CN116548980B

CN116548980B - Long-time-interval electrocardiograph classification method and system based on pulse neural network

Info

Publication number: CN116548980B
Application number: CN202310830798.1A
Authority: CN
Inventors: 李亚; 孙紫琦
Original assignee: Guangdong Polytechnic Normal University
Current assignee: Guangdong Polytechnic Normal University
Priority date: 2023-07-07
Filing date: 2023-07-07
Publication date: 2023-09-01
Anticipated expiration: 2043-07-07
Also published as: CN116548980A

Abstract

The application discloses a long-time-range electrocardio classification method and system based on a pulse neural network, which can be locally deployed on wearable equipment. Firstly, empirical mode decomposition is carried out on long-time electrocardiographic data, a picture is generated by utilizing a low-order eigenmode function, then the picture is input into a pulse neural network, and the training is directly carried out by using a substitution gradient method, so that the trained model can be locally deployed on wearable equipment to carry out arrhythmia classification. Compared with the existing neural network method, the power consumption of the method is greatly reduced, so that the method can be deployed locally on the wearable equipment; considerable accuracy can be achieved with lower delays relative to other impulse neural network approaches.

Description

Long-time-interval electrocardiograph classification method and system based on pulse neural network

Technical Field

The application relates to the technical field of electrocardiograph classification, in particular to a long-time-interval electrocardiograph classification method and system based on a pulse neural network.

Background

Cardiovascular disease is currently the leading cause of morbidity and mortality worldwide. Electrocardiogram monitoring is an effective means of achieving early detection of cardiovascular disease because it allows for non-invasive, almost non-radiative and immediate interpretation of the electrical state produced by the heart beat. However, since arrhythmia has sudden and sporadic characteristics, it is difficult to capture abnormal heartbeats in a short time. Thus, long-term automatic diagnosis of electrocardiographic signals is critical to improving the efficiency of routine cardiovascular disease care.

Currently, there are many high-performance automatic classification algorithms for electrocardiographic signals developed based on deep learning in the prior art, but almost all of them require transmitting signals to a connected smart phone or remote cloud server to perform a large number of calculations related to the ECG classification algorithm. In addition to privacy concerns, continuous cardiac monitoring with such solutions is also limited by the availability, speed and energy consumption of wireless connections. On the other hand, most of the algorithms classify single heart beats, the acquired electrocardiosignals are easily affected by noise in the wearable equipment scene, the data affected by the noise can be misjudged by the single heart beat classification algorithm, and the classification frequency is too high, so that the low-power-consumption implementation is not facilitated. Thus, there is a need to develop long-term electrocardiographic classification algorithms that can be deployed locally on a wearable device.

Disclosure of Invention

In order to solve the technical problems, the application provides a long-time-range electrocardio classification method and system based on a pulse neural network.

The first aspect of the application provides a long-term electrocardiograph classification method based on a pulse neural network, which comprises the following steps:

acquiring an electrocardiographic record, and extracting an electrocardiographic signal based on the electrocardiographic record;

dividing the electrocardiosignal into long-time-interval fragments, setting an overlapping window according to the data volume, adding labels to the long-time-interval fragments, and generating a training set and a testing set by utilizing the electrocardiosignal fragments containing the labels;

the electrocardiosignal is decomposed and processed through an empirical mode, and is decomposed into two signal components IMF1 and IMF2, and the IMF1 and IMF2 are combined into a two-dimensional image;

encoding the image combined by the IMF1 and the IMF2 into pulses, and constructing an electrocardiographic classification model by using a pulse neural network with a convolution structure for replacing gradient training;

and classifying the electrocardiographic sample to be classified by using the trained electrocardiographic classification model.

In this scheme, will the electrocardiosignal is cut apart into long-term section, sets up overlapping window according to the data volume, will long-term section adds the label, utilizes the electrocardiosignal section that contains the label to generate training set and test set, specifically does:

acquiring a preset number of electrocardiographic records through an MIT-BIH arrhythmia data set, extracting electrocardiographic signals, and screening normal heart beats, left branch block heart beats, right branch block heart beats and pacing heart beats from the electrocardiographic signals for marking;

dividing the electrocardiosignal into long-term fragments, and setting an overlapping window between adjacent long-term fragments to increase the sample data volume when the data volume corresponding to the long-term fragments is smaller than a preset threshold value;

obtaining a label of the long-time-range fragment by using a marked heart beat, and marking the long-time-range fragment by the label;

80% of the long-term fragments containing the labels are randomly selected as training sets, and 20% are selected as test sets.

In the scheme, the label of the long-term segment is obtained by using the marked heart beat, and the long-term segment is marked by the label, specifically:

the method comprises the steps of distributing labels for long-time-interval fragments according to marked heart beats, and marking the heart beats as normal fragments when all heart beats in one fragment are normal;

if the normal heart beat and the abnormal heart beat exist in one segment at the same time, the label of the abnormal heart beat is given to the segment;

if a plurality of types of anomalies exist in one segment, the segment is assigned with the anomaly label with the largest occurrence frequency, and if the occurrence frequency of the anomalies is the same, the segment is assigned with the anomaly label which occurs first;

if the label finally assigned to the segment is not within the classification range of the model, the segment is discarded.

In the scheme, the electrocardiosignals are processed through empirical mode decomposition, and the empirical mode decomposition comprises the following steps:

s1: determining an electrical centering signalDrawing a lower envelope of the electrocardiosignal based on the minimum point therein +.>Drawing an upper envelope line according to the maximum point>；

S2: averaging the upper and lower envelopes；

S3: solving for intermediate signals，/>Judging->Whether the condition of the eigenmode function IMF is satisfied, if so, it is defined as IMF1 (/ -)>) If not, then->Repeating the steps S1-S3 on the basis until the IMF1 meeting the condition is obtained;

s4: solving a secondary intermediate signal，/>Will be->Repeating steps S1-S3 as original signal until IMF2 (/ so) is obtained>）；

The combined signal components IMF1 and IMF2, i.e. when both functions take the same t, willAs an x-axis coordinate of the x-axis,as y-axis coordinates, points (x, y) are formed until all (x, y) within the range of t-values are taken, and the long-term segment is converted into a two-dimensional image by the above method.

In the scheme, an image of the combination of IMF1 and IMF2 is encoded into pulses, and a pulse neural network with a convolution structure is trained by using a substitution gradient to construct an electrocardiographic classification model, which comprises the following specific steps:

acquiring pixel values of a two-dimensional image generated by combining the IMF1 and the IMF2, normalizing the pixel values, setting a time step to be 100 by using a clock-driven simulation strategy, generating a random number between 0 and 1 in each time step, and generating a pulse in the current time step when the random number is larger than the normalized value;

using leaky integrate discharge LIF neurons as an activation function of a pulse neural network, wherein the neurons emit pulses when the membrane potential is greater than a preset membrane potential threshold;

pulse neural network is trained by using back propagation and gradient descent based on pulses by an alternative gradient method, the relationship between pulse and membrane potential being replaced by a fast sigmoid function during back propagation：

；

Where k is a constant value and where,represents the membrane potential, +.>Representing a membrane potential threshold;

by means ofCalculating the gradient of the loss relative to the weight in the impulse neural network, and obtaining the gradient of the loss relative to the weight in each time step>Adding the gradients of each time step to obtain a global gradient +.>，Indicating synaptic weight, +.>Representing time step->Is>。

In the scheme, the electrocardiographic sample to be classified is classified by using a trained electrocardiographic classification model, and the method specifically comprises the following steps:

the electrocardio classification model comprises a convolution layer, a pooling layer, an LIF neuron layer and a full connection layer, an electrocardio sample to be classified is input into the electrocardio classification model,

the method comprises the steps that feature extraction is carried out on two-dimensional image input by convolution layers, each convolution layer comprises a plurality of feature images, different features in the two-dimensional image are detected, neurons in the same feature images detect the same feature at different positions of the image, local features are extracted by shallow convolution layers, and global features are extracted by deep convolution layers;

the pooling layer obtains translational invariance by carrying out maximum pooling operation on adjacent neurons, and downsamples intermediate features to reduce the calculated amount;

and importing the pooled features into a full-connection layer, classifying by the full-connection layer by integrating the features extracted in the front, and acquiring a classification result of an electrocardio sample to be classified according to LIF neurons which send the most pulses.

The second aspect of the present application also provides a long-term electrocardiograph classification system based on a pulsed neural network, the system comprising: the long-time-course electrocardio classification method program based on the pulse neural network is executed by the processor and comprises the following steps:

The application solves the technical problems in the background technology and has the following beneficial effects:

because neurons in the impulse neural network only send impulses to subsequent neurons when the membrane potential is higher than a threshold value, the consumed energy is far lower than that of the traditional artificial neural network, and therefore, the electrocardiograph analysis model is different from the model which needs to be deployed on a cloud server in the past, and can be deployed on wearable equipment locally.

The application uses the alternative gradient training pulse neural network, and directly applies the back propagation and gradient descent algorithm to pulse-based training, the obtained pulse neural network model can obtain higher performance under the condition of low delay, and the reasoning power consumption is further reduced because the delay is reduced.

Unlike the current practice of most automatic electrocardiosignal analysis algorithms (classifying individual heart beats after slicing), the present application segments the electrocardiographic records into long-term segments and classifies them. Therefore, misjudgment of noise during single-beat classification can be avoided, the classification times can be greatly reduced, and classification is more efficient.

Drawings

FIG. 1 shows a flow chart of a long-term electrocardiographic classification method based on a pulsed neural network of the present application;

FIG. 2 is a flow chart of a method of empirical mode decomposition in accordance with the present application;

FIG. 3 shows a flow chart of a method for classifying an electrocardiographic sample to be classified using an electrocardiographic classification model;

fig. 4 shows a block diagram of a long-term electrocardiographic classification system based on a pulsed neural network according to the present application.

Detailed Description

In order that the above-recited objects, features and advantages of the present application will be more clearly understood, a more particular description of the application will be rendered by reference to the appended drawings and appended detailed description. It should be noted that, without conflict, the embodiments of the present application and features in the embodiments may be combined with each other.

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application, however, the present application may be practiced in other ways than those described herein, and therefore the scope of the present application is not limited to the specific embodiments disclosed below.

Fig. 1 shows a flow chart of a long-term electrocardiographic classification method based on a pulsed neural network according to the present application.

As shown in fig. 1, a first aspect of the present application provides a long-term electrocardiographic classification method based on a pulse neural network, including:

s102, acquiring an electrocardiographic record, and extracting electrocardiographic signals based on the electrocardiographic record;

s104, dividing the electrocardiosignal into long-term segments, setting an overlapping window according to the data volume, adding labels to the long-term segments, and generating a training set and a testing set by using the electrocardiosignal segments containing the labels;

s106, processing electrocardiosignals through empirical mode decomposition, decomposing the electrocardiosignals into two signal components IMF1 and IMF2, and combining the IMF1 and IMF2 into a two-dimensional image;

s108, coding the image combined by the IMF1 and the IMF2 into pulses, and constructing an electrocardiographic classification model by using a pulse neural network with a convolution structure for replacing gradient training;

s110, classifying the electrocardio samples to be classified by using the trained electrocardio classification model.

It should be noted that the assessment was performed using an MIT-BIH arrhythmia dataset comprising 48 30 minute Electrocardiographic (ECG) recordings involving 15 different types of beats. Acquiring a preset number of electrocardiographic records through an MIT-BIH arrhythmia data set, extracting electrocardiographic signals, and screening normal heart beats (N), left branch block heart beats (LBBB), right branch block heart beats (RBBB) and pacing heart beats (P) in the electrocardiographic signals for marking;

dividing the electrocardiosignal into long-term segments, and setting an overlapping window between adjacent long-term segments to increase the sample data volume when the data volume corresponding to the long-term segments is smaller than a preset threshold value, specifically, for one long-term ECG segment(X is a complete ECG record of a patient), define +.>Where the nth sample is indexed from 0, the larger O the smaller the overlap of adjacent segments, and T represents the length of each segment. Obtaining a label of the long-time-range fragment by using a marked heart beat, and marking the long-time-range fragment by the label; 80% of the long-term fragments containing the labels are randomly selected as training sets, and 20% are selected as test sets.

The method comprises the steps of obtaining a label of a long-time-course fragment by using a marked heart beat, and marking the long-time-course fragment by the label, wherein the method comprises the following steps:

Figure 2 shows a flow chart of a method of empirical mode decomposition in the present application.

According to the embodiment of the application, the electrocardiosignals are processed through empirical mode decomposition, and the empirical mode decomposition comprises the following steps:

S2: averaging the upper and lower envelopes；

s4: solving a secondary intermediate signal，/>Will be->Repeating steps S1-S3 as the original signal until IMF2 is acquired（/>）；

It should be noted that Empirical Mode Decomposition (EMD) can decompose the original signal into a plurality of eigenmode functions (IMFs). In the decomposition process, the IMF needs to satisfy the following two conditions: 1) In the whole data segment, the number of extreme points and the number of zero crossing points must be equal or the difference between the extreme points and the zero crossing points cannot exceed one at most; 2) At any time, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis.

According to the embodiment of the application, a two-dimensional image generated before is converted into a size of 32×32, and then the two-dimensional image is input into a pulse neural network through rate coding, specifically:

the leaky integrate-and-discharge LIF neurons are used as an activation function of a impulse neural network, in particular, LIF neurons satisfy the following relation:

；

wherein U [ t ]]And U [ t-1 ]]Respectively representing the membrane potential of the current time step and the previous timeThe membrane potential of the interval, X t]S [ t-1 ] is the input of the current time step of LIF neurons]For the pulse output of the previous time step, beta is the membrane potential decay rate,is synaptic weight, ++>Is a membrane potential threshold. The membrane potential of the previous time step is input after natural attenuation and weighting, and then the membrane potential of the current time step can be obtained by performing soft threshold resetting operation according to the pulse output of the previous time step.

The conditions for firing LIF neurons are as follows:

；

the neuron sends out pulses when the membrane potential is greater than a preset membrane potential threshold value;

the pulsed neural network is trained by an alternative gradient approach based on pulse use counter-propagation and gradient descent, resulting in better performance at low delays. Training the neural network using back propagation and gradient descent, requiring the loss and its gradient relative to the weight; the membrane potential of the output layer neurons was softmax activated and cross entropy loss was calculated. The gradient of loss relative to weight in the impulse neural network can be calculated by the chain derivative rule, wherein the partial derivative termDue to total of 0%Except when used) has a dead neuron problem that is addressed by using an alternative gradient. Specifically, the relation between pulse and membrane potential is replaced by fast sigmoid function during back propagation>：

；

Where k is a constant, in the example set to 25, the greater k the function is atThe steeper the vicinity is +.>Represents the membrane potential, +.>Representing a membrane potential threshold;

by means ofReplace->Calculating the gradient of the loss relative to the weight in the impulse neural network, and obtaining the gradient of the loss relative to the weight in each time step>Adding the gradients of each time step to obtain a global gradient，/>Indicating synaptic weight, +.>Representing time step->Is>。

Fig. 3 shows a flow chart of a method for classifying an electrocardiographic sample to be classified using an electrocardiographic classification model.

According to the embodiment of the application, the electrocardiographic sample to be classified is classified by using a trained electrocardiographic classification model, and the electrocardiographic sample to be classified is specifically:

s302, inputting an electrocardiographic sample to be classified into an electrocardiographic classification model, wherein the electrocardiographic classification model comprises a convolution layer, a pooling layer, an LIF neuron layer and a full connection layer;

s304, performing feature extraction on two-dimensional image input by using convolution layers, wherein each convolution layer comprises a plurality of feature images, different features in the two-dimensional image are detected, neurons in the same feature images detect the same features in different positions of the image, local features are extracted by using shallow convolution layers, and global features are extracted by using deep convolution layers;

s306, the pooling layer obtains translational invariance by carrying out maximum pooling operation on adjacent neurons, and downsamples intermediate features to reduce the calculated amount;

and S308, importing the pooled features into a full-connection layer, classifying by the full-connection layer by integrating the features extracted in the front, and acquiring a classification result of an electrocardio sample to be classified according to LIF neurons which send the most pulses.

The second aspect of the present application also provides a long-term electrocardiographic classification system 4 based on a pulsed neural network, the system comprising: a memory 41, and a processor 42, where the memory includes a long-term electrocardiographic classification method program based on a pulse neural network, and when the long-term electrocardiographic classification method program based on the pulse neural network is executed by the processor, the following steps are implemented:

S2: averaging the upper and lower envelopes；

The combined signal components IMF1 and IMF2, i.e. when both functions take the same t, willAs an x-axis coordinate of the x-axis,as y-axis coordinates, points (x, y) are formed until all (x, y) within the range of t values are taken,

by the method, the long-term fragments are converted into two-dimensional images.

；

wherein U [ t ]]And U [ t-1 ]]Representing the membrane potential of the current time step and the membrane potential of the last time step, X [ t ]]S [ t-1 ] is the input of the current time step of LIF neurons]For the pulse output of the previous time step, beta is the membrane potential decay rate,is synaptic weight, ++>Is a membrane potential threshold. The membrane potential of the previous time step is input after natural attenuation and weighting, and then the membrane potential of the current time step can be obtained by performing soft threshold resetting operation according to the pulse output of the previous time step.

The conditions for firing LIF neurons are as follows:

；

the impulse neural network is trained by an alternative gradient method based on impulse use of back propagation and gradient descent,

better performance is obtained with low latency. Training the neural network using back propagation and gradient descent, requiring the loss and its gradient relative to the weight; the membrane potential of the output layer neurons was softmax activated and cross entropy loss was calculated. The gradient of loss relative to weight in the impulse neural network can be calculated by the chain derivative rule, wherein the partial derivative termDue to total 0 ()>Except when used) has a dead neuron problem that is addressed by using an alternative gradient. Specifically, the relation between pulse and membrane potential is replaced by fast sigmoid function during back propagation>：

；

Where k is a constant, set to 25 in the example, the greater k is the function is atThe steeper the vicinity is +.>Represents the membrane potential, +.>Representing a membrane potential threshold;

the electrocardio classification model comprises a convolution layer, a pooling layer, an LIF neuron layer and a full connection layer, and an electrocardio sample to be classified is input into the electrocardio classification model;

The third aspect of the present application also provides a computer readable storage medium, wherein the computer readable storage medium includes a long-term electrocardiographic classification method program based on a pulse neural network, and when the long-term electrocardiographic classification method program based on the pulse neural network is executed by a processor, the steps of a long-term electrocardiographic classification method based on the pulse neural network are implemented.

In the several embodiments provided by the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The above described device embodiments are only illustrative, e.g. the division of the units is only one logical function division, and there may be other divisions in practice, such as: multiple units or components may be combined or may be integrated into another system, or some features may be omitted, or not performed. In addition, the various components shown or discussed may be coupled or directly coupled or communicatively coupled to each other via some interface, whether indirectly coupled or communicatively coupled to devices or units, whether electrically, mechanically, or otherwise.

The units described above as separate components may or may not be physically separate, and components shown as units may or may not be physical units; can be located in one place or distributed to a plurality of network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated in one processing unit, or each unit may be separately used as one unit, or two or more units may be integrated in one unit; the integrated units may be implemented in hardware or in hardware plus software functional units.

Those of ordinary skill in the art will appreciate that: all or part of the steps for implementing the above method embodiments may be implemented by hardware related to program instructions, and the foregoing program may be stored in a computer readable storage medium, where the program, when executed, performs steps including the above method embodiments; and the aforementioned storage medium includes: a mobile storage device, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk or an optical disk, or the like, which can store program codes.

Alternatively, the above-described integrated units of the present application may be stored in a computer-readable storage medium if implemented in the form of software functional modules and sold or used as separate products. Based on such understanding, the technical solutions of the embodiments of the present application may be embodied in essence or a part contributing to the prior art in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a removable storage device, ROM, RAM, magnetic or optical disk, or other medium capable of storing program code.

The foregoing is merely illustrative of the present application, and the present application is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. The long-time-interval electrocardio classification method based on the impulse neural network is characterized by comprising the following steps of:

classifying the electrocardiographic sample to be classified by using the trained electrocardiographic classification model;

dividing the electrocardiosignal into long-term segments, setting an overlapping window according to data volume, adding labels to the long-term segments, and generating a training set and a testing set by utilizing the electrocardiosignal segments containing the labels, wherein the training set and the testing set are specifically as follows:

randomly selecting 80% of the long-term fragments containing the labels as training sets and 20% as test sets;

the electrocardiosignals are processed through empirical mode decomposition, and the empirical mode decomposition comprises the following steps:

s1: determining an electrical centering signalDrawing a lower envelope of the electrocardiosignal according to the minimum value pointDrawing an upper envelope line according to the maximum point>；

S2: averaging the upper and lower envelopes；

The combined signal components IMF1 and IMF2, i.e. when both functions take the same t, willAs x-axis coordinates, +.>Forming points (x, y) as y-axis coordinates until all (x, y) in the value range of t are obtained, and converting the long-term fragments into two-dimensional images by the method;

the image of the combination of IMF1 and IMF2 is encoded into pulses, and a pulse neural network with a convolution structure is trained by using a substitution gradient to construct an electrocardiographic classification model, which comprises the following specific steps:

using the leaky integrate discharge neuron as an activation function of a pulse neural network, and sending out a pulse when the membrane potential is greater than a preset membrane potential threshold value;

；

by means ofCalculating the gradient of the loss relative to the weight in the impulse neural network, and obtaining the gradient of the loss relative to the weight in each time step>Adding the gradients of each time step to obtain a global gradient +.>，/>Indicating synaptic weight, +.>Representing time step->Is>；

Classifying the electrocardiographic sample to be classified by using the trained electrocardiographic classification model, specifically:

2. The long-term electrocardiographic classification method based on the pulse neural network according to claim 1, wherein the label of the long-term segment is obtained by using a marked heart beat, and the long-term segment is marked by the label, specifically:

3. A long-term electrocardiographic classification system based on a pulsed neural network, the system comprising: the long-time-course electrocardio classification method program based on the pulse neural network is executed by the processor and comprises the following steps:

S2: averaging the upper and lower envelopes；

；