CN111345814B - Analysis method, device, equipment and storage medium for electrocardiosignal center beat - Google Patents

Abstract

The embodiment of the present invention discloses an analysis method, device, equipment and storage medium of a heartbeat of an electrocardiographic signal. The method includes: preprocessing the collected ECG signals to obtain several sample signals with a set length; performing feature extraction on the sample signals to generate feature vectors of the sample signals on three scales; constructing a preset according to the feature vectors The length of the anchor point, calculate the reference point position of the heartbeat according to the anchor point; input the feature vector of each heartbeat within the corresponding width range to the type recognition module, and then go through the convolution layer, activation layer, and batch normalization of the type recognition module The calculation of the normalized layer and the linear fully connected layer outputs a five-dimensional vector, and the five-dimensional vector is used to represent the probability that the heartbeat corresponds to the five heartbeat types. This embodiment realizes automatic synchronous training of the modules, which saves training time.

Description

A method, device, equipment and storage medium for analyzing the heartbeat of an electrocardiogram signal

Technical Field

The embodiments of the present invention relate to the field of signal processing technology, and in particular to a method, device, equipment and storage medium for analyzing the heartbeat of an electrocardiogram signal.

Background Art

At present, the number of cardiovascular disease patients in my country is about 260 million, and cardiovascular mortality ranks first among urban and rural residents' disease deaths, and the number of patients is still increasing. ECG signals reflect the electrophysiological process of cardiac activity. Since ECG-based inspections are low-cost and easy to use, they are widely used in the inspection and diagnosis of cardiovascular diseases. Figure 1 shows a typical ECG waveform. A normal ECG signal is generally composed of a P wave, a QRS complex wave and a T wave, and sometimes a U wave. The P wave represents the electrical activity of atrial contraction, and the QRS wave and T represent the electrical activity of ventricular contraction and relaxation, respectively.

ECG analysis algorithms are an important component of diagnostic and analysis equipment such as electrocardiographs and electrocardiograph monitors. ECG analysis algorithms can detect a variety of diseases and issue alarms in a timely manner by measuring and classifying ECG signals. In actual use, ECG analysis usually first performs heartbeat position detection (also known as QRS detection), then performs reference point detection (such as QRS wave starting point, end point, etc.) and type identification (normal, supraventricular abnormality, ventricular abnormality, etc.) on the detected heartbeat position, and finally outputs the analysis results. Once the heartbeat position output by the heartbeat position detection module changes, it will seriously affect the performance of subsequent reference point detection and heartbeat type identification. Existing algorithms usually design different functional modules for different steps and test them separately. Its advantage is that the development process of a single module is simple. After specifying the input and output methods of each module, the performance of each module can be tested by marking information, but the disadvantage is that the joint debugging process between modules is cumbersome.

Summary of the invention

The present invention provides an analysis method, device, equipment and storage medium for determining the heartbeat in an electrocardiogram signal, so as to solve the technical problem of the cumbersome joint debugging process between multiple modules based on step refinement in the prior art.

In a first aspect, an embodiment of the present invention provides a method for analyzing a heartbeat in an electrocardiogram signal, comprising:

Preprocessing the collected ECG signals to obtain a number of sample signals of set lengths;

Performing feature extraction on the sample signal to generate feature vectors of the sample signal at three scales, wherein the lengths of the feature vectors at the three scales are respectively one-half, one-quarter, and one-eighth of the set length;

Constructing an anchor point of a preset length according to the feature vector corresponding to the sample signal, and calculating the reference point position of the heartbeat according to the anchor point;

The feature vector of each heartbeat within the corresponding reference point position range is input into the type recognition module, and is calculated in sequence by the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module to output a five-dimensional vector, which is used to characterize the probability that the heartbeat belongs to the five heartbeat types.

The step of constructing an anchor point of a preset length according to the feature vector corresponding to the sample signal and calculating the reference point position of the heartbeat according to the anchor point includes:

For each feature vector corresponding to the scale, an anchor point with a length of 9 sample points is constructed with each sample point on the feature vector as the center, and the score of the anchor point is the value of the central sample point of the anchor point;

Input the feature vector corresponding to the anchor point into the heartbeat position detection and measurement module to obtain the score of the heartbeat probability, the change from the anchor point to the bounding box, and the duration ratio of each band;

Sort all the anchor points by scores from large to small to generate a checklist;

The anchor points in the inspection list are sequentially screened for candidate regions, and the screened anchor points are added to the candidate list as candidate regions. The anchor points and other anchor points within 0.2 seconds of the anchor points are deleted from the inspection list, and the screened anchor points are the anchor points in the inspection list with the highest scores, starting points greater than 0, and end points less than the length of the corresponding feature vector;

Calculate the midpoint and width of the heartbeat bounding box according to the midpoint and width of the anchor point in the candidate list and the variation, and calculate the start point and end point of the heartbeat bounding box according to the midpoint and width of the heartbeat bounding box;

The positions of five reference points of the heartbeat are calculated according to the starting point, width and time ratio of the heartbeat boundary box.

The step of extracting features from the sample signal to generate feature vectors of the sample signal at three scales, wherein the lengths of the feature vectors at the three scales are respectively one-half, one-quarter, and one-eighth of the set length, includes:

The sample signal is input into an initial feature extraction module to obtain an initial feature extraction result, and the feature extraction modules corresponding to the three scales perform feature extraction on the initial feature extraction result according to the corresponding scale levels, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length;

The feature extraction result corresponding to each scale is output to the corresponding convolution layer and the feature extraction results at adjacent scales are integrated to generate the feature vectors of the sample signal at three scales.

The total loss of the type recognition module during training is calculated by the following formula:

loss=det_loss+bbx_loss+fiducial_loss+cls_loss

Among them, loss is the total loss, det_loss is the heartbeat position detection loss, bbx_loss is the heartbeat bounding box range detection loss, fiducial_loss is the reference point position detection loss and cls_loss is the heartbeat type classification loss.

The heartbeat position detection loss is calculated by the following formula:

Wherein, score _i represents the output score of the i-th anchor point in the heart beat position detection and measurement module (0<score _i <1), _fi represents the type of the anchor point, and fi = 1 when the absolute value of the difference between the anchor point _position and the true heart beat position is less than 0.15 seconds, otherwise _fi = 0.

The heartbeat bounding box range detection loss bbx_loss is calculated by the following formula:

The fiducial_loss is calculated by the following formula:

和

第j个锚点对应的P波起点、P波终点、QRS波起点、QRS波终点和T波终点分别为

和

Among them, the starting point and end point of the heartbeat corresponding to the jth anchor point are

and

The P wave start point, P wave end point, QRS wave start point, QRS wave end point, and T wave end point corresponding to the jth anchor point are

and

Among them, the type classification loss is calculated by the following formula:

表示第j个锚点属于第k类型心拍的概率

表示该锚点是否属于第k类型，如果属于该类型则

反之

in

represents the probability that the jth anchor point belongs to the kth type of heartbeat

Indicates whether the anchor point belongs to the kth type. If it does, then

on the contrary

In a second aspect, an embodiment of the present invention further provides a device for analyzing a heartbeat in an electrocardiogram signal, comprising:

A preprocessing unit, used to preprocess the collected ECG signals to obtain a number of sample signals of set lengths;

A feature extraction unit, used to extract features from the sample signal to generate feature vectors of the sample signal at three scales, wherein the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length;

A reference point detection unit, configured to construct an anchor point of a preset length according to a feature vector corresponding to the sample signal, and calculate a reference point position of a heartbeat according to the anchor point;

The heartbeat classification unit is used to input the feature vector of each heartbeat within the corresponding reference point position range into the type recognition module, and sequentially pass the calculations of the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module to output a five-dimensional vector, wherein the five-dimensional vector is used to characterize the probability that the heartbeat corresponds to the five heartbeat types.

Wherein, the reference point detection unit includes:

An anchor point calculation module, used for constructing an anchor point with a length of 9 sample points, with each sample point on the feature vector as the center, for each feature vector corresponding to the scale, wherein the score of the anchor point is the value of the central sample point of the anchor point;

A data output module, used to input the feature vector corresponding to the anchor point into the heartbeat position detection and measurement module to obtain the score of the heartbeat probability, the change from the anchor point to the bounding box and the duration ratio of each band;

An anchor point sorting module, used to sort all the anchor points from large to small according to the scores, and generate a check list;

A candidate area screening module is used to sequentially screen the anchor points in the inspection list for candidate areas, add the screened anchor points each time as candidate areas to the candidate list, delete the anchor point and other anchor points within 0.2 seconds of the anchor point in the inspection list, and the screened anchor point each time is the anchor point in the inspection list with the highest score, a starting point greater than 0, and an end point less than the length of the corresponding feature vector;

A range calculation module, used to calculate the midpoint and width of the heartbeat bounding box according to the midpoint and width of the anchor point in the candidate list and the variation, and calculate the start point and end point of the heartbeat bounding box according to the midpoint and width of the heartbeat bounding box;

The reference point calculation module is used to calculate the positions of five reference points of the heartbeat according to the starting point, width and time ratio of the heartbeat boundary box.

Wherein, the feature extraction unit comprises:

A feature extraction module, used for inputting the sample signal into an initial feature extraction module to obtain an initial feature extraction result, wherein the feature extraction modules corresponding to the three scales perform feature extraction on the initial feature extraction result according to the corresponding scale levels, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length;

The feature vector generation module is used to output the feature extraction result corresponding to each scale to the corresponding convolution layer and integrate the feature extraction results at adjacent scales to generate the feature vectors of the sample signal at three scales.

The total loss of the type recognition module during training is calculated by the following formula:

loss＝det_loss+bbx_loss+fiducial_loss+cls_loss

Among them, loss is the total loss, det_loss is the heartbeat position detection loss, bbx_loss is the heartbeat bounding box range detection loss, fiducial_loss is the reference point position detection loss and cls_loss is the heartbeat type classification loss.

The heartbeat position detection loss is calculated by the following formula:

Wherein, score _i represents the output score of the i-th anchor point in the heart beat position detection and measurement module (0<score _i <1), _fi represents the type of the anchor point, and fi = 1 when the absolute value of the difference between the anchor point _position and the true heart beat position is less than 0.15 seconds, otherwise _fi = 0.

The heartbeat bounding box range detection loss bbx_loss is calculated by the following formula:

The fiducial_loss is calculated by the following formula:

和

第j个锚点对应的P波起点、P波终点、QRS波起点、QRS波终点和T波终点分别为

和

Among them, the starting point and end point of the heartbeat corresponding to the jth anchor point are

and

The P wave start point, P wave end point, QRS wave start point, QRS wave end point, and T wave end point corresponding to the jth anchor point are

and

Among them, the type classification loss is calculated by the following formula:

表示第j个锚点属于第k类型心拍的概率

表示该锚点是否属于第k类型，如果属于该类型则

反之

in,

represents the probability that the jth anchor point belongs to the kth type of heartbeat

Indicates whether the anchor point belongs to the kth type. If it does, then

on the contrary

In a third aspect, an embodiment of the present invention further provides a device, the device comprising:

one or more processors;

a memory for storing one or more programs,

When the one or more programs are executed by the one or more processors, the one or more processors implement the method for analyzing the heartbeat in the electrocardiogram signal as described in any one of the first aspects.

In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for analyzing the heartbeat in an electrocardiogram signal as described in any one of the first aspects.

The above-mentioned heartbeat analysis method, device, equipment and storage medium of the ECG signal preprocess the collected ECG signal to obtain several segments of sample signals of set length; extract features from the sample signals to generate feature vectors of the sample signals at three scales, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length; construct anchor points of preset lengths according to the feature vectors corresponding to the sample signals, and calculate the reference point position of the heartbeat according to the anchor points; input the feature vectors of each heartbeat within the corresponding reference point position range into the type recognition module, and sequentially pass through the convolution layer, activation layer, batch normalization layer and linear full connection layer of the type recognition module to output a five-dimensional vector, which is used to characterize the probability that the heartbeat corresponds to the five heartbeat types. This embodiment is based on a deep learning method, and outputs the type of each heartbeat reference point position through an end-to-end fully automatic analysis process, so as to realize automatic synchronous training of all modules, simplify the training process while maintaining the comprehensive performance of each link, and save training time.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the structure of an electrocardiogram signal;

FIG2 is a flow chart of a method for analyzing the heartbeat of an electrocardiogram signal provided in Embodiment 1 of the present invention;

3 to 6 are schematic diagrams showing changes in an electrocardiogram signal processing process provided by the first embodiment of the present invention;

7 is a schematic diagram of data flow in the process of processing ECG signals provided in the first embodiment of the present invention;

FIG8 is a flow chart of feature extraction of a method for analyzing the heartbeat of an electrocardiogram signal provided in Embodiment 2 of the present invention;

FIG9 is a schematic diagram of data flow during feature extraction in Embodiment 2 of the present invention;

10-13 are schematic diagrams of the data processing flow of the feature extraction module in the second embodiment of the present invention;

14 is a flowchart of reference point position calculation in Embodiment 2 of the present invention;

15 is a schematic diagram of a data processing flow for calculating a reference point position in a second embodiment of the present invention;

FIG16 is a schematic diagram of a data processing flow for center beat classification according to Embodiment 2 of the present invention;

FIG17 is a schematic diagram of the structure of a device for analyzing the heartbeat of an electrocardiogram signal provided in Embodiment 3 of the present invention;

FIG18 is a schematic diagram of the structure of a device provided in Embodiment 4 of the present invention.

DETAILED DESCRIPTION

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only the parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

Embodiment 1

FIG2 is a flow chart of a method for analyzing the heartbeat of an ECG signal provided in Embodiment 1 of the present invention. The method for analyzing the heartbeat of an ECG signal provided in the embodiment can be performed by a device for detecting and analyzing the QRS wave in the ECG signal, and the device for detecting the QRS wave in the ECG signal can be implemented by software and/or hardware, and the device for analyzing the heartbeat of the ECG signal can be composed of two or more physical entities, or can be composed of one physical entity. For example, the device for analyzing the heartbeat of an ECG signal can be a mobile phone, an industrial control computer, etc.

As shown in FIG2 , the method for analyzing the heartbeat of an electrocardiogram signal provided in the first embodiment includes the following steps:

Step S110: pre-processing the collected ECG signal to obtain a number of sample signals of set length.

Preprocessing mainly includes resampling, filtering and other waveform adjustments to the signal. Specifically, for the most original ECG signal, the ECG signal is first resampled to 256Hz (fs=256), and then bandpass filtered using a filter with a passband range of 0.5Hz to 40Hz. For an individual's ECG signal, the length of the ECG signal segment used for analysis is generally about 10 seconds, that is, for an individual's ECG signal, there are 2560 sample points in a 10-second ECG signal. In this embodiment, the resampled ECG signal is rounded to s _i , i=1,…,n (n=2560). Figures 3 and 4 show the ECG signals before and after filtering. By comparing Figures 3 and 4, it can be found that the bandpass filter filters out the "burr" part of the signal.

Step S120: extracting features from the sample signal to generate feature vectors of the sample signal at three scales, wherein the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length.

The lengths of the feature vectors at the three scales are half, one quarter, and one eighth of the set length, respectively. For simplicity, they are defined as the first, second, and third layers, respectively. The first layer is used for the detection of pacemaker signals, the second layer is used for the detection of normal width heartbeats, and the third layer is used for the detection of wide and deformed heartbeats, such as ventricular premature beats. The collection of these three types constitutes all heartbeats.

Step S130: constructing anchor points of a preset length according to the feature vector corresponding to the sample signal, and calculating the reference point position of the heartbeat according to the anchor points.

According to the value of the anchor point of the set length, each sample signal can be scored and sorted, and the candidate area of the suspected heartbeat can be roughly identified first, as shown in Figure 5, in which the positions of several suspected QRS waves are first identified and marked with "×". In addition to the part with a larger peak value, there are also quite a few ECG waves with smaller peak values that are marked.

After the suspected QRS wave is initially confirmed, the corresponding candidate area and feature vector are input into the candidate area detection module, which can further accurately determine whether the signal in the candidate area is a QRS wave. Based on the candidate area shown in Figure 5, the signal marked with "×" in Figure 6 is determined to be a QRS wave. In Figures 3 to 6, the wave corresponding to "·" is the actual QRS wave. It can be clearly seen that after the above processing, all QRS waves are accurately determined in Figure 6.

During the heartbeat detection process, the candidate area detection module will also more accurately determine the width range corresponding to each heartbeat. The width range is expressed as data through reference points, and the heartbeat type is subsequently determined based on the corresponding feature vector within the width range.

Step S140: Input the feature vector of each heartbeat within the corresponding reference point position range into the type recognition module, and sequentially pass through the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module to output a five-dimensional vector, which is used to characterize the probability that the heartbeat corresponds to one of the five heartbeat types.

The feature vectors within the reference point position range corresponding to each heartbeat are input into the type recognition module. After calculations by the convolution layer, activation layer, batch normalization layer and linear fully connected layer in sequence, each heartbeat can obtain a five-dimensional vector result. The five-dimensional vector indicates the probability that the heartbeat may correspond to one of the five heartbeat types. Based on the five-dimensional vector, the type with the highest probability can be used as the heartbeat type of the heartbeat. If the probabilities corresponding to multiple types are close, the heartbeat can be marked separately for manual judgment.

FIG7 further illustrates the above data flow process, starting with obtaining the original ECG signal, which is then preprocessed to obtain a preliminary signal for feature extraction. As for the result of feature extraction, the heartbeat and the corresponding width detection are combined to achieve heartbeat classification by combining the results of heartbeat detection, width measurement and adjustment extraction.

In general, by preprocessing the collected ECG signals, several segments of sample signals of set lengths are obtained; feature extraction is performed on the sample signals to generate feature vectors of the sample signals at three scales, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length; anchor points of preset lengths are constructed according to the feature vectors corresponding to the sample signals, and the reference point positions of the heartbeats are calculated according to the anchor points; the feature vectors of each heartbeat within the corresponding reference point position range of the calculated heartbeats are input into the type recognition module, and the convolution layer, activation layer, batch normalization layer and linear full connection layer of the type recognition module are calculated in turn to output a five-dimensional vector, which is used to characterize the probability that the heartbeats correspond to the five types of heartbeats. This embodiment is based on a deep learning method, and outputs the type of each heartbeat reference point position through an end-to-end fully automatic analysis process, so as to realize automatic synchronous training of all modules, simplify the training process while maintaining the comprehensive performance of each link, and save training time.

Embodiment 2

This embodiment is concretized on the basis of the above-mentioned embodiment, especially the concretization of step S120 and step S130. It should be noted that in this embodiment, the concretization of step S120 and step S130 is presented at the same time, which does not mean that the two must be implemented at the same time, but is an integrated processing for describing the scheme. In the actual processing process, the concretization of step S120 and step S130 can exist as an independent implementation method, or it can be an integration of the two. This embodiment makes a complete description of the entire signal processing process at a more detailed level. On the whole, it includes step S110, the related steps in Figures 8 and 14, and step S140.

Step S110: pre-processing the collected ECG signal to obtain a number of sample signals of set length.

Step S121: Input the sample signal into the initial feature extraction module to obtain an initial feature extraction result. The feature extraction modules corresponding to the three scales perform feature extraction on the initial feature extraction result according to the corresponding scale levels. The lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length.

Step S122: outputting the feature extraction result corresponding to each scale to the corresponding convolution layer and integrating the feature extraction results at adjacent scales to generate feature vectors of the sample signal at three scales.

The data processing process in step S121 and step S122 can refer to FIG9, where C0, C1, C2 and C3 represent the initial feature extraction module and the feature extraction modules corresponding to the three scales, respectively. After the sample signal enters C0 (initial feature extraction module), it flows in the direction shown in FIG9, and is processed accordingly in the flow process as shown in FIG9. Taking the first layer as an example, after the sample signal enters C0, it follows the data processing flow shown in FIG10 (including convolution layer, batch normalization layer, activation layer and pooling layer); the extraction result of C0 is output to C1 (first layer extraction module), and C1 obtains the corresponding processing result according to the data processing flow shown in FIG11 (including two convolution layers, two batch normalization layers and two activation layers), and the processing result is output to C2 and the convolution layer corresponding to C1, and then after two convolutions shown in FIG9 and one upsampling data synthesis, the feature vector _f1 corresponding to the first layer is obtained. The feature extraction of the second and third layers refers to FIG12 and FIG13 respectively. The data processing process of the second and third layers is similar, except that some of the processing parameters involved are different. For example, the number of channels (c), convolution kernel size (k), step size (s) and padding size (p) of the convolution layer C0 are 64, 7, 1 and 3 respectively. The parameter combinations of the two convolution layers of C1, C2 and C3 may be different. For details, please refer to Figures 11-13. In the feature vector finally extracted, the jth sample point (i.e. position j) on the feature vector of the i-th scale (i-th layer) is recorded as

Step S131: for each feature vector corresponding to the scale, an anchor point with a length of 9 sample points is constructed with each sample point on the feature vector as the center, and the score of the anchor point is the value of the central sample point of the anchor point.

Step S132: input the feature vector corresponding to the anchor point into the heartbeat position detection and measurement module to obtain the score of the heartbeat probability, the change from the anchor point to the bounding box, and the duration ratio of each band.

Step S133: sort all the anchor points according to the scores from large to small to generate a check list.

Step S134: Screen the candidate areas for the anchor points in the inspection list in turn, add the screened anchor point each time as a candidate area to the candidate list, delete the anchor point and other anchor points within 0.2 seconds from the anchor point in the candidate list, and the anchor point screened out each time is the anchor point with the highest score in the inspection list, a starting point greater than 0 and an end point less than the length of the corresponding feature vector, and the score of the anchor point is used as the score of the corresponding candidate area.

为中心，构建长度为9个样本点的锚点

例如样本点

所对应的锚点为

锚点的分数为

(锚点中心处样本)的值。(第1层特征向量上的锚点对应于原始信号上的17个样本点长度，即0.067秒的候选区域，用于起搏器信号的检测；第2层特征向量上的锚点对应于原始信号上的33个样本点长度，即0.13秒的候选区域，用于正常宽度心拍的检测；第3层特征向量上的锚点对应于原始信号上的65个样本点长度，即0.25秒的候选区域，用于宽大畸形的心拍的检测)b.将全部锚点按分数由大到小进行排序，生成检查列表。c.在检查列表中选择分数最高，且锚点起点大于0，锚点终点小于特征图长度(l_i)的锚点作为候选区域加入候选列表，并在检查列表中删除该锚点，以及与该锚点相距离0.2秒(正常情况下两次心跳的最短间隔为0.2秒)之内的所有其他锚点。d.重复步骤c直到检查列表为空，将候选列表中的锚点作为候选区域的位置，锚点的值作为候选区域的分数并输出。The determination of the candidate region in steps S131 to S133 can be specifically described by the following mathematical language: a. Taking each sample point on the feature vector

As the center, construct an anchor point with a length of 9 sample points

For example, sample points

The corresponding anchor point is

The score of the anchor point is

(sample at the center of the anchor point). (The anchor points on the first-layer feature vector correspond to 17 sample points on the original signal, i.e., a candidate region of 0.067 seconds, which is used for the detection of pacemaker signals; the anchor points on the second-layer feature vector correspond to 33 sample points on the original signal, i.e., a candidate region of 0.13 seconds, which is used for the detection of normal width heartbeats; the anchor points on the third-layer feature vector correspond to 65 sample points on the original signal, i.e., a candidate region of 0.25 seconds, which is used for the detection of wide and deformed heartbeats) b. Sort all anchor points by scores from large to small to generate a check list. c. Select the anchor point with the highest score in the check list, whose starting point is greater than 0 and whose end point is less than the feature map length (l _i ), as a candidate region and add it to the candidate list, and delete the anchor point in the check list, as well as all other anchor points within 0.2 seconds (the shortest interval between two heartbeats is 0.2 seconds under normal circumstances) from the anchor point. d. Repeat step c until the check list is empty, take the anchor point in the candidate list as the position of the candidate region, and the value of the anchor point as the score of the candidate region and output it.

Step S135: Calculate the midpoint and width of the heartbeat bounding box according to the midpoint and width of the anchor points in the candidate list and the variation, and calculate the start point and end point of the heartbeat bounding box according to the midpoint and width of the heartbeat bounding box.

Step S136: Calculate the positions of five reference points of the heartbeat according to the starting point, width and time ratio of the heartbeat boundary box.

The overall calculation process of the heartbeat based on the feature vector is as follows: first, anchor points are established on the feature vector, and then 1) the probability score of each anchor point belonging to the heartbeat (score), 2) the change from the anchor point to the heartbeat bounding box (delta), and 3) the ratio of the duration of the four parts within the heartbeat bounding box (duration rate): P wave start to P wave end ( _Pdur ), P wave end to QRS wave start ( _PRseg ), QRS wave start to QRS wave end ( _QRSint ), QRS wave end to T wave end ( _STint ). Then, the beat location is obtained according to the probability score of each anchor point, and the range of the beat bounding box is obtained according to the change of the bounding box. Finally, the five fiducial points of the current beat are obtained by using the range of the beat bounding box and the ratio of the duration of each part: P wave start (P _onset ), P wave end (P _offset ), QRS wave start (QRS _onset ), QRS wave end (QRS _offset ) and T wave end (T _offset ). The specific process is as follows:

为中心，构建长度为9个样本点的锚点

例如样本点

所对应的锚点为

锚点的分数为

(锚点中心处样本)的值。(第1层特征向量上的锚点对应于原始信号上的17个样本点长度，即0.067秒的候选区域，用于起搏器信号的分析；第2层特征向量上的锚点对应于原始信号上的33个样本点长度，即0.13秒的候选区域，用于正常心拍和室上性异常心拍的分析；第3层特征向量上的锚点对应于原始信号上的65个样本点长度，即0.25秒的候选区域，用于室性异常心拍和融合心拍的分析)。a. Take each sample point on the feature vector

As the center, construct an anchor point with a length of 9 sample points

For example, sample points

The corresponding anchor point is

The score of the anchor point is

(Sample at the center of the anchor point). (The anchor points on the first-layer feature vector correspond to 17 sample points on the original signal, i.e., a candidate region of 0.067 seconds, which is used for the analysis of pacemaker signals; the anchor points on the second-layer feature vector correspond to 33 sample points on the original signal, i.e., a candidate region of 0.13 seconds, which is used for the analysis of normal heartbeats and supraventricular abnormal heartbeats; the anchor points on the third-layer feature vector correspond to 65 sample points on the original signal, i.e., a candidate region of 0.25 seconds, which is used for the analysis of ventricular abnormal heartbeats and fusion heartbeats).

b. Input the feature vector corresponding to each anchor point into the heartbeat position detection and measurement module, and obtain the heartbeat probability score, the change value from the anchor point to the bounding box, and the ratio of the duration of each band.

c. For the predicted heartbeat probability scores, first sort the probability scores from large to small to generate a check list. Select the anchor point with the highest score in the check list as the candidate area and add it to the candidate list. Delete the anchor point and all other anchor points within 0.2 seconds (the shortest interval between two heartbeats is 0.2 seconds) from the anchor point in the check list. Repeat the above steps until the check list is empty or the remaining anchor point values are all less than 0.5. Finally, the midpoint of the anchor point in the candidate list is used as the heartbeat position.

和宽度

最后求出预测的心拍边界框的起点

和终点

位置，如公式(2)所示。d. For the change from the anchor point to the heartbeat bounding box, first calculate the midpoint of the predicted heartbeat bounding box based on the midpoint (anchor _center ) and width (anchor _width ) of the anchor point and its change (delta _center and delta _width )

and width

Finally, find the starting point of the predicted heartbeat bounding box

and end point

Position, as shown in formula (2).

e. Based on the starting point and width of the heartbeat bounding box and the ratio of the duration of the four parts within the bounding box, use formula (3) to predict the positions of the five reference points:

Step S140: Input the feature vector of each heartbeat within the corresponding width range into the type recognition module, and sequentially pass through the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module to output a five-dimensional vector, which is used to characterize the probability that the heartbeat corresponds to the five heartbeat types.

For the heartbeat position detected by the heartbeat position detection and measurement module, the feature vector of the corresponding position on the feature extraction module is first extracted, and then sent to the heartbeat type recognition module for classification. After calculation by the convolution layer, activation layer, batch normalization layer and linear full connection layer, a 5-dimensional vector is output, which respectively represents the probability that the heartbeat belongs to 5 types: sinus beat, supraventricular abnormal heartbeat, ventricular abnormal heartbeat, fusion heartbeat and other types of heartbeat. Finally, the type with the highest probability is taken as the type of heartbeat.

Embodiment 3

FIG17 is a schematic diagram of the structure of a device for analyzing the heartbeat of an electrocardiogram signal provided by Embodiment 3 of the present invention. Referring to FIG17 , the device for analyzing the heartbeat of an electrocardiogram signal includes: a preprocessing unit 310 , a feature extraction unit 320 , a reference point detection unit 330 and a heartbeat classification unit 340 .

Among them, the preprocessing unit 310 is used to preprocess the collected ECG signal to obtain several segments of sample signals of set length; the feature extraction unit 320 is used to extract features from the sample signal to generate feature vectors of the sample signal at three scales, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length; the reference point detection unit 330 is used to construct anchor points of preset length according to the feature vectors corresponding to the sample signal, and calculate the reference point position of the heartbeat according to the anchor points; the heartbeat classification unit 340 is used to input the feature vector of each heartbeat within the corresponding reference point position range into the type recognition module, and output a five-dimensional vector after calculation by the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module in sequence, and the five-dimensional vector is used to characterize the probability that the heartbeat corresponds to five heartbeat types.

Based on the above embodiment, the reference point detection unit 330 includes:

An anchor point calculation module, used for constructing an anchor point with a length of 9 sample points, with each sample point on the feature vector as the center, for each feature vector corresponding to the scale, wherein the score of the anchor point is the value of the central sample point of the anchor point;

A data output module, used to input the feature vector corresponding to the anchor point into the heartbeat position detection and measurement module to obtain the score of the heartbeat probability, the change from the anchor point to the bounding box and the duration ratio of each band;

An anchor point sorting module, used to sort all the anchor points from large to small according to the scores, and generate a check list;

A candidate area screening module is used to sequentially screen the anchor points in the inspection list for candidate areas, add the screened anchor points each time as candidate areas to the candidate list, delete the anchor point and other anchor points within 0.2 seconds of the anchor point in the inspection list, and the screened anchor point each time is the anchor point in the inspection list with the highest score, a starting point greater than 0, and an end point less than the length of the corresponding feature vector;

A range calculation module, used to calculate the midpoint and width of the heartbeat bounding box according to the midpoint and width of the anchor point in the candidate list and the variation, and calculate the start point and end point of the heartbeat bounding box according to the midpoint and width of the heartbeat bounding box;

The reference point calculation module is used to calculate the positions of five reference points of the heartbeat according to the starting point, width and time ratio of the heartbeat boundary box.

Based on the above embodiment, the feature extraction unit 320 includes:

A feature extraction module, used for inputting the sample signal into an initial feature extraction module to obtain an initial feature extraction result, wherein the feature extraction modules corresponding to the three scales perform feature extraction on the initial feature extraction result according to the corresponding scale levels, and the lengths of the feature vectors at the three scales are respectively one-half, one-quarter and one-eighth of the set length;

The feature vector generation module is used to output the feature extraction result corresponding to each scale to the corresponding convolution layer and integrate the feature extraction results at adjacent scales to generate the feature vectors of the sample signal at three scales.

Based on the above embodiment, the total loss of the type recognition module during the training process is calculated by the following formula:

loss=det_loss+bbx_loss+fiducial_loss+cls_loss

Among them, loss is the total loss, det_loss is the heartbeat position detection loss, bbx_loss is the heartbeat bounding box range detection loss, fiducial_loss is the reference point position detection loss and cls_loss is the heartbeat type classification loss.

Based on the above embodiment, the heartbeat position detection loss is calculated by the following formula:

Wherein, score _i represents the output score of the i-th anchor point in the heart beat position detection and measurement module (0<score _i <1), _fi represents the type of the anchor point, and fi = 1 when the absolute value of the difference between the anchor point _position and the true heart beat position is less than 0.15 seconds, otherwise _fi = 0.

Based on the above embodiment, the heartbeat bounding box range detection loss bbx_loss is calculated by the following formula:

The fiducial_loss is calculated by the following formula:

和

第j个锚点对应的P波起点、P波终点、QRS波起点、QRS波终点和T波终点分别为

和

Among them, the starting point and end point of the heartbeat corresponding to the jth anchor point are

and

The P wave start point, P wave end point, QRS wave start point, QRS wave end point, and T wave end point corresponding to the jth anchor point are

and

Based on the above embodiment, the type classification loss is calculated by the following formula:

表示第j个锚点属于第k类型心拍的概率

表示该锚点是否属于第k类型，如果属于该类型则

反之

in,

represents the probability that the jth anchor point belongs to the kth type of heartbeat

Indicates whether the anchor point belongs to the kth type. If it does, then

on the contrary

The device for analyzing the heartbeat of an electrocardiogram signal provided in an embodiment of the present invention is included in a device for detecting the heartbeat of an electrocardiogram signal, and can be used to execute the method for analyzing the heartbeat of an electrocardiogram signal provided in any of the above embodiments, and has corresponding functions and beneficial effects.

Embodiment 4

FIG18 is a schematic diagram of the structure of a device provided in the fourth embodiment of the present invention. The device can be various electrocardiographs and electrocardiograph monitors in specific product presentation. More specifically, it can be a device that applies the method for analyzing the heartbeat of an electrocardiograph signal described in the aforementioned embodiment. As shown in FIG18 , the device includes a processor 410, a memory 420, an input device 430, an output device 440, and a communication device 450; the number of processors 410 in the device can be one or more, and FIG18 takes one processor 410 as an example; the processor 410, the memory 420, the input device 430, the output device 440, and the communication device 450 in the device for detecting the heartbeat of an electrocardiograph signal can be connected via a bus or other means, and FIG18 takes the connection via a bus as an example.

The memory 420, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions/modules corresponding to the method for analyzing the heartbeat of an electrocardiogram signal in an embodiment of the present invention (for example, the preprocessing unit 310, the feature extraction unit 320, the reference point detection unit 330 and the heartbeat classification unit 340 in the device for analyzing the heartbeat of an electrocardiogram signal). The processor 410 executes various functional applications and data processing of the device by running the software programs, instructions and modules stored in the memory 420, that is, realizes the above-mentioned method for analyzing the heartbeat of an electrocardiogram signal.

The memory 420 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the device, etc. In addition, the memory 420 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 420 may further include a memory remotely arranged relative to the processor 410, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The input device 430 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the device. The output device 440 may include a display device such as a display screen. The communication device 450 is used to communicate data with the image capture module.

The above-mentioned device includes a device for analyzing the heartbeat of an electrocardiogram signal, which can be used to execute a method for analyzing the heartbeat of any electrocardiogram signal and has corresponding functions and beneficial effects.

Embodiment 5

An embodiment of the present invention further provides a storage medium comprising computer executable instructions, wherein the computer executable instructions are used to execute a method for analyzing the heartbeat in an electrocardiogram signal when executed by a computer processor, the method comprising:

Preprocessing the collected ECG signals to obtain a number of sample signals of set lengths;

Performing feature extraction on the sample signal to generate feature vectors of the sample signal at three scales, wherein the lengths of the feature vectors at the three scales are respectively one-half, one-quarter, and one-eighth of the set length;

Constructing an anchor point of a preset length according to the feature vector corresponding to the sample signal, and calculating the reference point position of the heartbeat according to the anchor point;

The feature vector of each heartbeat within the corresponding reference point position range is input into the type recognition module, and is calculated in sequence by the convolution layer, activation layer, batch normalization layer and linear fully connected layer of the type recognition module to output a five-dimensional vector, which is used to characterize the probability that the heartbeat belongs to the five heartbeat types.

Of course, the computer executable instructions of a storage medium including computer executable instructions provided in an embodiment of the present invention are not limited to the method operations described above, and can also execute related operations in the method for analyzing the heartbeat of an electrocardiogram signal provided in any embodiment of the present invention.

Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

It is worth noting that in the embodiment of the above-mentioned device for analyzing the cardiac beat of the electrocardiogram signal, the various units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be achieved; in addition, the specific names of the various functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims (9)

1. An analysis method for electrocardiosignal center beat is characterized by comprising the following steps:

preprocessing the acquired electrocardiosignals to obtain a plurality of sections of sample signals with set lengths;

performing feature extraction on the sample signal to generate feature vectors of the sample signal on three scales, wherein the lengths of the feature vectors on the three scales are one half, one quarter and one eighth of the set length respectively;

constructing an anchor point with a preset length according to the characteristic vector corresponding to the sample signal, and calculating the reference point position of the heartbeat according to the anchor point;

inputting the characteristic vector of each heart beat in the corresponding reference point position range into a type identification module, sequentially calculating a convolution layer, an activation layer, a batch normalization layer and a linear full-connection layer of the type identification module, and outputting a five-dimensional vector, wherein the five-dimensional vector is used for representing the probability that the heart beat corresponds to five heart beat types;

wherein, the performing feature extraction on the sample signal to generate feature vectors of the sample signal on three scales, the lengths of the feature vectors on the three scales are respectively one half, one quarter and one eighth of the set length, and the method includes:

inputting the sample signal into an initial feature extraction module to obtain an initial feature extraction result, carrying out feature extraction on the initial feature extraction result according to corresponding scale grades by the feature extraction modules corresponding to three scales, wherein the lengths of feature vectors on the three scales are respectively one half, one quarter and one eighth of the set length;

and outputting the feature extraction result corresponding to each scale to the corresponding convolution layer and generating the feature vectors of the sample signals on the three scales by integrating the feature extraction results on the adjacent scales.

2. The analysis method according to claim 1, wherein the constructing an anchor point with a preset length according to the feature vector corresponding to the sample signal, and calculating the reference point position of the heartbeat according to the anchor point comprises:

for the feature vector corresponding to each scale, constructing an anchor point with the length of 9 sample points by taking each sample point on the feature vector as a center, wherein the fraction of the anchor point is the value of the central sample point of the anchor point;

inputting the feature vector corresponding to the anchor point into a heart beat position detection and measurement module to obtain the fraction of the heart beat probability, the variable quantity from the anchor point to the boundary box and the continuous time proportion of each wave band;

sequencing all the anchor points from large to small according to the scores to generate a check list;

sequentially screening candidate regions of the anchor points in the inspection list, adding the anchor points screened each time into the candidate list as candidate regions, deleting the anchor points and other anchor points within 0.2 seconds of the anchor points in the inspection list, wherein the anchor points screened each time are the anchor points with the highest score, the starting points of the anchor points are more than 0, and the end points of the anchor points are smaller than the length of the corresponding feature vectors in the inspection list;

calculating the midpoint and the width of a heart beat boundary box according to the midpoint and the width of the anchor points in the candidate list and the variation, and calculating the starting point and the end point of the heart beat boundary box according to the midpoint and the width of the heart beat boundary box;

and calculating the positions of the five datum points of the heartbeat according to the starting point, the width and the time proportion of the heartbeat boundary frame.

3. The analysis method according to claim 1, wherein the total loss of the type recognition module during the training process is calculated by the following formula:

loss＝det_loss+bbx_loss+fiducial_loss+cls_loss

wherein, loss is total loss, det _ loss is heart beat position detection loss, bbx _ loss is heart beat boundary box range detection loss, fit _ loss is reference point position detection loss and cls _ loss is heart beat type classification loss.

4. The analysis method according to claim 3, wherein the heart beat position detection loss is calculated by the following formula:

wherein, score _i Represents the output score (0) of the ith anchor point at the heart beat position detection and measurement module<score _i <1)，f _i Indicating the type of anchor point, f when the absolute value of the difference between the anchor point position and the true heart beat position is less than 0.15 second _i =1, otherwise f _i ＝0。

5. The analysis method according to claim 3, wherein the heart beat bounding box range detection loss bbx _ loss is calculated by the following formula:

the fiducial position detection loss, fidual _ loss, is calculated by the following formula:

wherein, the starting point and the end point of the real solid beat corresponding to the jth anchor point are respectively

And &>

The jth anchor point corresponds to the real P wave starting point, P wave end point, QRS wave starting point, QRS wave end point and T wave end point respectively

And &>

The jth anchor point corresponds to the start point and the end point of the predicted heartbeat and is/are respectively->

And &>

The corresponding predicted P wave starting point, P wave end point, QRS wave starting point, QRS wave end point and T wave end point of the jth anchor point are ^ and/or ^>

And &>

6. The analytical method of claim 3, wherein the type classification penalty is calculated by the formula:

wherein,

indicating the probability that the jth anchor belongs to a type k heartbeat>

Represents whether the anchor point belongs to the kth type and if so->

Or vice versa>

7. An analysis device for cardiac signal center beat, comprising:

the preprocessing unit is used for preprocessing the acquired electrocardiosignals to obtain a plurality of sections of sample signals with set lengths;

the characteristic extraction unit is used for extracting characteristics of the sample signal and generating characteristic vectors of the sample signal on three scales, and the lengths of the characteristic vectors on the three scales are respectively one half, one quarter and one eighth of the set length;

the reference point detection unit is used for constructing an anchor point with a preset length according to the characteristic vector corresponding to the sample signal and calculating the reference point position of the heartbeat according to the anchor point;

the heart beat classification unit is used for inputting the feature vector of each heart beat in the corresponding reference point position range into the type identification module, sequentially calculating a convolution layer, an activation layer, a batch normalization layer and a linear full-connection layer of the type identification module, and outputting a five-dimensional vector, wherein the five-dimensional vector is used for representing the probability that the heart beat corresponds to five heart beat types;

wherein the feature extraction unit includes:

the characteristic extraction module is used for inputting the sample signal into the initial characteristic extraction module to obtain an initial characteristic extraction result, the characteristic extraction modules corresponding to three scales perform characteristic extraction on the initial characteristic extraction result according to corresponding scale grades, and the lengths of characteristic vectors on the three scales are respectively one half, one quarter and one eighth of the set length;

and the characteristic vector generation module is used for outputting the characteristic extraction result corresponding to each scale to the corresponding convolution layer and synthesizing the characteristic extraction results on adjacent scales to generate the characteristic vectors of the sample signal on three scales.

8. An apparatus, characterized in that the apparatus comprises:

one or more processors;

a memory for storing one or more programs,

when executed by the one or more processors, cause the one or more processors to implement a method of analyzing a center beat of cardiac electrical signals as recited in any of claims 1-6.

9. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a method for analyzing a center beat of an electrocardiogram signal as claimed in any one of claims 1 to 6.

Images