CN111887834A - Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization - Google Patents

Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization Download PDF

Info

Publication number
CN111887834A
CN111887834A CN202010682668.4A CN202010682668A CN111887834A CN 111887834 A CN111887834 A CN 111887834A CN 202010682668 A CN202010682668 A CN 202010682668A CN 111887834 A CN111887834 A CN 111887834A
Authority
CN
China
Prior art keywords
signal
heartbeat
template signal
training sample
sample set
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.)
Granted
Application number
CN202010682668.4A
Other languages
Chinese (zh)
Other versions
CN111887834B (en
Inventor
焦昶哲
程家馨
刘源洁
缑水平
毛莎莎
李阳阳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202010682668.4A priority Critical patent/CN111887834B/en
Publication of CN111887834A publication Critical patent/CN111887834A/en
Application granted granted Critical
Publication of CN111887834B publication Critical patent/CN111887834B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • A61B5/024—Detecting, measuring or recording pulse rate or heart rate
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
    • A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
    • A61B5/1102—Ballistocardiography
    • 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
    • A61B5/7225—Details of analog processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation
    • 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
    • A61B5/7264—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)
  • Artificial Intelligence (AREA)
  • Medical Informatics (AREA)
  • Physiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Public Health (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Cardiology (AREA)
  • Power Engineering (AREA)
  • Dentistry (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention discloses a beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization, and mainly solves the problems that in the prior art, the dependency on artificial markers is high and the BCG signal heart rate estimation accuracy is low. The scheme is as follows: collecting an original heart impact tracing signal and a finger electric signal; extracting the heartbeat signal characteristics of the ballistocardiogram signal, dividing the heartbeat signal into a positive packet and a negative packet, and then dividing the heartbeat signal into a training set and a test set: learning the training sample set to obtain an initialized heartbeat template signal, and reducing the dimension of the initialized heartbeat template signal to obtain a dimension-reduced heartbeat template signal; performing iterative optimization on the heart beat template signal subjected to dimension reduction to obtain an optimal heart beat template signal; and carrying out classification detection on the test sample set by using the optimal heartbeat template signal to obtain a final heart rate detection result. The method improves the accuracy of estimating the BCG signal heart rate, has low requirement on the initialization of the heart beat characteristics and low cost of manual marking, and can be used for detecting the heart beat of the non-accurately marked ballistocardiogram signal.

Description

Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization
Technical Field
The invention belongs to the technical field of medical signal processing, and further relates to a heart rate estimation method which can be used for heartbeat detection of non-accurately marked ballistocardiogram signals.
Background
Heart rate is one of the important vital signs to assess the physical health of people, especially patients with cardiovascular diseases. Most patients with heart disease need to take medicine for a lifetime, and even if clinical symptoms disappear, the heart problem may recur at any time. Real-time heart rate monitoring is an essential means of preventing heart disease. Current methods for measuring heart rate are: electrocardiography ECG, photoplethysmography PPG, phonocardiography PCG and ballistocardiography BCG. Each method determines heart rate by measuring different phenomena occurring in the human body during the heart beat or heart cycle.
Compared with ECG, BCG records ballistocardiogram signals in a non-invasive and comfortable way, has higher precision than PPG and PCG, has the advantages of non-invasiveness, easy operation, arrangement of sensors and embedded systems and the like, and is more suitable for long-term monitoring at home. A hydraulic bed using cardiac shock imaging with four sensors, the signal obtained being a superposition of BCG and breathing signals, can be used to collect heartbeat data of a person during sleep. It has four hydraulic pressure sensors placed in parallel below the mattress, provides a flexible, low-cost and stable long-term monitoring solution for heart rate and other vital signs, such as blood pressure and respiratory rate.
The multi-example learning is a weak supervision method for solving the problem of inaccurate marking, and is a learning problem taking a multi-example packet as a training unit. In multi-instance learning, a training set consists of a set of multi-instance packets with class labels, each of which contains several instances without class labels. If the multi-instance package contains at least one positive instance, the package is marked as a positive-class multi-instance package, i.e., a positive package. If all instances of the multi-instance package are negative instances, the package is marked as a negative class multi-instance package, i.e., a negative package. The purpose of multi-instance learning is to build a multi-instance classifier through learning of multi-instance packets with classification labels and apply the classifier to predictions of unknown multi-instance packets. In multi-instance learning, the labels of the multi-instance packages are known and the labels of the instances are unknown.
Currently, the research methods for BCG signal heart rate estimation mainly include a frequency analysis-based method, a clustering-based method, a deep learning network-based method, and the like. For example:
the iean electronic science and technology university provides a heart rate estimation method based on a bidirectional cyclic neural network and a regression network in the patent of 'a deep regression heart rate estimation method of ballistocardiogram signals' (patent application number: CN201910688377.3, publication number: CN 110420019A). The heart rate estimation method provided by the method is characterized in that a bidirectional cyclic neural network is used for obtaining the periodic characteristics and the amplitude characteristics of the heart beat signals, and the heart rate value is estimated by simultaneously utilizing the periodic characteristics and the amplitude characteristics of the ballistocardiogram signals through a regression network, so that the heart rate estimation steps of the ballistocardiogram signals are simplified, and the estimation error of the heart rate of the ballistocardiogram signals is effectively reduced. However, the method is a supervised method, and a large amount of labor is required to obtain an accurately labeled data set, so that the method is limited in practical application.
The Guilin electronic science and technology university provides a heart beat signal classification method based on a high-order spectrum and a convolutional neural network in the patent of 'a classification method of a heart attack signal based on a high-order spectrum and a convolutional neural network' (patent application number: CN201910690101.9, publication number: CN 110327055A). The method for classifying the cardiac shock signals comprises the following steps: denoising and preprocessing the obtained data set training sample by adopting a Chebyshev and wavelet transform filtering method to obtain a pure heart impact signal; performing high-order spectral feature analysis on the cardiac shock signal to obtain feature information of the amplitude and the phase of the signal; and constructing a convolutional neural network model, and taking the characteristic information of the amplitude and the phase of the step signal as the input of the convolutional neural network model to obtain a classification result. The method utilizes the time-shift invariance, the scale variability and the phase retentivity of a high-order spectrum to extract the characteristics, not only can more signal information be reserved, but also the colored Gaussian noise can be inhibited, and therefore the classification performance of the heart attack signals is improved; meanwhile, the method has better generalization performance and effectively solves the problem of two-dimensional template matching of high-order spectrum application. But the method has the following disadvantages: the same good classification accuracy is expected to be obtained for different data, a large number of feature extraction and feature selection are required, the workload of scientific researchers is obviously greatly increased, and the classification result of the method is influenced by the difference of the feature selection.
Disclosure of Invention
The invention aims to provide a heart rate estimation method based on multi-example learning and evolutionary optimization aiming at the defects of the prior art so as to realize accurate estimation of the heart rate of a ballistocardiogram signal, fully utilize the characteristics of multi-example learning weak supervision and multi-target optimization algorithm, reduce the selection of different characteristics of the signal, reduce the labor cost and improve the accuracy of BCG signal heartbeat classification and the accuracy of heart rate detection.
In order to achieve the purpose, the technical scheme of the invention comprises the following steps:
(1) collecting original ballistocardiograph signals and finger electric signals, wherein the sampling frequency is 100Hz, and filtering the signals to obtain filtered ballistocardiograph signals b and finger electric signals f;
(2) extracting a heartbeat signal characteristic fb of the ballistocardiogram signal b;
(3) forming a multi-example positive packet by using the wave crests of the ballistocardiogram signal b at the same time of each wave crest of the finger electric signal f and the heartbeat signal characteristics corresponding to the left and right wave crests of the ballistocardiogram signal b at the same time, recording the heartbeat signal characteristics corresponding to the rest wave crests of the ballistocardiogram signal b between the two positive packets as negative packets, and dividing the positive and negative packets into a training sample set x and a test sample y according to the proportion of 1: 1;
(4) learning the training sample set to obtain an initialized heartbeat template signal s:
(4a) calculating the mean value mu of examples in all negative packets of the training sample set xbSum variance σbCovariance matrix of all training samples
Figure BDA0002586406570000031
And by aligning the
Figure BDA0002586406570000032
Decomposing the eigenvalue to obtain an eigenvector U and an eigenvalue D;
(4b) according to the result in (4a), preprocessing whitening and normalization processing is carried out on the training sample set x in sequence to obtain a preprocessed training sample set
Figure BDA0002586406570000033
(4c) Setting the heartbeat template signal as s, and sequentially carrying out whitening and normalization preprocessing on s according to the result in the step (4a) to obtain a preprocessed heartbeat template signal
Figure BDA0002586406570000034
(4d) Calculating a preprocessed heartbeat template signal
Figure BDA0002586406570000035
And training sample set
Figure BDA0002586406570000036
Cosine similarity statistic of mid-forward packet
Figure BDA0002586406570000037
And
Figure BDA0002586406570000038
and training sample set
Figure BDA0002586406570000039
Cosine similarity statistic of medium and negative packets
Figure BDA00025864065700000310
To be provided with
Figure BDA00025864065700000311
And
Figure BDA00025864065700000312
the similarity of the middle and positive packet examples is maximum, the similarity of the middle and positive packet examples is minimum, the target equation is used, the module value of the initialized heartbeat template signal s 'is 1, the target equation is solved, and the initialized heartbeat template signal s' is obtained;
(5) carrying out dimension reduction processing on the initialized heartbeat template signal s 'to obtain a heartbeat template signal s' after dimension reduction;
(6) iterative optimization is carried out on the heart beat template signal s' after dimension reduction by using a constraint evolutionary algorithm to obtain an optimal heart beat template signal ubest；
(7) Using the optimal heartbeat template signal ubestAnd carrying out classification detection on the test sample set y to obtain a final heart rate detection result.
Compared with the prior art, the invention has the following advantages:
1) the manual labeling cost is low
The invention learns based on the concept of the packet in the training process, is a learning method for non-precise marking data, does not need to precisely mark each heartbeat position of BCG data signals, and greatly reduces the cost of manual marking;
2) the heartbeat classification result is more accurate, and the heart rate estimation result is more accurate
According to the invention, the initialized heartbeat template signal obtained by multi-instance learning is optimized by evolutionary optimization, and the advantage that a multi-objective evolutionary algorithm has a large search space and is not easy to fall into a local optimal solution is utilized, so that a global optimal solution with excellent performance is easier to find, and the heartbeat position and the non-heartbeat position can be accurately classified, thereby obtaining a more accurate heart rate detection result.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
Fig. 2 is a conceptual illustration of multiple example positive and negative packets in solution 3.
Detailed Description
Embodiments and effects of the present invention will be described in further detail below with reference to the accompanying drawings.
Referring to fig. 1, the implementation steps of the invention are as follows:
step 1, collecting an original cardiac shock tracing signal and a finger electric signal.
The original ballistocardiographic signal is obtained by using v hydraulic sensors at a sampling frequency fcAcquiring v ballistocardiogram signals of a length T of a subject;
the finger electric signal is obtained by a finger-clipped pulse sensor at a sampling frequency fcCollecting a heart pulse signal with the length T of a subject;
simultaneously acquiring v original ballistocardiogram signals and finger electric signals of a subject, and adopting a lower cutoff frequency limit of f to reduce the influence of respiratory components and high-frequency noise in the ballistocardiogram signals on the heart rate estimation performance1Upper limit of f2The second-order Butterworth band-pass filter respectively filters the v ballistocardiograph signals and the finger electric signals to obtain v ballistocardiograph signals b and finger electric signals f after filtering, wherein f1The value range of (a) is based on the upper limit of the frequency of the respiratory component in the ballistocardiogram signal being about 0.3 Hz-0.8 Hz, f2The value range of (a) is based on the lower frequency limit of the high frequency noise component in the ballistocardiogram signal being about 8Hz to 12Hz, and the example is not limited to f1＝0.4Hz，f2＝10Hz，v＝4，fc＝100Hz,T＝60000。
And 2, extracting the heartbeat signal characteristics of the ballistocardiogram signal.
From all the peaks of the detected ballistocardiograph signal b, 45 sampling points are respectively taken from the left and the right by taking the time position of each peak as a central point, and a plurality of signal segments with the length of 91 are formed, namely the heartbeat signal characteristic fb.
And 3, dividing the heartbeat signal characteristics into positive packets and negative packets, and dividing the positive packets and the negative packets into a training set and a test set.
Referring to fig. 2, a multi-example positive packet is formed by using the heartbeat signal characteristics of each wave peak of the finger electric signal f at the same time of carving the wave peak of the ballistocardiogram signal b and corresponding to the left wave peak and the right wave peak of the ballistocardiogram signal b, and the heartbeat signal characteristics corresponding to the rest wave peaks of the ballistocardiogram signal b between the two positive packets are recorded as a negative packet;
the positive and negative packets are divided into a training sample set x and a test sample y according to a ratio of 1: 1.
And 4, learning the training sample set to obtain an initialized heartbeat template signal s.
(4.1) calculate the mean μ of the examples in all negative packets of the training sample set xbSum variance σbThe formula is as follows:
Figure BDA0002586406570000041
Figure BDA0002586406570000042
wherein x isiI is more than or equal to 1 and less than or equal to n, and n is the sum of the samples of all negative packets in the training sample set;
(4.2) calculating covariance matrices of all training samples
Figure BDA0002586406570000051
Is represented as follows:
Figure BDA0002586406570000052
wherein the content of the first and second substances,
Figure BDA0002586406570000053
representing the covariance between the p-th and q-th samples in the training sample set x, i.e.
Figure BDA0002586406570000054
The p is more than or equal to 1 and less than or equal to m, q is more than or equal to 1 and less than or equal to m, m is the total number of samples in the training sample set x, x ispzIs the z-th value in the p-th sample, z is more than or equal to 1 and less than or equal to k, k is the total number of the median values in each sample,
Figure BDA0002586406570000055
is the sample mean of the p-th sample, xqzIs as followsThe z-th value of the q samples,
Figure BDA0002586406570000056
is the sample mean of the q sample;
(4.3) covariance matrix for the training sample
Figure BDA0002586406570000057
Decomposing the eigenvalue to obtain an eigenvector U and an eigenvalue D;
(4.4) according to the result in (4.1), whitening the training sample set x to obtain the whitened training sample set
Figure BDA0002586406570000058
Then to
Figure BDA0002586406570000059
Carrying out normalization processing to obtain a training sample set after preprocessing
Figure BDA00025864065700000510
The formula is expressed as follows:
Figure BDA00025864065700000511
Figure BDA00025864065700000512
wherein the content of the first and second substances,
Figure BDA00025864065700000513
is the evolution of the characteristic value D, UTIs a transpose of the feature vector U,
Figure BDA00025864065700000514
is that
Figure BDA00025864065700000515
The L1 norm;
(4.5) setting the heartbeat template signal as s, carrying out whitening processing on s according to the result in (4.1),after whitening treatment
Figure BDA00025864065700000516
Then to
Figure BDA00025864065700000517
Carrying out normalization processing to obtain a preprocessed heartbeat template signal
Figure BDA00025864065700000518
The formula is expressed as follows:
Figure BDA00025864065700000519
Figure BDA00025864065700000520
wherein the content of the first and second substances,
Figure BDA00025864065700000521
is the evolution of the characteristic value D, UTIs a transpose of the feature vector U,
Figure BDA00025864065700000522
is that
Figure BDA00025864065700000523
The L1 norm;
(4.6) calculating the preprocessed heartbeat template signal
Figure BDA00025864065700000524
Respectively with training sample sets
Figure BDA00025864065700000525
Cosine similarity statistic of mid-forward packet
Figure BDA0002586406570000061
The formula is as follows:
Figure BDA0002586406570000062
wherein the content of the first and second substances,
Figure BDA0002586406570000063
for the h positive packet of training sample set x
Figure BDA0002586406570000064
Neutralizing the preprocessed heartbeat template signal
Figure BDA0002586406570000065
In the most similar example of the present invention,
Figure BDA0002586406570000066
for the v-th example in the h-th positive packet, 1 ≦ h ≦ N+，N+For training sample sets
Figure BDA0002586406570000067
The number of the middle and positive packets;
(4.7) calculating the preprocessed heartbeat template signal
Figure BDA0002586406570000068
Respectively with training sample sets
Figure BDA0002586406570000069
Cosine similarity statistic of medium and negative packets
Figure BDA00025864065700000610
The formula is as follows:
Figure BDA00025864065700000611
wherein the content of the first and second substances,
Figure BDA00025864065700000612
for the g-th example of the kth negative packet in the training sample set x, k is more than or equal to 1 and less than or equal to N-，N-The number of the negative packets is the number of the negative packets,
Figure BDA00025864065700000613
representing the number of instances within the kth negative packet;
(4.8) to obtain the initialized heartbeat template signal s', the preprocessed heartbeat template signal
Figure BDA00025864065700000614
And training sample set
Figure BDA00025864065700000615
The similarity of the medium and positive packet examples is maximum, the similarity of the medium and positive packet examples is minimum, an object equation is constructed, and meanwhile, the module value of the initialized heartbeat template signal s' is 1 to serve as a constraint condition, and the constraint condition aims to prevent the object equation from being solved
Figure BDA00025864065700000616
Tending towards infinity, solving the objective equation yields s', expressed as follows:
Figure BDA00025864065700000617
wherein the content of the first and second substances,
Figure BDA00025864065700000618
is a preprocessed heartbeat template signal, s'TTo initialize the transposition of the heartbeat template signal s',
Figure BDA00025864065700000619
for preprocessed heartbeat template signals
Figure BDA00025864065700000620
And training sample set
Figure BDA00025864065700000621
The cosine similarity statistic of the medium positive packet,
Figure BDA00025864065700000622
for preprocessed heartbeat template signals
Figure BDA00025864065700000623
And training sample set
Figure BDA00025864065700000624
Cosine similarity statistic of mid-forward packet
Figure BDA00025864065700000625
And 5, reducing the dimension of the initialized heartbeat template signal to obtain the heartbeat template signal after dimension reduction.
Because the dimension of the heartbeat template signal is too large, the heartbeat template signal is difficult to converge to an optimal solution in the evolution optimization process, so that the dimension reduction processing needs to be carried out on the initialized heartbeat template signal s 'to obtain the heartbeat template signal s' after the dimension reduction, and the process is as follows:
(5.1) initializing the heartbeat template signal s' at the sampling frequency fsPerforming N-point fourier transform to obtain the amplitudes, frequencies and phases of all frequency domain components after the transform, where N is 91, f in this examples＝128；
(5.2) sorting all the frequency domain components after the transformation from large to small according to amplitude, and taking the sorted L-dimensional frequency domain components to obtain a heart beat template signal s' after the dimension reduction:
Figure BDA0002586406570000071
wherein a isl，wl，θlL is more than or equal to 1 and less than or equal to L, and L is more than or equal to 2 and less than or equal to 12.
Step 6, carrying out iterative optimization on the heart beat template signal s' subjected to dimension reduction by using a constraint evolutionary algorithm to obtain an optimal heart beat template signal ubest。
The existing constrained evolution algorithm has a penalty function method, a feasible solution preference method and a multi-objective optimization method. The present example uses, but is not limited to, an evolutionary algorithm for multi-objective optimization, which is implemented as follows:
(6.1) setting the maximum iteration number tmaxRandomly generating an initial population A of a heartbeat template signal with the population scale of M-100t；
(6.2) setting a heartbeat template signal population AtIn (4d), the objective equation in (4d) is rewritten as a fitness function of the individual, G (u) ═ F (u), p (u)]Wherein:
Figure BDA0002586406570000072
in the formula (I), the compound is shown in the specification,
Figure BDA0002586406570000073
for the h positive packet of training sample set x
Figure BDA0002586406570000074
Neutralizing the preprocessed heartbeat template signal
Figure BDA0002586406570000075
Most similar example, 1 ≦ h ≦ N+，N+For training sample sets
Figure BDA0002586406570000076
The number of the middle and positive packets is increased,
Figure BDA0002586406570000077
is composed of
Figure BDA0002586406570000078
The transpose of (a) is performed,
Figure BDA0002586406570000079
for the g-th example of the kth negative packet in the training sample set x, k is more than or equal to 1 and less than or equal to N-，N-The number of the negative packets is the number of the negative packets,
Figure BDA00025864065700000710
indicating the number of instances within the kth negative packet,
Figure BDA00025864065700000711
is composed of
Figure BDA00025864065700000712
Transposing;
(6.3) obtaining a mutated population B through a mutation operation of a non-uniform mutation operatort：
Because the function of the traditional mutation operator has no direct relation with the mutation algebra, when the evolution reaches a certain algebra, the optimal solution is difficult to obtain due to the lack of local search or the same whole search range; the variation range of the non-uniform variation operator is relatively larger at the initial stage of evolution, and the variation range is smaller and smaller as the evolution reaches the later stage, so that the effect of fine adjustment on the system is achieved, and therefore the non-uniform variation operator is adopted in the embodiment to perform fine adjustment on the initial population AtPerforming mutation operation to obtain a mutated population Bt；
(6.4) generating a progeny population C by crossover operationst：
Since the simplex crossover operator SPX generates offspring based on uniform probability distribution, any fitness information is not needed, and the calculation complexity is low, the simplex crossover operator SPX is adopted in the embodiment to carry out mutation on the population BtPerforming crossover operation to generate offspring population Ct；
(6.5) starting population AtAnd progeny population CtMerging to obtain a collection CAt＝Ct∪At；
(6.6) calculating CA according to the individual fitness function G (u) in (6.2)tFitness of individuals in the population;
(6.7) aggregating CA on the initial population and the offspring populationtNSES by CA using evolution strategytSelecting the fitness of individuals in the population to obtain a t +1 generation population At+1；
(6.8) for the current population AtAnd (4) judging:
if A istSatisfies the maximum number of iterations t in (6.1)maxThen the iteration is stopped and population A is selected according to the pareto optimal ruletTo select the optimal heartbeat template signal ubest；
Otherwise, the t +1 generation population At+1Go to (6.3) and continue the iteration.
Step 7 utilizing the optimal heartbeat template signal ubestAnd carrying out classification detection on the test sample set y to obtain a final heart rate detection result.
The effects of the present invention can be further illustrated by the following simulations.
1. Simulation conditions
The simulation was performed on windows10 professional edition with a CPU fundamental frequency of 3.3GHZ x 2, using MATLAB R2019a software.
2. Emulated content
Simulation I, performing heart rate estimation on 6 subjects by using the conventional ballistocardiogram signal heart rate estimation method HT based on Hilbert transform, respectively obtaining the estimated heart rate of each test set sample of the subjects, and respectively calculating heart rate estimation errors.
And simulating two, performing heart rate estimation on 6 subjects by using the existing energy-based method EN, and respectively acquiring the estimated heart rate of each test set sample of the subjects.
And thirdly, learning an optimal heartbeat signal template for each subject by using the method, respectively obtaining the estimated heart rate of each subject test set sample, independently operating for ten times, and averaging the results.
The heart rate estimation error is calculated by the formula:
Figure BDA0002586406570000081
wherein P is the total number of samples in the test set of subjects, yiAnd
Figure BDA0002586406570000082
the true heart rate and the estimated heart rate, respectively, for the ith sample in the test set, represent absolute value operations.
The three simulations described above tested the heart rate estimation error for each individual, as shown in table 1.
TABLE 1 comparison of heart rate estimation errors of the method and HT and EN algorithms
Test subject Existing HT methods Existing EN method The method of the invention
1 0.85 14.69 0.19±0.04
2 0.79 1.83 0.66±0.16
3 0.48 1.65 0.60±0.38
4 1.26 0.77 0.27±0.10
5 1.07 0.53 0.087±0.09
6 1.05 0.61 0.32±0.05
3. Simulation effect analysis
As can be seen from Table 1, the mean heart rate estimation error of the method of the present invention is 0.35 in the test set samples of six subjects, while the mean heart rate estimation error of the HT method is 0.92 in the test set samples of six subjects, and the mean heart rate estimation error of the EN method is 3.35 in the test set samples of six subjects, which is obviously smaller than the estimation errors of the HT method and the EN method. And it can be seen from the above table that the heart rate estimation error variance of the invention is smaller, which shows that the robustness of the method of the invention is stronger.
The experimental results and experimental analysis show that the discriminant concept can be learned from the data which are not accurately marked by the method, the evolutionary algorithm can search a wider solution space, so that the solution is not easy to fall into local optimum, a better solution than other traditional iterative optimization methods can be obtained, the effective classification of BCG signal heartbeat positions can be realized, and the superiority and robustness of the method in heartbeat data detection are proved.

Claims (10)

1. A beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization is characterized by comprising the following steps:
(1) collecting original ballistocardiograph signals and finger electric signals, wherein the sampling frequency is 100Hz, and filtering the signals to obtain filtered ballistocardiograph signals b and finger electric signals f;
(2) extracting a heartbeat signal characteristic fb of the ballistocardiogram signal b;
(3) forming a multi-example positive packet by using the wave crests of the ballistocardiogram signal b at the same time of each wave crest of the finger electric signal f and the heartbeat signal characteristics corresponding to the left and right wave crests of the ballistocardiogram signal b at the same time, recording the heartbeat signal characteristics corresponding to the rest wave crests of the ballistocardiogram signal b between the two positive packets as negative packets, and dividing the positive and negative packets into a training sample set x and a test sample y according to the proportion of 1: 1;
(4) learning the training sample set to obtain an initialized heartbeat template signal s:
(4a) calculating the mean value mu of examples in all negative packets of the training sample set xbSum variance σbCovariance matrix of all training samples
Figure FDA0002586406560000011
And by aligning the
Figure FDA0002586406560000012
Decomposing the eigenvalue to obtain an eigenvector U and an eigenvalue D;
(4b) according to the result in (4a), preprocessing whitening and normalization processing is carried out on the training sample set x in sequence to obtain a preprocessed training sample set
Figure FDA0002586406560000013
(4c) Setting the heartbeat template signal as s, and sequentially carrying out whitening and normalization preprocessing on s according to the result in the step (4a) to obtain a preprocessed heartbeat template signal
Figure FDA0002586406560000014
(4d) Calculating a preprocessed heartbeat template signal
Figure FDA0002586406560000015
And training sample set
Figure FDA0002586406560000016
Cosine similarity statistic of mid-forward packet
Figure FDA0002586406560000017
And
Figure FDA0002586406560000018
and training sample set
Figure FDA0002586406560000019
Cosine phase of middle and negative packetSimilarity statistic
Figure FDA00025864065600000110
To be provided with
Figure FDA00025864065600000111
And
Figure FDA00025864065600000112
the similarity of the middle and positive packet examples is maximum, the similarity of the middle and positive packet examples is minimum, the target equation is used, the module value of the initialized heartbeat template signal s 'is 1, the target equation is solved, and the initialized heartbeat template signal s' is obtained;
(5) performing dimensionality reduction on the initialized heartbeat template signal s 'to obtain a heartbeat template signal s' subjected to dimensionality reduction;
(6) iterative optimization is carried out on the heart beat template signal s' after dimension reduction by using a constraint evolutionary algorithm to obtain an optimal heart beat template signal ubest；
(7) Using the optimal heartbeat template signal ubestAnd carrying out classification detection on the test sample set y to obtain a final heart rate detection result.
2. The method according to claim 1, wherein the heartbeat signal feature fb of the ballistocardiogram signal b is extracted in (2) by taking 45 sampling points from the left and right sides of all peaks of the detected ballistocardiogram signal b by taking the time position of each peak as the center, and forming a plurality of signal segments with the length of 91, namely the heartbeat signal feature fb.
3. The method of claim 1, wherein (4a) the mean μ of the samples in all negative packets of the training sample set x is calculatedbSum variance σbThe formula is as follows:
Figure FDA0002586406560000021
Figure FDA0002586406560000022
wherein x isiI is more than or equal to 1 and less than or equal to n, and n is the sum of the samples of all the negative packets in the training sample set.
4. The method of claim 1, wherein the covariance matrix of all training samples in (4a)
Figure FDA0002586406560000023
Is represented as follows:
Figure FDA0002586406560000024
wherein, σ (x)p,xq) To represent
Figure FDA0002586406560000025
For the covariance between the p sample and the q sample in the training sample set x, i.e.
Figure FDA0002586406560000026
The p is more than or equal to 1 and less than or equal to m, q is more than or equal to 1 and less than or equal to m, m is the total number of samples in the training sample set x, x ispzIs the z-th value in the p-th sample, z is more than or equal to 1 and less than or equal to k, k is the total number of the median values in each sample,
Figure FDA0002586406560000031
is the sample mean of the p-th sample, xqzFor the z-th value in the q-th sample,
Figure FDA0002586406560000032
is the sample mean of the q sample.
5. The method of claim 1, wherein the preprocessing of whitening and normalizing the training sample set x in (4b) is performed sequentially according to the following formula:
Figure FDA0002586406560000033
Figure FDA0002586406560000034
wherein the content of the first and second substances,
Figure FDA0002586406560000035
is the evolution of the characteristic value D, UTIs a transpose of the feature vector U,
Figure FDA00025864065600000324
is the result of whitening processing on the training sample set x,
Figure FDA0002586406560000036
is that
Figure FDA0002586406560000037
The L1 norm of (a),
Figure FDA0002586406560000038
is to
Figure FDA0002586406560000039
And normalizing the processed result.
6. The method of claim 1, wherein the preprocessing of whitening and normalizing the heart skipping template signal s in (4c) is performed sequentially by the following formula:
Figure FDA00025864065600000310
Figure FDA00025864065600000311
wherein the content of the first and second substances,
Figure FDA00025864065600000312
is the evolution of the characteristic value D, UTIs a transpose of the feature vector U,
Figure FDA00025864065600000313
is the result of whitening the heart skipping template signal s,
Figure FDA00025864065600000314
is that
Figure FDA00025864065600000315
The L1 norm of (a),
Figure FDA00025864065600000316
is to
Figure FDA00025864065600000317
And normalizing the processed result.
7. The method of claim 1, wherein the preprocessed heartbeat template signal is calculated in (4d)
Figure FDA00025864065600000318
Respectively with training sample sets
Figure FDA00025864065600000319
Cosine similarity statistic of mid-forward packet
Figure FDA00025864065600000320
Cosine similarity statistic of sum-minus packet
Figure FDA00025864065600000321
The formula is as follows:
Figure FDA00025864065600000322
Figure FDA00025864065600000323
wherein the content of the first and second substances,
Figure FDA0002586406560000041
for the h positive packet of training sample set x
Figure FDA0002586406560000042
Neutralizing the preprocessed heartbeat template signal
Figure FDA0002586406560000043
In the most similar example of the present invention,
Figure FDA0002586406560000044
for the v-th example in the h-th positive packet, 1 ≦ h ≦ N+，N+For training sample sets
Figure FDA0002586406560000045
The number of the middle and positive packets is increased,
Figure FDA0002586406560000046
for the g-th example of the kth negative packet in the training sample set x, k is more than or equal to 1 and less than or equal to N-，N-The number of the negative packets is the number of the negative packets,
Figure FDA0002586406560000047
Figure FDA0002586406560000048
indicating the number of instances within the kth negative packet.
8. The method of claim 1, wherein the objective equation in (4d) is solved to obtain the initialized heartbeat template signal s', and the formula is as follows:
Figure FDA0002586406560000049
wherein the content of the first and second substances,
Figure FDA00025864065600000410
is a preprocessed heartbeat template signal, s'TTo initialize the transposition of the heartbeat template signal s',
Figure FDA00025864065600000411
for preprocessed heartbeat template signals
Figure FDA00025864065600000412
And training sample set
Figure FDA00025864065600000413
The cosine similarity statistic of the medium positive packet,
Figure FDA00025864065600000414
for preprocessed heartbeat template signals
Figure FDA00025864065600000415
And training sample set
Figure FDA00025864065600000416
Cosine similarity statistic of mid-forward packet
Figure FDA00025864065600000417
9. The method according to claim 1, wherein the initialized heartbeat template signal s' is subjected to dimension reduction in (5) by:
(5a) sampling frequency f for initialized heartbeat template signal ssPerforming N-point Fourier transform to obtain all transformed frequenciesAmplitude, frequency and phase of the domain component, where N is 91, fs＝128；
(5b) Sorting all the frequency domain components after transformation from large to small according to amplitude values, and taking the sorted L-dimensional frequency domain components to obtain a heart beat template signal after dimension reduction:
Figure FDA00025864065600000418
wherein a isl，wl，θlL is more than or equal to 1 and less than or equal to L, and L is more than or equal to 2 and less than or equal to 12.
10. The method according to claim 1, wherein in (6), the heart beat template signal s "after being subjected to the dimension reduction is iteratively optimized by using a multi-objective evolutionary algorithm to obtain the heart beat template signal s" after being subjected to the dimension reduction, and the method is realized as follows:
(6a) setting a maximum number of iterations tmaxRandomly generating an initial population A of a heartbeat template signal with the population scale of M-100t；
(6b) Set heartbeat template signal population AtIn (4d), the objective equation in (4d) is rewritten as a fitness function of the individual, G (u) ═ F (u), p (u)]Wherein:
Figure FDA0002586406560000051
in the formula (I), the compound is shown in the specification,
Figure FDA0002586406560000052
for the h positive packet of training sample set x
Figure FDA0002586406560000053
Neutralizing the preprocessed heartbeat template signal
Figure FDA0002586406560000054
Most similar example, 1 ≦ h ≦ N+，N+For training sample sets
Figure FDA0002586406560000055
The number of the middle and positive packets is increased,
Figure FDA0002586406560000056
is composed of
Figure FDA0002586406560000057
The transpose of (a) is performed,
Figure FDA0002586406560000058
for the g-th example of the kth negative packet in the training sample set x, k is more than or equal to 1 and less than or equal to N-，N-The number of the negative packets is the number of the negative packets,
Figure FDA0002586406560000059
Figure FDA00025864065600000510
indicating the number of instances within the kth negative packet,
Figure FDA00025864065600000511
is composed of
Figure FDA00025864065600000512
Transposing;
(6c) using non-uniform variation operator to initial population AtPerforming mutation operation to obtain a mutated population Bt；
(6d) Adopting simplex crossover operator SPX to pair the mutated population BtPerforming crossover operation to generate offspring population Ct；
(6e) Initial population AtAnd progeny population CtMerging to obtain a collection CAt＝Ct∪At；
(6f) Calculating CA according to the individual fitness function G (u) in (6b)tFitness of individuals in the population;
(6g) for initial population and offspring population, CAtAccording to the evolution strategy NSESSelecting the rows to obtain a t +1 generation population At+1；
(6f) For the current population AtAnd (4) judging:
if A istSatisfies the maximum number of iterations t in (6a)maxThen the iteration is stopped and population A is selected according to the pareto optimal ruletTo select the optimal heartbeat template signal ubest；
Otherwise, the t +1 generation population At+1Go to (6c) and continue the iteration.
CN202010682668.4A 2020-07-15 2020-07-15 Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization Active CN111887834B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010682668.4A CN111887834B (en) 2020-07-15 2020-07-15 Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010682668.4A CN111887834B (en) 2020-07-15 2020-07-15 Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization

Publications (2)

Publication Number Publication Date
CN111887834A true CN111887834A (en) 2020-11-06
CN111887834B CN111887834B (en) 2021-11-02

Family

ID=73193070

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010682668.4A Active CN111887834B (en) 2020-07-15 2020-07-15 Beat-to-beat heart rate detection method based on multi-example learning and evolutionary optimization

Country Status (1)

Country Link
CN (1) CN111887834B (en)

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106202256A (en) * 2016-06-29 2016-12-07 西安电子科技大学 Propagate based on semanteme and mix the Web graph of multi-instance learning as search method
CN108154927A (en) * 2017-12-21 2018-06-12 华子昂 A kind of intelligence traditional Chinese medical science robot software's framework
US20180285774A1 (en) * 2017-03-31 2018-10-04 Yahoo! Inc. Collaborative personalization via simultaneous embedding of users and their preferences
US10217488B1 (en) * 2017-12-15 2019-02-26 Snap Inc. Spherical video editing
CN109977994A (en) * 2019-02-02 2019-07-05 浙江工业大学 A kind of presentation graphics choosing method based on more example Active Learnings
CN110420019A (en) * 2019-07-29 2019-11-08 西安电子科技大学 A kind of depth recurrence heart rate estimation method of ballistocardiography signal
US10561253B2 (en) * 2016-07-29 2020-02-18 Bryte, Inc. Adaptive sleep system using data analytics and learning techniques to improve individual sleep conditions
CN111241965A (en) * 2020-01-06 2020-06-05 重庆邮电大学 Target tracking method for occlusion detection based on multi-example learning
US20200178887A1 (en) * 2016-04-29 2020-06-11 Fitbit, Inc. Sleep monitoring system with optional alarm functionality

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200178887A1 (en) * 2016-04-29 2020-06-11 Fitbit, Inc. Sleep monitoring system with optional alarm functionality
CN106202256A (en) * 2016-06-29 2016-12-07 西安电子科技大学 Propagate based on semanteme and mix the Web graph of multi-instance learning as search method
US10561253B2 (en) * 2016-07-29 2020-02-18 Bryte, Inc. Adaptive sleep system using data analytics and learning techniques to improve individual sleep conditions
US20180285774A1 (en) * 2017-03-31 2018-10-04 Yahoo! Inc. Collaborative personalization via simultaneous embedding of users and their preferences
US10217488B1 (en) * 2017-12-15 2019-02-26 Snap Inc. Spherical video editing
CN108154927A (en) * 2017-12-21 2018-06-12 华子昂 A kind of intelligence traditional Chinese medical science robot software's framework
CN109977994A (en) * 2019-02-02 2019-07-05 浙江工业大学 A kind of presentation graphics choosing method based on more example Active Learnings
CN110420019A (en) * 2019-07-29 2019-11-08 西安电子科技大学 A kind of depth recurrence heart rate estimation method of ballistocardiography signal
CN111241965A (en) * 2020-01-06 2020-06-05 重庆邮电大学 Target tracking method for occlusion detection based on multi-example learning

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
CHANGZHE JIAO ET.AL.: "Heart Beat Characterization from Ballistocardiogram Signals using Extended Functions of Multiple Instances", 《2016 38TH ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY (EMBC)》 *
CHANGZHE JIAO,ET.AL: "Multiple Instance Dictionary Learning for Beat-to-Beat Heart Rate Monitoring from Ballistocardiograms", 《IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING》 *
PELLEGRINI T , CANCES L .: "Cosine-similarity penalty to discriminate sound classes in weakly-supervised sound event detection", 《2019 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN)》 *
安苏阳: "基于多示例学习的计算机辅助肺结节检测研究", 《中国优秀硕士学位论文全文数据库 (医药卫生科技辑)》 *
王敏: "基于智能床垫的心冲击图信号处理及其在心血管健康评估中的应用研究", 《中国优秀硕士学位论文全文数据库》 *

Also Published As

Publication number Publication date
CN111887834B (en) 2021-11-02

Similar Documents

Publication Publication Date Title
Atal et al. Arrhythmia classification with ECG signals based on the optimization-enabled deep convolutional neural network
Zhang et al. Automated detection of myocardial infarction using a gramian angular field and principal component analysis network
Sangaiah et al. An intelligent learning approach for improving ECG signal classification and arrhythmia analysis
Li et al. Classification of ECG signals based on 1D convolution neural network
CN111449645B (en) Intelligent classification and identification method for electrocardiogram and heartbeat
Ai et al. Classification of parkinsonian and essential tremor using empirical mode decomposition and support vector machine
Chen et al. Atrial fibrillation detection using a feedforward neural network
Singh et al. Evaluation of electrocardiogram for biometric authentication
Belgacem et al. ECG based human authentication using wavelets and random forests
Wu et al. A novel method to detect multiple arrhythmias based on time-frequency analysis and convolutional neural networks
CN113397555A (en) Arrhythmia classification algorithm of C-LSTM for physiological parameter monitoring
Li et al. Robust ECG biometrics using GNMF and sparse representation
Zhang et al. ECG signal classification with deep learning for heart disease identification
Altan et al. ECG based human identification using second order difference plots
Kayikcioglu et al. Time-frequency approach to ECG classification of myocardial infarction
Pal et al. Increasing the accuracy of ECG based biometric analysis by data modelling
CN113901893A (en) Electrocardiosignal identification and classification method based on multiple cascade deep neural network
Prakash et al. A system for automatic cardiac arrhythmia recognition using electrocardiogram signal
Philip et al. Identifying arrhythmias based on ecg classification using enhanced-PCA and enhanced-SVM methods
Singh et al. Short and noisy electrocardiogram classification based on deep learning
Alam et al. Wearable respiration monitoring: interpretable inference with context and sensor biomarkers
Ganguly et al. A non-invasive approach for fetal arrhythmia detection and classification from ecg signals
Zhang et al. Automated localization of myocardial infarction from vectorcardiographic via tensor decomposition
Allam et al. A deformable CNN architecture for predicting clinical acceptability of ECG signal
Wu et al. ECG identification based on neural networks

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