Disclosure of Invention
In order to solve the technical problem of low training accuracy of the existing arrhythmia recognition model, the invention aims to provide an arrhythmia recognition method and system based on a lightweight neural network model, and the adopted technical scheme is as follows:
in a first aspect of the present invention, there is provided an arrhythmia recognition method based on a lightweight neural network model, comprising:
Dividing an electrocardiogram fluctuation curve of a subject to be monitored into a plurality of heartbeat cycle wave patterns, and clustering to obtain a plurality of clusters, thereby obtaining a heartbeat fluctuation abnormal cluster;
acquiring a body index fluctuation curve which is the same time as each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster;
screening to obtain a target heartbeat cycle waveform diagram based on waveform differences of adjacent body index fluctuation curves, wherein the target heartbeat cycle waveform diagram represents a heartbeat cycle waveform diagram influenced by movement;
Forming continuous heartbeat cycle wave patterns in the heartbeat fluctuation abnormal cluster in time sequence into continuous heartbeat cycle wave patterns, and obtaining real motion abnormal probability of the continuous heartbeat cycle wave patterns according to the number and the density of target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns;
And obtaining the weight of the waveform characteristic of the waveform chart of the continuous heartbeat cycle based on the real motion abnormality probability.
In an exemplary embodiment, the process for obtaining the heartbeat fluctuation abnormal cluster includes:
Obtaining the fluctuation abnormal degree of each wave in the cluster, wherein the wave types comprise P waves, QRS wave groups and T waves;
fusing the fluctuation anomaly degree of all the seed waves of the cluster to obtain the fluctuation anomaly degree of the cluster;
and taking the cluster corresponding to the fluctuation abnormality degree which is greater than or equal to the preset fluctuation abnormality degree threshold as the heartbeat fluctuation abnormality cluster.
In one exemplary embodiment, the process of obtaining the degree of fluctuation anomaly for each wave in the cluster includes:
acquiring characteristic values of various waves in each heartbeat cycle oscillogram in the cluster;
calculating the average value of the characteristic values of the same wave of all heartbeat cycle oscillograms in a cluster to obtain a cluster single wave characteristic value of each wave of the cluster;
and obtaining the fluctuation anomaly degree of each wave of the cluster based on the difference between the single wave characteristic value of the cluster and the corresponding preset normal characteristic value range of each wave of the cluster.
In an exemplary embodiment, the filtering process of the target heartbeat cycle waveform graph includes:
Obtaining the fluctuation burst degree of a first body index fluctuation curve according to the waveform difference of the first body index fluctuation curve and a second body index fluctuation curve, wherein the first body index fluctuation curve is a body index fluctuation curve which is the same as any heartbeat cycle waveform chart in the heartbeat fluctuation abnormal cluster, and the second body index fluctuation curve is a body index fluctuation curve section which is adjacent to the left side of the first body index fluctuation curve and has the same length;
And taking a heartbeat cycle waveform diagram corresponding to the fluctuation burst degree which is larger than or equal to a preset fluctuation burst degree threshold value as the target heartbeat cycle waveform diagram in the heartbeat fluctuation abnormal cluster.
In one exemplary embodiment, the process for obtaining the fluctuation burst degree of the first body index fluctuation curve includes:
acquiring the data mean value difference between the data mean value of the first body index fluctuation curve and the data mean value of the second body index fluctuation curve;
Acquiring Euclidean distance between a first data set and a second data set, wherein the first data set consists of maximum data and minimum data in the first body index fluctuation curve, and the second data set consists of maximum data and minimum data in the second body index fluctuation curve;
And obtaining the fluctuation burst degree of the first body index fluctuation curve according to the data mean value difference and the Euclidean distance, wherein the fluctuation burst degree is in direct proportion to the data mean value difference and the Euclidean distance.
In an exemplary embodiment, the process of obtaining the intensity level includes obtaining a distance between any two adjacent target heartbeat cycle waveform graphs in the continuous heartbeat cycle waveform graphs, and calculating an average value of the distances to obtain the intensity level of the target heartbeat cycle waveform graphs in the continuous heartbeat cycle waveform graphs.
In an exemplary embodiment, the process for obtaining the probability of real motion abnormality includes:
acquiring the number ratio of the target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagrams;
And obtaining the real motion abnormality probability of the continuous heartbeat cycle waveform graph according to the quantity duty ratio and the intensity, wherein the real motion abnormality probability is in direct proportion to the quantity duty ratio and in inverse proportion to the intensity.
In one exemplary embodiment, the process of acquiring the weights of the waveform features of the waveform diagrams of the continuous heartbeat cycle includes:
Obtaining the importance degree of the continuous heartbeat cycle waveform diagram according to the real motion abnormality probability of the continuous heartbeat cycle waveform diagram, wherein the importance degree is inversely proportional to the real motion abnormality probability;
and obtaining the weight of the waveform characteristics of the waveform diagrams of the continuous heartbeat cycles according to the importance degree.
In an exemplary embodiment, if a plurality of electrocardiographic wave curves of a subject to be monitored are acquired, a process for acquiring weights of waveform features of waveform diagrams of successive heart cycles in each electrocardiographic wave curve includes:
obtaining the number of clusters in each electrocardiogram wave curve, and combining the real motion abnormality probability of the continuous heartbeat cycle wave patterns in each electrocardiogram wave curve to obtain the importance degree of the continuous heartbeat cycle wave patterns in each electrocardiogram wave curve, wherein the importance degree is in direct proportion to the number of clusters and in inverse proportion to the real motion abnormality probability;
And obtaining the weight of the waveform characteristics of the continuous heartbeat cycle waveform graphs in each electrocardiogram fluctuation curve according to the importance degree of the continuous heartbeat cycle waveform graphs in each electrocardiogram fluctuation curve.
In a second aspect of the invention, an arrhythmia recognition system based on a lightweight neural network model is provided, and the arrhythmia recognition system comprises a memory and a processor, wherein the memory is connected with the processor, is used for storing program instructions, and is used for realizing the arrhythmia recognition method based on the lightweight neural network model when the program instructions are executed.
The method has the advantages that the heart-wave fluctuation curve of the main body to be monitored is divided into a plurality of heart-wave cycle wave patterns and clustered to obtain a plurality of clustered clusters, the heart-wave fluctuation abnormal clustered clusters can be obtained, subsequent abnormal recognition and processing are facilitated, then a certain fluctuation exists on the basis of body indexes during movement, therefore, whether the abnormality of each heart-wave cycle wave pattern is caused by movement or not is determined on the basis of the relation between the body index fluctuation curve with each heart-wave cycle wave pattern in the heart-wave fluctuation abnormal clustered clusters at the same time, and the target heart-wave cycle wave patterns are obtained through screening.
Detailed Description
In order to further describe the technical means and effects adopted by the invention to achieve the preset aim, the following detailed description of the specific implementation, structure, characteristics and effects of the invention is given with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" means that the embodiments are not necessarily the same. Furthermore, the particular features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The data information acquired by the method is fully authorized to be acquired, and the collection, the use and the processing of the related information are required to comply with related laws and regulations and standards of related countries and regions.
The embodiment provides an arrhythmia recognition method based on a lightweight neural network model, as shown in fig. 1, comprising the following steps:
dividing an electrocardiogram fluctuation curve of a main body to be monitored into a plurality of heartbeat cycle wave patterns, and clustering to obtain a plurality of cluster clusters, thereby obtaining a heartbeat fluctuation abnormal cluster;
Step 2, acquiring a body index fluctuation curve which is the same time as each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster;
Step 3, screening to obtain a target heartbeat cycle waveform diagram based on waveform differences of adjacent body index fluctuation curves, wherein the target heartbeat cycle waveform diagram represents a heartbeat cycle waveform diagram influenced by movement;
Step 4, forming continuous heartbeat cycle wave patterns in the heartbeat fluctuation abnormal cluster in time sequence into continuous heartbeat cycle wave patterns, and obtaining real motion abnormal probability of the continuous heartbeat cycle wave patterns according to the number and the density of target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns;
and 5, obtaining the weight of the waveform characteristics of the waveform diagrams of the continuous heartbeat cycles based on the real motion abnormality probability.
The steps will be specifically described below with reference to the drawings.
And step 1, dividing an electrocardiogram fluctuation curve of a main body to be monitored into a plurality of heartbeat cycle oscillograms, and clustering to obtain a plurality of clustering clusters, thereby obtaining a heartbeat fluctuation abnormal clustering cluster.
The subject to be monitored generally refers to a patient, and in one specific application scenario, the patient wears a wearable device (such as a smart watch or a smart bracelet with an electrocardiogram acquisition function, or a professional electrocardiogram acquisition device) in a hospital or at home to acquire an electrocardiogram fluctuation curve in a monitoring period. The time length of the monitoring time period is set by actual needs, and because the electrocardiographic fluctuation curve of the patient is used as a training sample, the lightweight neural network model is trained according to the electrocardiographic fluctuation curve of the patient, so that the arrhythmia recognition model is obtained, and the monitoring time period can be set longer in order to improve the accuracy of the training sample. Furthermore, the present embodiment may acquire one electrocardiographic wave, or may acquire a plurality of electrocardiographic waves. An electrocardiographic wave is taken as an example as follows.
Meanwhile, the wearable device also collects relevant body indexes of the patient, wherein the body indexes particularly can generate corresponding fluctuation body indexes such as respiratory frequency, body temperature, blood pressure and the like when the patient moves. And acquiring the body index in the patient monitoring time period according to the preset sampling frequency, so as to obtain a body index fluctuation curve. In this embodiment, the body index takes the respiratory rate as an example, and then the respiratory rate in the patient monitoring period is obtained according to the preset sampling frequency, so as to obtain a respiratory rate fluctuation curve.
Since the electrocardiographic wave includes a plurality of heart cycles, the electrocardiographic wave is divided into a plurality of heart cycle waveform diagrams. Each heartbeat cycle waveform is composed of a P wave, QRS complex, T wave, etc., as shown in fig. 2. In fig. 2, PR interval, ST segment, QT interval, etc. are also included. Accordingly, in the present embodiment, the kinds of waves included in the heartbeat cycle waveform diagram include P waves, QRS complexes, and T waves. In an exemplary embodiment, the existing P-wave detection method, QRS complex detection method, and T-wave detection method may be used to detect the P-wave, QRS complex, and T-wave in each heartbeat cycle waveform diagram, respectively, and at the same time, may obtain the feature value corresponding to each wave. The eigenvalues corresponding to each wave are shown in table 1.
TABLE 1
And then clustering the plurality of heartbeat cycle oscillograms to obtain a plurality of clusters. In one exemplary embodiment, the present embodiment employs a K-means clustering method to cluster the heartbeat cycle waveform map. And acquiring DTW (DYNAMIC TIME WARPING) distances of every two heartbeat cycle wave patterns, and clustering all the heartbeat cycle wave patterns based on the DTW distances by adopting a K-means clustering method to obtain a plurality of clusters. Wherein the number of clusters is determined by the contour coefficients. The larger the number of clusters obtained, the more the number of changes in the heart rate fluctuation of the patient in the electrocardiogram wave curve.
Referring to table 2, table 2 shows characteristic values corresponding to each wave of a plurality of patients.
TABLE 2
Reference is made to fig. 10, 10 for a schematic diagram of an ECG signal of patient 1, 11 for an ECG signal of patient 2, 12 for an ECG signal of patient 3, 13 for an ECG signal of patient 4, 14 for an ECG signal of patient 5, 15 for an ECG signal of patient 6, 16 for an ECG signal of patient 7, 17 for an ECG signal of patient 8, 17 for an ECG signal of patient 18, 18 for an ECG signal of patient 9, 19 for an ECG signal of patient 10, 20 for an ECG signal of patient 11, 21 for an ECG signal of patient 12, 22 for a loss function, 23 for a change in test accuracy and training accuracy, and 23 for a change in the loss function.
And finally, determining and obtaining the heartbeat fluctuation abnormal cluster from the obtained clusters based on the normal condition of the characteristic value of each wave in the clusters. In an exemplary embodiment, as shown in fig. 3, a specific acquisition procedure of the heartbeat fluctuation anomaly cluster is given as follows:
and step 1-1, obtaining the fluctuation anomaly degree of each wave in the cluster.
With each wave as an analysis object, the larger the difference between the characteristic value of each wave in the cluster and the corresponding preset normal characteristic value range is, the larger the possibility of abnormal fluctuation of the wave is. Accordingly, as shown in fig. 4, a specific acquisition procedure of the fluctuation anomaly degree of each wave in the cluster is given as follows:
and step 1-1, acquiring characteristic values of various waves in each heartbeat cycle oscillogram in the cluster.
For convenience of explanation, the first cluster is set to be any one cluster, the first heartbeat cycle waveform chart is any one heartbeat cycle waveform chart in the first cluster, and the first wave is any one wave in the first heartbeat cycle waveform chart. And acquiring the characteristic value of the first wave in the first heartbeat cycle waveform diagram in the first cluster.
And 1-1-2, calculating the average value of the characteristic values of the same wave of all the heartbeat cycle oscillograms in the cluster to obtain a cluster single wave characteristic value of each wave of the cluster.
Since the first cluster includes a plurality of heartbeat cycle wave patterns, the characteristic value of the first wave of each heartbeat cycle wave pattern in the first cluster is obtained, so that the average value of the characteristic values of the first waves of all heartbeat cycle wave patterns in the first cluster is calculated and used as the cluster single wave characteristic value of the first wave. Thereby obtaining the cluster single wave characteristic value of each wave in the first cluster.
And 1-1-3, obtaining the fluctuation abnormal degree of each wave of the cluster based on the difference between the single wave characteristic value of the cluster and the corresponding preset normal characteristic value range of each wave of the cluster.
It should be understood that each wave has a normal characteristic value range, and thus, a normal characteristic value range is preset for each wave. Taking the first wave as an example, when the first wave is in the normal characteristic value range corresponding to the first wave, the first wave is normal, when the first wave is not in the normal characteristic value range corresponding to the first wave, the first wave is abnormal, and when the difference between the first wave and the normal characteristic value range of the first wave is larger, the degree of abnormality of the first wave is larger.
Therefore, the fluctuation anomaly degree of each wave of the first cluster is obtained based on the difference between the cluster single-wave characteristic value of each wave of the first cluster and the corresponding preset normal characteristic value range.
If the cluster single wave characteristic value of the first wave is in the normal characteristic value range corresponding to the first wave, the fluctuation abnormality degree of the first wave of the first cluster is set to be 0, otherwise, the difference characteristic of the fluctuation abnormality degree of the first wave is obtained, wherein the difference characteristic is obtained in the process that if the cluster single wave characteristic value of the first wave is larger than the upper limit value of the normal characteristic value range corresponding to the first wave, the difference value is calculated and is the difference characteristic, and if the cluster single wave characteristic value of the first wave is smaller than the lower limit value of the normal characteristic value range corresponding to the first wave, the difference value is calculated and is the difference characteristic.
And then normalizing the difference characteristics of the first wave of the first cluster, wherein the normalized result is the fluctuation anomaly degree of the first wave of the first cluster. The normalization method comprises the steps of obtaining the maximum value and the minimum value in the difference characteristics of the first wave of all the clusters, and normalizing the difference characteristics of the first wave of the first cluster by adopting a maximum value and minimum value normalization method.
It will be appreciated that when arrhythmias are present, abnormalities such as atrial fibrillation, ventricular fibrillation, conduction block etc. may occur, and that the abnormalities exhibited by different abnormalities on the electrocardiogram wave pattern are different, such as atrial fibrillation being characterized by the disappearance of p-waves. Therefore, in the above manner, the degree of fluctuation abnormality of each wave of each cluster is obtained.
And 1-2, fusing the fluctuation anomaly degree of all the seed waves of the cluster to obtain the fluctuation anomaly degree of the cluster.
Fusing the fluctuation anomaly degree of all the seed waves of the first cluster, and in one exemplary embodiment, calculating an average value of the fluctuation anomaly degree of all the seed waves of the first cluster as the fluctuation anomaly degree of the first cluster. By adopting the mode, the fluctuation abnormity degree of each cluster is obtained.
And 1-3, taking the cluster corresponding to the fluctuation abnormality degree which is greater than or equal to the preset fluctuation abnormality degree threshold as a heartbeat fluctuation abnormality cluster.
The degree of fluctuation abnormality reflects the abnormality of each wave of the cluster, and the higher the degree of fluctuation abnormality, the more serious the abnormality of each wave of the cluster. In an exemplary embodiment, a threshold level of fluctuation abnormality is preset. The value range of the preset fluctuation abnormal degree threshold is 0-1, and the specific value of the preset fluctuation abnormal degree threshold is set by actual conditions. If more reliable fluctuation abnormality degree judgment logic is needed, the preset fluctuation abnormality degree threshold can be set to be smaller, so that more clusters satisfy the judgment logic, and in this embodiment, 0.3 is taken as an example.
Comparing the fluctuation abnormality degree of each cluster with a preset fluctuation abnormality degree threshold to obtain a cluster corresponding to the fluctuation abnormality degree greater than or equal to the preset fluctuation abnormality degree threshold, and taking the cluster corresponding to the fluctuation abnormality degree greater than or equal to the preset fluctuation abnormality degree threshold as a heartbeat fluctuation abnormality cluster. The abnormal occurrence of the fluctuation of the heartbeat cycle waveform chart in the heartbeat fluctuation abnormal cluster can be the manifestation of arrhythmia.
And 2, acquiring a body index fluctuation curve which is at the same time as each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster.
It should be understood that the number of the abnormal cluster of heartbeat fluctuation may be only one or may be plural, and any abnormal cluster of heartbeat fluctuation is taken as an example. If the plurality of heartbeat fluctuation abnormal clustering clusters exist, the data processing process of each heartbeat fluctuation abnormal clustering cluster is the same as the data processing process of one heartbeat fluctuation abnormal clustering cluster.
The cluster of beat fluctuation anomalies may be due to movement or the like, and therefore it is necessary to determine whether the fluctuation anomalies are due to movement or the like in combination with the patient's physical index fluctuation curve. Because the electrocardiogram fluctuation curve and the body index fluctuation curve are synchronously acquired, the acquisition time of each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster is obtained after the heartbeat fluctuation abnormal cluster is obtained, and therefore, the body index fluctuation curve which is the same in time as each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster is obtained. The number of the heartbeat cycle wave patterns in the heartbeat fluctuation abnormal cluster is the same as that of the body index wave curves, the heartbeat cycle wave patterns and the body index wave curves are in one-to-one correspondence, and each heartbeat cycle wave pattern has the body index wave curve at the same time. And setting that the heartbeat fluctuation abnormal cluster comprises U heartbeat cycle wave patterns, wherein the body index fluctuation curves are also U, and the U-th heartbeat cycle wave patterns and the U-th body index fluctuation curves in the heartbeat fluctuation abnormal cluster are in the same time, and U is equal to [1, U ].
And 3, screening to obtain a target heartbeat cycle waveform diagram based on waveform differences of adjacent body index fluctuation curves, wherein the target heartbeat cycle waveform diagram represents a heartbeat cycle waveform diagram influenced by movement.
Based on the waveform differences of the adjacent body index fluctuation curves, whether the fluctuation of the body index is sudden fluctuation or continuous fluctuation is determined, so that a heartbeat cycle waveform diagram influenced by movement, namely a target heartbeat cycle waveform diagram, is determined. In an exemplary embodiment, as shown in fig. 5, a specific screening procedure of the target heartbeat cycle waveform is given as follows:
And 3-1, obtaining the fluctuation burst degree of the first body index fluctuation curve according to the waveform difference of the first body index fluctuation curve and the second body index fluctuation curve.
For convenience of explanation, the first body index fluctuation curve is set to be a body index fluctuation curve with the same time as any one of the heartbeat cycle waveform diagrams in the heartbeat fluctuation abnormal cluster.
Because the body index can have certain fluctuation condition when the patient moves, the body index fluctuation curve section which is adjacent to the left side of the first body index fluctuation curve and has the same length is acquired and is defined as a second body index fluctuation curve. Because the left adjacent is earlier than the first body index fluctuation curve in time, the acquisition time of the second body index fluctuation curve is earlier than the first body index fluctuation curve, the starting time of the first body index fluctuation curve is the ending time of the second body index fluctuation curve, and the lengths of the first body index fluctuation curve and the second body index fluctuation curve are the same. For example, the starting time and the ending time of the second body index wave curve are respectively 10 points 1 minute 0 seconds and 10 points 2 minutes 0 seconds, and the starting time and the ending time of the first body index wave curve are respectively 10 points 2 minutes 0 seconds and 10 points 3 minutes 0 seconds. The first body index profile and the second body index profile are each 1 minute in length. It should be understood that if the first body index fluctuation curve does not have a preceding body index fluctuation curve in time sequence, that is, does not have a body index fluctuation curve adjacent to the left, the first body index fluctuation curve is no longer involved in the acquisition of the subsequent fluctuation burst degree, and the acquisition of the subsequent fluctuation burst degree is only performed on other body index fluctuation curves adjacent to the left.
Based on the waveform difference of the first body index fluctuation curve and the second body index fluctuation curve, the fluctuation burst degree of the first body index fluctuation curve is obtained, and the larger the waveform difference of the first body index fluctuation curve and the second body index fluctuation curve is, the larger the fluctuation burst degree of the first body index fluctuation curve is.
In an exemplary embodiment, as shown in fig. 6, a specific acquisition procedure of the fluctuation burst degree of the first body index fluctuation curve is given as follows:
And step 3-1-1, obtaining the data average difference between the data average value of the first body index fluctuation curve and the data average value of the second body index fluctuation curve.
Because the first body index fluctuation curve and the second body index fluctuation curve are composed of a plurality of data points, when the body index takes the respiratory frequency as an example, the first body index fluctuation curve and the second body index fluctuation curve are composed of respiratory frequency values at all sampling moments. Then, an average value of the data values of the plurality of data points included in the first body indicator fluctuation curve is calculated to obtain a first data average value, and an average value of the data values of the plurality of data points included in the second body indicator fluctuation curve is calculated to obtain a second data average value. And acquiring the data mean value difference of the first data mean value and the second data mean value, and particularly acquiring the absolute value of the difference value of the first data mean value and the second data mean value. The larger the difference of the data means, the larger the difference between the first body index fluctuation curve and the second body index fluctuation curve.
And 3-1-2, acquiring Euclidean distances of the first data set and the second data set.
The method comprises the steps of obtaining a fluctuation range of a first body index fluctuation curve, namely obtaining a maximum data value and a minimum data value in the first body index fluctuation curve, and forming a first data group by the maximum data value and the minimum data value in the first body index fluctuation curve, obtaining a fluctuation range of a second body index fluctuation curve, namely obtaining the maximum data value and the minimum data value in the second body index fluctuation curve, and forming a second data group by the maximum data value and the minimum data value in the second body index fluctuation curve.
The Euclidean distance between the first data group and the second data group is obtained, specifically, the difference value between the minimum data value in the first data group and the minimum data value in the second data group is calculated, the difference value between the maximum data value in the first data group and the maximum data value in the second data group is calculated, then the square sum of the two difference values is calculated, and the obtained result is the Euclidean distance between the first data group and the second data group.
The larger the euclidean distance of the first data set and the second data set, the larger the difference between the first body index profile and the second body index profile.
And 3-1-3, obtaining the fluctuation burst degree of the first body index fluctuation curve according to the data mean value difference and the Euclidean distance.
According to the steps, the larger the difference of the data mean values of the first data set and the second data set is, the larger the difference of the first body index fluctuation curve and the second body index fluctuation curve is, the larger the fluctuation burst degree of the first body index fluctuation curve is, the data mean value difference is in direct proportion to the fluctuation burst degree, the larger the Euclidean distance between the first data set and the second data set is, the larger the difference of the first body index fluctuation curve and the second body index fluctuation curve is, the larger the fluctuation burst degree of the first body index fluctuation curve is, and the Euclidean distance is in direct proportion to the fluctuation burst degree. Therefore, the fluctuation burst degree of the first body index fluctuation curve can be obtained according to the data mean value difference and the Euclidean distance of the first data set and the second data set.
In an exemplary embodiment, the product of the difference of the data means of the first data set and the second data set and the euclidean distance is calculated, and then normalized, and the normalized result is the fluctuation burst degree of the first body index fluctuation curve. The normalization method can be that the maximum value and the minimum value in the products of the data mean value difference and the Euclidean distance of the body index fluctuation curves corresponding to all the heartbeat cycle oscillograms in the heartbeat fluctuation abnormal cluster are obtained, and then the maximum value and the minimum value normalization method is adopted to normalize the products corresponding to the first body index fluctuation curves.
By adopting the process, the fluctuation burst degree of the body index fluctuation curve corresponding to each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster is obtained, and the fluctuation burst degree of each heartbeat cycle oscillogram in the heartbeat fluctuation abnormal cluster is obtained.
It should be understood that if multiple types of body index wave curves, such as respiratory rate and body temperature, are acquired, the wave burst degree of each body index wave curve is acquired respectively, and then an average value is calculated to obtain the final wave burst degree.
And 3-2, taking a heartbeat cycle waveform diagram corresponding to the fluctuation burst degree which is greater than or equal to a preset fluctuation burst degree threshold as a target heartbeat cycle waveform diagram in the heartbeat fluctuation abnormal cluster.
The greater the degree of fluctuation burst, the more the abnormality caused by the movement. Presetting a fluctuation burst degree threshold, wherein the value range of the preset fluctuation burst degree threshold is 0-1, and the specific value of the preset fluctuation burst degree threshold is set by actual conditions. The preset fluctuation burst degree threshold is used for screening and obtaining the heartbeat cycle waveform diagram of the motion abnormality, if stricter judgment logic is needed, the preset fluctuation burst degree threshold can be set to be smaller, so that the heartbeat cycle waveform diagram meeting the requirement is easier to obtain, omission is avoided, and the embodiment takes 0.5 as an example.
Comparing the fluctuation burst degree of each heartbeat cycle waveform chart in the heartbeat fluctuation abnormal cluster with a preset fluctuation burst degree threshold value, acquiring a heartbeat cycle waveform chart corresponding to the fluctuation burst degree which is greater than or equal to the preset fluctuation burst degree threshold value, and taking the heartbeat cycle waveform chart corresponding to the fluctuation burst degree which is greater than or equal to the preset fluctuation burst degree threshold value as a target heartbeat cycle waveform chart in the heartbeat fluctuation abnormal cluster. The target heart cycle waveform chart characterizes the fluctuation anomalies due to motion.
And 4, forming continuous heartbeat cycle wave patterns in the heartbeat fluctuation abnormal cluster in time sequence into continuous heartbeat cycle wave patterns, and obtaining the real motion abnormal probability of the continuous heartbeat cycle wave patterns according to the number and the density of the target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns.
In general, the movement is a continuous process, and thus, the heartbeat cycle waveform patterns which are continuous in time series in the heartbeat fluctuation abnormal cluster are composed into continuous heartbeat cycle waveform patterns. In an exemplary embodiment, all the heartbeat cycle wave patterns obtained by dividing the electrocardiographic fluctuation curve can be sequenced according to time sequence to obtain the sequence numbers of the heartbeat cycle wave patterns, then the heartbeat cycle wave patterns with continuous sequence numbers in the heartbeat fluctuation abnormal cluster are combined into continuous heartbeat cycle wave patterns, for example, the sequence numbers of all the heartbeat cycle wave patterns are respectively 1-10, the sequence numbers of the heartbeat cycle wave patterns in the heartbeat fluctuation abnormal cluster are respectively 2, 3, 6, 7 and 9, and then the 2 nd heartbeat cycle wave pattern and the 3 rd heartbeat cycle wave pattern are combined into one continuous heartbeat cycle wave pattern, and the 6 th heartbeat cycle wave pattern and the 7 th heartbeat cycle wave pattern are combined into one continuous heartbeat cycle wave pattern.
It should be understood that, for an isolated heartbeat cycle waveform map in the heartbeat fluctuation abnormal cluster, that is, in the heartbeat fluctuation abnormal cluster, there is no adjacent heartbeat cycle waveform map before and after, then, the division of the continuous heartbeat cycle waveform map and the acquisition of the true motion abnormality probability of the continuous heartbeat cycle waveform map, such as the 9 th heartbeat cycle waveform map in the above example, are not performed.
And then, obtaining the real motion abnormality probability of the continuous heartbeat cycle waveform diagram according to the quantity and the intensity of the target heartbeat cycle waveform diagram in the continuous heartbeat cycle waveform diagram fluctuation.
The degree of intensity characterizes the aggregation of the target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagrams, and in an exemplary embodiment, a distance between any two adjacent target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagrams is obtained, where the distance may be a difference value between sequence numbers of any two adjacent target heartbeat cycle waveform diagrams, for example, if the sequence numbers of the two adjacent target heartbeat cycle waveform diagrams are 2 and 4 respectively, the distance between the two adjacent target heartbeat cycle waveform diagrams is 4-2=2. And then calculating the average value of the distances between all two adjacent target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns as the density degree of the target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns.
In an exemplary embodiment, as shown in fig. 7, a specific acquisition procedure of the true motion abnormality probability is given as follows:
and 4-1, acquiring the number duty ratio of the target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagrams.
And for any one continuous heartbeat cycle waveform diagram, acquiring the number of target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagram and the total number of the heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagram, and calculating the ratio of the number of the target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagram to the total number of the heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagram to obtain the number of the target heartbeat cycle waveform diagrams in the continuous heartbeat cycle waveform diagram.
And 4-2, obtaining the real motion abnormality probability of the continuous heartbeat cycle oscillogram according to the quantity duty ratio and the density degree.
The larger the number ratio is, the more the number of the target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns is, the more the motion is based on the process which is continuous, the more the possibility of motion abnormality of the continuous heartbeat cycle wave patterns is, namely the larger the probability of real motion abnormality is, the probability of real motion abnormality is in proportion to the number ratio, the smaller the density degree is, the smaller the distance between the target heartbeat cycle wave patterns in the continuous heartbeat cycle wave patterns is, the more the motion abnormality of the continuous heartbeat cycle wave patterns is based on the process which is continuous, namely the probability of real motion abnormality is larger, and the probability of real motion abnormality is in inverse proportion to the density degree.
In one exemplary embodiment, the degree of intensity is normalized by a negative correlation, such as exp (-x), where x is the object that requires negative correlation normalization and exp is an exponential function based on a natural constant e. And then calculating the product of the quantity duty ratio and the density degree after the negative correlation normalization, wherein the product is the real motion abnormality probability of the waveform diagrams of the continuous heartbeat cycles.
By adopting the process, the real motion abnormality probability of each continuous heartbeat cycle waveform chart is obtained.
It should be understood that if there is no target heartbeat cycle waveform diagram in the continuous heartbeat cycle waveform diagram, the true motion abnormality probability of the continuous heartbeat cycle waveform diagram is 0, and if there is only one target heartbeat cycle waveform diagram in the continuous heartbeat cycle waveform diagram, the true motion abnormality probability of the continuous heartbeat cycle waveform diagram is also set to 0.
And 5, obtaining the weight of the waveform characteristics of the waveform diagrams of the continuous heartbeat cycles based on the real motion abnormality probability.
Because the larger the probability of the real motion abnormality is, the larger the probability of the continuous heartbeat cycle waveform diagram affected by the motion is, in order to reduce the influence of the continuous heartbeat cycle waveform diagram on the accuracy of the training result of the lightweight neural network model, the smaller the weight is required to be set for the continuous heartbeat cycle waveform diagram, so that the training influence of the lightweight neural network model is reduced. Thus, the true motion anomaly probability of the continuous heartbeat cycle waveform map is inversely proportional to the weight of the waveform features of the continuous heartbeat cycle waveform map.
If only one electrocardiogram wave curve is obtained, the weight of the waveform characteristics of each continuous heartbeat cycle wave pattern is obtained according to the real motion abnormality probability of each continuous heartbeat cycle wave pattern in the electrocardiogram wave curve. In an exemplary embodiment, as shown in fig. 8, a specific acquisition procedure is given as follows:
And 5-1, obtaining the importance degree of the continuous heartbeat cycle waveform diagram according to the real motion abnormality probability of the continuous heartbeat cycle waveform diagram.
For convenience of explanation, the first continuous heartbeat cycle waveform chart is set to be any continuous heartbeat cycle waveform chart, and the importance degree of the first continuous heartbeat cycle waveform chart is obtained according to the real motion abnormality probability of the first continuous heartbeat cycle waveform chart. The importance level is inversely proportional to the probability of true motion anomalies. In an exemplary embodiment, a difference between the value 1 and the true motion abnormality probability of the first continuous heartbeat cycle waveform map is calculated, and the obtained difference is the importance degree of the first continuous heartbeat cycle waveform map.
And 5-2, obtaining the weight of the waveform characteristics of the waveform diagrams of the continuous heartbeat cycles according to the importance degree.
The weight of the waveform characteristic of the continuous heartbeat cycle waveform graph is obtained according to the importance degree, and the importance degree is higher, and in an exemplary embodiment, the importance degree of the obtained continuous heartbeat cycle waveform graph is taken as the weight of the waveform characteristic of the continuous heartbeat cycle waveform graph.
If at least two electrocardiographic wave curves are obtained, the real motion abnormality probability of each continuous heartbeat cycle wave diagram in each electrocardiographic wave curve is obtained according to the process of the step 1-4. Then, the weights of the waveform characteristics of each continuous heartbeat cycle waveform graph in each electrocardiogram wave curve are obtained for the actual motion abnormality probability of each continuous heartbeat cycle waveform graph in each electrocardiogram wave curve. In an exemplary embodiment, as shown in fig. 9, a specific acquisition procedure is given as follows:
and 5-3, obtaining the number of clusters in each electrocardiogram fluctuation curve, and combining the real motion abnormality probability of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve to obtain the importance degree of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve.
For ease of explanation, any one of the electrocardiographic curves is characterized as a first electrocardiographic curve as follows.
The method comprises the steps of obtaining the number of clusters in a first electrocardiogram fluctuation curve, and obtaining the importance degree of each continuous heartbeat cycle waveform diagram in the first electrocardiogram fluctuation curve by combining the actual motion abnormality probability of each continuous heartbeat cycle waveform diagram in the first electrocardiogram fluctuation curve.
The more the number of clusters is, the more the number of change times of heart rate fluctuation in the first electrocardiogram fluctuation curve is, the more the waveform characteristics are rich, and the corresponding identification of arrhythmia is more important, so that the importance degree of the first electrocardiogram fluctuation curve is higher, and the importance degree is in direct proportion to the number of clusters. The importance of each successive heart beat cycle waveform in the first electrocardiogram wave trace is inversely proportional to the probability of actual motion abnormality of each successive heart beat cycle waveform in the first electrocardiogram wave trace.
In one exemplary embodiment, the total number of clusters in all electrocardiographic wave curves is obtained. The ratio of the number of clusters in the first electrocardiogram wave curve to the total number of clusters is calculated as the feature of the number of clusters in the first electrocardiogram wave curve.
And calculating the product of the cluster number characteristic in the first electrocardiogram wave curve and the actual motion abnormality probability of each continuous heartbeat cycle wave form graph in the first electrocardiogram wave curve as the importance degree of each continuous heartbeat cycle wave form graph in the first electrocardiogram wave curve.
By adopting the process, the importance degree of each continuous heartbeat cycle waveform chart in each electrocardiogram fluctuation curve is obtained.
And 5-4, obtaining the weight of the waveform characteristics of the continuous heartbeat cycle waveform graphs in each electrocardiogram fluctuation curve according to the importance degree of the continuous heartbeat cycle waveform graphs in each electrocardiogram fluctuation curve.
The weight of the waveform characteristic of each continuous heartbeat cycle waveform graph in each electrocardiographic wave curve is obtained according to the importance degree of each continuous heartbeat cycle waveform graph in each electrocardiographic wave curve, and in an exemplary embodiment, the importance degree of each continuous heartbeat cycle waveform graph in each electrocardiographic wave curve is taken as the weight of the waveform characteristic of each continuous heartbeat cycle waveform graph in each electrocardiographic wave curve.
The weight of the waveform characteristic of each heartbeat cycle waveform chart in the continuous heartbeat cycle waveform charts is taken as the weight of the waveform characteristic of each heartbeat cycle waveform chart in the continuous heartbeat cycle waveform charts. The weight is set to 1 for each of the other heartbeat cycle waveform patterns except for each of the continuous heartbeat cycle waveform patterns in the electrocardiographic fluctuation curve. Thus, the weight of the waveform characteristic of each heartbeat cycle waveform chart in the electrocardiogram fluctuation curve is obtained.
In one exemplary embodiment, 80% of all heart cycle waveforms in all electrocardiogram wave patterns are used as training sets and 20% are used as test sets in subsequent specific applications.
Waveform characteristics of each heartbeat cycle waveform diagram in the training set are determined, wherein the waveform characteristics specifically refer to characteristic values of each wave in the heartbeat cycle waveform diagram. And multiplying the waveform characteristics of each heartbeat cycle waveform chart in the training set by the corresponding weight to realize weighting, and inputting the waveform characteristics of each weighted heartbeat cycle waveform chart into the lightweight neural network model as the training set to train. Wherein the loss function in the training process is a cross entropy loss function. Wherein the training of the model is a well known technique and will not be described in detail here. The trained model is recorded as an arrhythmia recognition model based on a lightweight neural network model. Please refer to fig. 22, fig. 22 is a schematic diagram of a loss function, fig. 23 is a schematic diagram of a change of the test accuracy and the training accuracy. The horizontal axis of the coordinate axes in fig. 22 and 23 is a period, that is, eopch, and one period is a process in which all training samples are completely propagated forward and backward once.
Referring to table 3, table 3 is a training data table;
TABLE 3 Table 3
When the arrhythmia recognition is needed to be carried out on the patient subsequently, the waveform characteristics of each wave in the electrocardiogram fluctuation curve of the patient are input into the arrhythmia recognition model, and then the arrhythmia recognition result can be output.
The embodiment also provides an arrhythmia recognition system based on the lightweight neural network model, which comprises a memory and a processor, wherein the memory is connected with the processor and is used for storing program instructions, and the processor is used for realizing the steps in the arrhythmia recognition method embodiment based on the lightweight neural network model when the program instructions are executed.
In one exemplary embodiment, the invention provides a computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of the lightweight neural network model-based arrhythmia identification method embodiment described above.
It should be noted that the sequence of the embodiments of the present invention is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments.