less than 5mmHg, 10mmHg and 15mmHg, and the ratio is as high as 80%,99% and more than 1. Fig. 5 shows a consistency analysis of the blood pressure estimated by the video and the standard blood pressure measured by the experiment, and as can be seen from fig. 5 (a) and (c), the systolic blood pressure and the diastolic blood pressure estimated by the video and the standard systolic blood pressure and the diastolic blood pressure measured by the experiment have strong correlation, and the linear correlation coefficients reach 0.94. The Bland-Altman graphs in FIGS. 5 (b) and 5 (d) show that the mean value of the difference between the SBP estimated from the video and the standard SBP measured experimentally is 0.5mmHg, the mean value of the difference between the DBP estimated from the video and the standard DBP measured experimentally is 0.4mmHg, which are both closer to the dashed line representing the difference of 0, and the two blood pressure measurement methods are found to have stronger consistency.
TABLE 2 comparison of the video-based SBP/DBP prediction cumulative error percentage with the BHS Standard
While the invention has been described in detail in the foregoing general description and with reference to specific embodiments thereof, it will be apparent to one skilled in the art that modifications and improvements can be made thereto. Accordingly, such modifications or improvements may be made without departing from the spirit of the invention and are intended to be within the scope of the invention as claimed.
Claims (8)
1. A video stream-based non-contact continuous blood pressure measurement system, comprising:
The pulse wave signal processing and feature extracting module is used for preprocessing a pulse wave signal by adopting a discrete wavelet decomposition technology and a wavelet denoising method, analyzing the waveform features of the PPG and the expression of the physiological information of a reference human body in the PPG waveform, and extracting the time-frequency domain features of a pulse wave sequence;
The single-person model construction and screening module is used for establishing three single-person blood pressure prediction models according to the time-frequency domain characteristics of the pulse wave sequence extracted by the pulse wave signal processing and characteristic extraction module, wherein the three single-person blood pressure prediction models are respectively single-person blood pressure prediction models of an Elman neural network based on particle swarm optimization, and are based on the single-person blood pressure prediction model of a deep belief network and the single-person blood pressure prediction model of SVR, and the three single-person blood pressure prediction models are evaluated and screened to obtain an optimal model which is the single-person blood pressure prediction model of SVR;
The PPG signal feature selection module is used for carrying out feature sequencing according to the time-frequency domain features of the pulse wave sequence extracted by the pulse wave signal processing and feature extraction module to obtain a feature weight sequencing result;
The SVR general blood pressure model construction module is used for selecting time-frequency domain features of pulse wave sequences with different numbers according to the feature weight sequencing result to reconstruct an SVR single blood pressure prediction model to obtain an SVR general blood pressure prediction model, and carrying out consistency analysis on the prediction result and actual blood pressure to obtain and verify an optimal feature subset and stability of the optimal feature subset;
The video signal acquisition and processing module firstly acquires the face video data to be tested and the fingertip PPG signal by using a camera, extracts RGB sequences from the region of interest by using a frame selection technology for the acquired face video data, filters the RGB sequences by using a color distortion filtering algorithm, extracts video pulse wave signals from the filtered RGB sequences by using an independent component analysis algorithm, and carries out consistency estimation with the fingertip PPG signal;
the video blood pressure prediction module is used for extracting time-frequency domain features of a video pulse wave sequence by adopting a method in the pulse wave signal processing and feature extraction module, extracting a video pulse wave feature subset according to a pulse wave optimal feature subset, carrying out consistency estimation with the optimal pulse wave feature subset of a fingertip PPG signal, obtaining the optimal video pulse wave feature subset, taking the optimal video pulse wave feature subset as input, and carrying out blood pressure prediction by utilizing a SVR general blood pressure prediction model obtained through training.
2. The video-stream-based noncontact continuous blood pressure measurement system of claim 1, wherein the pulse wave signal processing and feature extraction module comprises:
Extracting time domain features, namely adopting a findpeaks function of Matlab, setting the minimum amplitude and the minimum distance between adjacent points to realize the positioning of pulse wave peaks, further searching the minimum value between the adjacent wave peaks to realize the positioning of the pulse wave peaks, adopting a trapz function of Matlab to realize the calculation of a rising area SS and a falling area DS, and adopting a diff function to realize the calculation of first-order differentiation of pulse wave signals to obtain the time domain features of the pulse waves;
Carrying out frequency domain feature extraction, namely carrying out Fourier transformation on each of left and right 10 pulse wave signals taking a current pulse wave as a center through a fft function of Matlab, adopting a max function of Matlab to realize the determination of the frequency and amplitude between 0.3-1.6Hz and 1.6-3Hz in fundamental waves and second harmonics of the pulse wave signals, adopting 'db6' wavelet as a mother wavelet to carry out 6-layer decomposition extraction on the pulse wave signals to obtain wavelet domain features, adopting an empirical mode decomposition tool box of Maltab to realize the decomposition of the pulse wave signals, and then carrying out Hilbert transformation on IMF components obtained by decomposition to obtain Hilbert transform domain features;
The co-extraction includes 78 time-frequency domain features, including time features, amplitude features, and area features.
3. The video-stream-based noncontact continuous blood pressure measurement system of claim 1, wherein the single person model construction and screening module comprises:
The method comprises the steps of establishing a single blood pressure prediction model of an Elman neural network with particle swarm optimization, adopting a single hidden layer, adopting a particle swarm optimization algorithm to find an optimal network weight and a threshold value, adopting mutation operation, and re-initializing particles with random probability after each particle update;
The SVR single blood pressure prediction model is built, a Libsvm library is adopted for training, an epsilon-SVR model of a radial basis function kernel is adopted, and the super parameters in the model are determined through a grid parameter optimizing function SVMcgForRegress for regression problems, namely: penalty coefficient C, tolerance E of kernel parameter g and termination criterion in kernel function;
The method comprises the steps of establishing a single blood pressure prediction model based on a deep belief network, pre-training by adopting a limited Boltzmann machine, keeping as much characteristic information as possible when characteristic parameters are reduced in dimension, initializing a neural network weight by using trained limited Boltzmann machine network parameters, and adjusting the weight by using training data in combination with an error back transmission algorithm;
When three single blood pressure prediction models are trained, 60% of data are used for model training by adopting a three-fold cross validation method, and 40% of data are used as test data; according to AAMI standard of blood pressure measurement, average absolute error of 3 times of obtained predicted blood pressure and actual blood pressure and average of standard deviation are used as evaluation indexes of model accuracy, and three modeling methods are compared and evaluated to obtain an optimal SVR single blood pressure prediction model.
4. The video-stream-based non-contact continuous blood pressure measurement system according to claim 1, wherein the PPG signal feature selection module further comprises a feature selection method as designed herein, specifically comprising:
using 65 tested data to participate in feature selection, wherein each tested data is about 3 data in different time periods, 190 groups of data are used, and each data length is ten minutes;
Selecting RRelieff, a filtering type feature selection method for neighbor component analysis and two embedded feature selection methods for supporting vector machine recursion feature elimination by a support vector machine and average influence value by a neural network, and adopting a stability aggregation technology based on ranking to perform integrated feature selection;
The four feature selection methods are adopted to respectively perform feature selection on each group of data, each group of data can obtain four feature weight sequences, firstly, 190 groups of feature sequences obtained by one feature selection method are subjected to stability aggregation to obtain feature weight sequences of the feature selection method, and then the feature weight sequences obtained by four different feature selection methods are subjected to stability aggregation to obtain a final feature weight sequence result.
5. The video-stream-based non-contact continuous blood pressure measurement system of claim 1, wherein the SVR generic blood pressure model building module comprises:
The SVR general blood pressure model construction data is random one group of data in 65 tested 3 groups of data in the feature selection module, test data is other two groups of data in 65 tested 3 groups of data selected for participation features and 37 tested data selected for participation features, 213 groups of data are used, the number of feature parameters are respectively valued for 3-20 and 30, 40, 50 and 60 for test, and random data in the 65 tested data are used for SVR general blood pressure model construction;
Taking the average absolute error, the mean value of standard deviation and the standard deviation of 213 groups of predicted data as evaluation standards of selected feature numbers, primarily screening out an optimal feature subset, obtaining a systolic pressure optimal feature subset of 6 and a diastolic pressure optimal feature subset of 7, adding 263 groups of data of 5 tested 50 groups acquired by an experimental platform, and testing and verifying a SVR general blood pressure model and the optimal feature subset number;
Selecting 15, 25, 35, 45 and 55 tested data to participate in feature selection, according to the determined optimal feature subset, comparing the first 6 features of the systolic pressure prediction retention feature weight result with the first 7 features of the diastolic pressure prediction retention feature weight result, and comparing the optimal feature subset obtained by selecting 65 tested data to participate in feature selection with the number of feature parameters, and verifying the stability of the optimal feature subset.
6. The video-stream-based noncontact continuous blood pressure measurement system of claim 1, wherein the video signal processing module includes:
Recording facial dynamic characteristics by using a camera, wherein the frame rate is 50fps, the distance is 1m, the fingertip PPG signals are synchronously collected, the sampling rate is 1000Hz, and the tested person keeps a static state in the collecting process;
Extracting the region of interest by adopting a frame selection technology, reading a video into frames by using VideoReader functions of matlab, selecting the region of interest by taking pictures of intermediate frames, and respectively calculating the gray average value of the region of interest of each frame of three RGB channels to obtain an RGB sequence;
Performing filtering processing on the video signal by using a color distortion filtering algorithm, performing time normalization processing on the RGB sequence, performing characteristic conversion after Fourier transformation, performing inverse Fourier transformation after energy calculation and weight calculation on the converted RGB sequence, and performing inverse time normalization to obtain the filtered RGB sequence;
Extracting components 1, 2 and 3 from the filtered RGB sequence by using an independent component analysis algorithm, automatically identifying three components by using spectrum analysis, selecting the component 1 with obvious PPG characteristics, and denoising the component 1 by using a Chebyshev I type IIR filter for band-pass filtering to obtain a video pulse wave signal;
And (3) carrying out accuracy estimation on the video pulse wave signal by utilizing the fingertip PPG signal, comparing and evaluating the similarity of the output signals by waveform, and comparing and evaluating the accuracy of the output signals containing physiological information by instantaneous heart rate.
7. The video-stream-based non-contact continuous blood pressure measurement system of claim 1, wherein the video blood pressure prediction module comprises:
Extracting time-frequency domain features of a video pulse wave sequence by adopting a method in a pulse wave signal processing and feature extraction module, extracting a video pulse wave feature subset according to a pulse wave optimal feature subset, evaluating the correctness of the video feature subset for blood pressure prediction by comparing the optimal feature subset of a fingertip PPG signal with the video pulse wave feature subset, and eliminating features with average absolute percentage error higher than 20% again to obtain the optimal video pulse wave feature subset as test data;
the optimal video pulse wave characteristic subset is 6 characteristics of P5, F1, F2, delta T, RP and RP6, based on experimental data of 5 tested video PPG signals collected by an experimental platform, each tested video PPG signal collects 15-30 groups of data in different time periods, and each group of data is about 2 minutes, and the total number of data is 100; and taking the optimal video pulse wave characteristic subset of 100 groups of data as input, and carrying out blood pressure prediction by using the SVR universal blood pressure prediction model obtained through training.
8. A video stream-based non-contact continuous blood pressure measurement method, comprising:
The pulse wave signal processing and feature extraction module pre-processes the pulse wave signal by adopting a discrete wavelet decomposition technology and a wavelet denoising method, analyzes the waveform features of the PPG and the expression of the physiological information of a reference human body in the PPG waveform, and extracts the time-frequency domain features of the pulse wave sequence;
the single-person model construction and screening module establishes three single-person blood pressure prediction models according to the time-frequency domain characteristics of the pulse wave sequence extracted by the pulse wave signal processing and characteristic extraction module, wherein the three single-person blood pressure prediction models are respectively single-person blood pressure prediction models of an Elman neural network based on particle swarm optimization, and are based on the single-person blood pressure prediction model of a deep belief network and the single-person SVR blood pressure prediction model, and the three single-person blood pressure prediction models are evaluated and screened to obtain an optimal model which is the single-person SVR blood pressure prediction model;
The PPG signal feature selection module performs feature sequencing according to the time-frequency domain features of the pulse wave sequence extracted by the pulse wave signal processing and feature extraction module to obtain a feature weight sequencing result;
The SVR general blood pressure model construction module is used for selecting time-frequency domain features of pulse wave sequences with different numbers according to the feature weight sequencing result to reconstruct an SVR single blood pressure prediction model to obtain an SVR general blood pressure prediction model, and carrying out consistency analysis on the prediction result and actual blood pressure to obtain and verify an optimal feature subset and stability of the optimal feature subset;
The video signal acquisition and processing module firstly acquires face video data to be tested and a fingertip PPG signal by using a camera, extracts RGB sequences from an interested region by using a frame selection technology for the acquired face video data, filters the RGB sequences by using a color distortion filtering algorithm, extracts video pulse wave signals from the filtered RGB sequences by using an independent component analysis algorithm, and carries out consistency estimation with the fingertip PPG signal;
the video blood pressure prediction module adopts a method in the pulse wave signal processing and feature extraction module to extract time-frequency domain features of a video pulse wave sequence, extracts a video pulse wave feature subset according to a pulse wave optimal feature subset, carries out consistency estimation with the optimal pulse wave feature subset of a fingertip PPG signal, obtains the optimal video pulse wave feature subset, takes the optimal video pulse wave feature subset as input, and carries out blood pressure prediction by utilizing a SVR general blood pressure prediction model obtained through training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210193095.8A CN114569096B (en) | 2022-02-28 | 2022-02-28 | Non-contact continuous blood pressure measurement method and system based on video stream |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210193095.8A CN114569096B (en) | 2022-02-28 | 2022-02-28 | Non-contact continuous blood pressure measurement method and system based on video stream |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114569096A CN114569096A (en) | 2022-06-03 |
| CN114569096B true CN114569096B (en) | 2024-05-24 |
Family
ID=81777330
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210193095.8A Active CN114569096B (en) | 2022-02-28 | 2022-02-28 | Non-contact continuous blood pressure measurement method and system based on video stream |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114569096B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115363598B (en) * | 2022-09-19 | 2023-03-24 | 深圳市太美亚电子科技有限公司 | Electrocardiosignal processing method and system |
| CN115886763A (en) * | 2022-11-03 | 2023-04-04 | 重庆大学 | Blood pressure characteristic information extraction method and blood pressure value estimation system |
| CN116299684B (en) * | 2023-05-17 | 2023-07-21 | 成都理工大学 | Novel microseismic classification method based on bimodal neurons in artificial neural network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20090052442A (en) * | 2007-11-21 | 2009-05-26 | 엘지전자 주식회사 | Non-invasive blood pressure measuring method and device |
| CN103908236A (en) * | 2013-05-13 | 2014-07-09 | 天津点康科技有限公司 | Automatic blood pressure measuring system |
| CN106037694A (en) * | 2016-05-13 | 2016-10-26 | 吉林大学 | Continuous blood pressure measuring device based on pulse waves |
| CN113288092A (en) * | 2021-05-06 | 2021-08-24 | 福建师范大学 | Blood pressure detection method and terminal for extracting pulse wave based on video |
| CN113303776A (en) * | 2021-05-28 | 2021-08-27 | 西安理工大学 | Non-contact blood pressure measuring method based on cyclic neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DK3554354T3 (en) * | 2016-12-19 | 2023-10-23 | Nuralogix Corp | System and method for contactless blood pressure determination |
-
2022
- 2022-02-28 CN CN202210193095.8A patent/CN114569096B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20090052442A (en) * | 2007-11-21 | 2009-05-26 | 엘지전자 주식회사 | Non-invasive blood pressure measuring method and device |
| CN103908236A (en) * | 2013-05-13 | 2014-07-09 | 天津点康科技有限公司 | Automatic blood pressure measuring system |
| CN106037694A (en) * | 2016-05-13 | 2016-10-26 | 吉林大学 | Continuous blood pressure measuring device based on pulse waves |
| CN113288092A (en) * | 2021-05-06 | 2021-08-24 | 福建师范大学 | Blood pressure detection method and terminal for extracting pulse wave based on video |
| CN113303776A (en) * | 2021-05-28 | 2021-08-27 | 西安理工大学 | Non-contact blood pressure measuring method based on cyclic neural network |
Non-Patent Citations (2)
| Title |
|---|
| 基于IPPG非接触式血压测量方法;孙晟珂;李凯扬;;信息技术;20201022(第10期);全文 * |
| 基于脉搏波的无创连续血压监测模型簇研究;吴海燕;季忠;李孟泽;;仪器仪表学报;20200831(第07期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114569096A (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114569096B (en) | Non-contact continuous blood pressure measurement method and system based on video stream | |
| CN108416367B (en) | Sleep staging method based on multi-sensor data decision-level fusion | |
| CN113349752B (en) | Wearable device real-time heart rate monitoring method based on sensing fusion | |
| Nirala et al. | Detection of type-2 diabetes using characteristics of toe photoplethysmogram by applying support vector machine | |
| CN109833035B (en) | Classification prediction data processing method of pulse wave blood pressure measuring device | |
| US20190295729A1 (en) | Universal non-invasive blood glucose estimation method based on time series analysis | |
| CN113729707A (en) | FECNN-LSTM-based emotion recognition method based on multi-mode fusion of eye movement and PPG | |
| CN110037668B (en) | System for judging age, health state and malignant arrhythmia identification by combining pulse signal time-space domain with model | |
| CN112190261A (en) | Autism electroencephalogram signal classification device based on resting brain network | |
| CN114781465B (en) | rPPG-based non-contact fatigue detection system and method | |
| Zhao et al. | An IoT-based wearable system using accelerometers and machine learning for fetal movement monitoring | |
| CN112806977B (en) | Physiological parameter measuring method based on multi-scale fusion network | |
| CN115211858A (en) | Emotion recognition method and system based on deep learning and storable medium | |
| CN116369877A (en) | Noninvasive blood pressure estimation method based on photoelectric volume pulse wave | |
| CN116451110A (en) | Blood glucose prediction model construction method based on signal energy characteristics and pulse period | |
| CN112120711B (en) | Noninvasive diabetes prediction system and method based on photoplethysmography pulse waves | |
| Yen et al. | Blood Pressure and Heart Rate Measurements Using Photoplethysmography with Modified LRCN. | |
| CN114387479A (en) | Non-contact heart rate measurement method and system based on face video | |
| CN116172531B (en) | Blood pressure change trend estimation method based on wavelet dispersion transformation | |
| CN112001862B (en) | Non-contact type apparent heart rate detection method for eliminating motion noise of video heart impact signal | |
| CN116211308A (en) | Method for evaluating body fatigue under high-strength exercise | |
| CN112274145A (en) | Method and device for processing near-infrared brain function imaging data and storage medium | |
| CN115089145A (en) | Intelligent blood pressure prediction method based on multi-scale residual error network and PPG signal | |
| CN115270847A (en) | Design decision electroencephalogram recognition method based on wavelet packet decomposition and convolutional neural network | |
| Suriani et al. | Non-contact facial based vital sign estimation using convolutional neural network approach |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |