An arrhythmia identification method and system based on a lightweight neural network model

Abstract

The present application relates to the technical field of data processing, in particular to a heart arrhythmia recognition method and system based on a light neural network model, which divides an electrocardiogram fluctuation curve into a plurality of heartbeat cycle waveform graphs and clusters them to obtain heartbeat fluctuation abnormality clustering clusters, obtains a target heartbeat cycle waveform graph based on waveform differences of adjacent body index fluctuation curves at the same time as the heartbeat cycle waveform graph, forms continuous heartbeat cycle waveform graphs by using heartbeat cycle waveform graphs in the heartbeat fluctuation abnormality clustering clusters that are continuous in time sequence, obtains a real motion abnormality probability of the continuous heartbeat cycle waveform graphs according to the number and density of the target heartbeat cycle waveform graphs in the continuous heartbeat cycle waveform graphs, obtains the weight of the waveform features of the continuous heartbeat cycle waveform graphs, sets the weight of the heartbeat cycle waveform graphs affected by motion to be smaller, reduces the influence on the training of the light neural network model, and thus improves the training accuracy of the heart arrhythmia recognition model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a heart arrhythmia recognition method and system based on a light neural network model. BACKGROUND

[0002] Heart arrhythmia is mainly caused by abnormal activation of the sinoatrial node or origin of heart activity other than the sinoatrial node during daily activities and conduction of electrocardiogram signals, resulting in abnormal behavior of heart activity, directly manifested as abnormal heart rate and rhythm. The current heart arrhythmia recognition method mainly uses a network model training method.

[0003] Generally, electrocardiogram (ECG) data is detected by a wearable device on the body of a monitored subject. Based on the characteristics of low power consumption, real-time performance and portability of wearable devices, a light neural network model (such as a GhostNet network) with small model parameters is usually selected for recognition when recognizing heart arrhythmia. The overall process of the existing heart arrhythmia recognition method based on a light neural network model is as follows: obtaining an electrocardiogram fluctuation curve (i.e., an ECG fluctuation curve), dividing the electrocardiogram fluctuation curve into a plurality of heartbeat cycle waveform graphs, taking 80% of the number of heartbeat cycle waveform graphs as a training set and 20% of the number of heartbeat cycle waveform graphs as a test set, and inputting the waveform features in each heartbeat cycle waveform graph in the training set into a light neural network model for training. The loss function in the training process can use a cross-entropy loss function. The trained model is recorded as a heart arrhythmia recognition model. The trained heart arrhythmia recognition model can be used for heart arrhythmia recognition.

[0004] However, when obtaining electrocardiogram data of a monitored subject through a wearable device, the motion behavior of the monitored subject may cause the obtained electrocardiogram data to be distorted, and the motion behavior may affect the electrocardiogram data, thereby affecting the training accuracy of the heart arrhythmia recognition model and the accuracy of heart arrhythmia recognition when training the light neural network model based on the sample set obtained therefrom. SUMMARY

[0005] To solve the technical problem of low training accuracy of the existing heart arrhythmia recognition model, the purpose of the present application is to provide a heart arrhythmia recognition method and system based on a light neural network model, and the technical solution adopted is as follows:

[0006] In a first aspect of the present application, a heart arrhythmia recognition method based on a light neural network model is provided, comprising:

[0007] The electrocardiogram fluctuation curve of the subject to be monitored is divided into multiple heartbeat cycle waveforms, and multiple clusters are obtained from them, from which the abnormal heartbeat fluctuation cluster is obtained;

[0008] Obtain the body indicator fluctuation curves at the same time as the heartbeat cycle waveforms in the abnormal heartbeat fluctuation cluster;

[0009] Based on the waveform differences of adjacent body indicator fluctuation curves, a target heart rate cycle waveform is obtained by screening. The target heart rate cycle waveform represents the heart rate cycle waveform affected by exercise.

[0010] The heartbeat cycle waveforms that are sequentially continuous in the heartbeat fluctuation abnormal cluster are combined into a continuous heartbeat cycle waveform, and the true motion abnormality probability of the continuous heartbeat cycle waveform is obtained based on the number and density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform.

[0011] The weights of waveform features in the continuous heartbeat cycle waveform are obtained based on the actual motion abnormality probability.

[0012] In an exemplary embodiment, the process of obtaining the abnormal heartbeat fluctuation cluster includes:

[0013] Obtain the wave anomaly sub-degree for each wave type in the cluster; where the wave types include P-waves, QRS groups, and T-waves;

[0014] The fluctuation anomaly degree of the cluster is obtained by fusing the fluctuation anomaly degrees of all wave types in the cluster.

[0015] Clusters corresponding to fluctuation abnormalities that are greater than or equal to a preset fluctuation abnormality threshold are designated as heartbeat fluctuation abnormality clusters.

[0016] In an exemplary embodiment, the process of obtaining the fluctuation anomaly degree of each wave in the cluster includes:

[0017] Obtain the feature values ​​of various waves in the waveform diagram of each heartbeat cycle in the cluster;

[0018] Calculate the average value of the characteristic values ​​of the same wave in all heartbeat cycle waveforms in the cluster to obtain the cluster single wave characteristic value of each wave in the cluster.

[0019] Based on the difference between the cluster single-wave characteristic value of each wave in the cluster and the corresponding preset normal characteristic value range, the fluctuation anomaly degree of each wave in the cluster is obtained.

[0020] In an exemplary embodiment, the screening process for the target heart rate cycle waveform includes:

[0021] The degree of sudden fluctuation of the first physical indicator fluctuation curve is obtained based on the waveform difference between the first physical indicator fluctuation curve and the second physical indicator fluctuation curve; the first physical indicator fluctuation curve is the physical indicator fluctuation curve at the same time as any heartbeat cycle waveform in the abnormal heartbeat fluctuation cluster, and the second physical indicator fluctuation curve is the physical indicator fluctuation curve segment that is adjacent to the first physical indicator fluctuation curve on the left and of the same length.

[0022] The heartbeat cycle waveform corresponding to a fluctuation suddenness level greater than or equal to a preset fluctuation suddenness level threshold is used as the target heartbeat cycle waveform in the heartbeat fluctuation anomaly cluster.

[0023] In an exemplary embodiment, the process of obtaining the degree of sudden fluctuation of the first physical indicator fluctuation curve includes:

[0024] Obtain the difference between the average data of the first body indicator fluctuation curve and the average data of the second body indicator fluctuation curve;

[0025] Obtain the Euclidean distance between the first data set and the second data set. The first data set consists of the maximum and minimum data in the first body index fluctuation curve, and the second data set consists of the maximum and minimum data in the second body index fluctuation curve.

[0026] Based on the difference in the mean of the data and the Euclidean distance, the degree of sudden fluctuation of the first body indicator fluctuation curve is obtained, and the degree of sudden fluctuation is proportional to both the difference in the mean of the data and the Euclidean distance.

[0027] In an exemplary embodiment, the process of obtaining the density includes: obtaining the distance between any two adjacent target heartbeat cycle waveforms in the continuous heartbeat cycle waveform diagram, and calculating the average value of the distance to obtain the density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform diagram.

[0028] In an exemplary embodiment, the process of obtaining the probability of the actual motion anomaly includes:

[0029] Obtain the percentage of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graphs;

[0030] Based on the quantity ratio and density, the true motion abnormality probability of the continuous heartbeat cycle waveform is obtained. The true motion abnormality probability is directly proportional to the quantity ratio and inversely proportional to the density.

[0031] In an exemplary embodiment, the process of obtaining the weights of waveform features in a continuous heartbeat cycle waveform includes:

[0032] The importance of the continuous heartbeat cycle waveform is obtained from the actual motion abnormality probability of the continuous heartbeat cycle waveform, and the importance is inversely proportional to the actual motion abnormality probability.

[0033] The weights of waveform features in a continuous heartbeat cycle waveform are determined based on their importance.

[0034] In an exemplary embodiment, if multiple electrocardiogram (ECG) fluctuation curves of the subject to be monitored are obtained, the process of obtaining the weights of the waveform features of the continuous heartbeat cycle waveforms in each ECG fluctuation curve includes:

[0035] The number of clusters in each electrocardiogram fluctuation curve is obtained, and the importance of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve is obtained by combining the actual motion abnormality probability of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve. The importance is directly proportional to the number of clusters and inversely proportional to the actual motion abnormality probability.

[0036] The weights of the waveform characteristics of the continuous heartbeat cycle waveforms in each electrocardiogram fluctuation curve are obtained based on the importance of the waveforms in each electrocardiogram fluctuation curve.

[0037] In a second aspect of the present invention, a cardiac arrhythmia recognition system based on a lightweight neural network model is provided, comprising: a memory and a processor; the memory is connected to the processor; the memory is used to store program instructions; the processor is used to implement the above-described cardiac arrhythmia recognition method based on a lightweight neural network model when the program instructions are executed.

[0038] This invention has the following beneficial effects: By dividing the electrocardiogram (ECG) fluctuation curve of the subject to be monitored into multiple heartbeat cycle waveforms and clustering them into multiple clusters, abnormal heartbeat fluctuation clusters can be obtained, facilitating subsequent anomaly identification and processing. Furthermore, based on the fact that body indicators fluctuate during exercise, the relationship between the simultaneous fluctuation curves of body indicators and the heartbeat cycle waveforms within the abnormal heartbeat fluctuation clusters can be used to determine whether the anomalies in each heartbeat cycle waveform are caused by exercise, thereby filtering out target heartbeat cycle waveforms. Since exercise is a continuous process, the abnormal heartbeat fluctuation clusters that are sequentially connected in time can be further analyzed. The continuous heart rate cycle waveforms are composed of consecutive heart rate cycle waveforms. Based on the number and density of target heart rate cycle waveforms in the continuous heart rate cycle waveforms, the true motion abnormality probability of the continuous heart rate cycle waveforms is obtained. Finally, based on the true motion abnormality probability, the weights of the waveform features of the continuous heart rate cycle waveforms are obtained. In this way, the samples in the sample set are divided into corresponding weights according to the motion state. The weights of the heart rate cycle waveforms affected by motion are set smaller, which reduces the impact on the training of the lightweight neural network model, thereby improving the training accuracy of the arrhythmia recognition model and the accuracy of arrhythmia recognition. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method for identifying arrhythmias based on a lightweight neural network model, provided in one embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the P wave, QRS complex and T wave in a heartbeat cycle waveform provided in an embodiment of the present invention;

[0041] Figure 3 This is a flowchart of the process for obtaining abnormal heartbeat fluctuation clusters according to an embodiment of the present invention;

[0042] Figure 4 This is a flowchart of obtaining the fluctuation anomaly degree of each wave in a cluster according to an embodiment of the present invention;

[0043] Figure 5 This is a flowchart of the process for obtaining a target heartbeat cycle waveform according to an embodiment of the present invention;

[0044] Figure 6 This is a flowchart of obtaining the degree of sudden fluctuation of the first body indicator fluctuation curve provided in one embodiment of the present invention;

[0045] Figure 7 This is a flowchart of the process for obtaining the probability of real motion anomalies provided in one embodiment of the present invention;

[0046] Figure 8This is a flowchart illustrating one method of obtaining weights according to an embodiment of the present invention;

[0047] Figure 9 This is another flowchart of obtaining weights provided in one embodiment of the present invention;

[0048] Figure 10 This is a schematic diagram of the ECG signal of patient 1 provided in one embodiment of the present invention;

[0049] Figure 11 This is a schematic diagram of the ECG signal of patient 2 provided in one embodiment of the present invention;

[0050] Figure 12 This is a schematic diagram of the ECG signal of patient 3 provided in one embodiment of the present invention;

[0051] Figure 13 This is a schematic diagram of the ECG signal of patient 4 provided in one embodiment of the present invention;

[0052] Figure 14 This is a schematic diagram of the ECG signal of patient 5 provided in one embodiment of the present invention;

[0053] Figure 15 This is a schematic diagram of the ECG signal of patient 6 provided in one embodiment of the present invention;

[0054] Figure 16 This is a schematic diagram of the ECG signal of patient 7 provided in one embodiment of the present invention;

[0055] Figure 17 This is a schematic diagram of the ECG signal of patient 8 provided in one embodiment of the present invention;

[0056] Figure 18 This is a schematic diagram of the ECG signal of patient 9 provided in one embodiment of the present invention;

[0057] Figure 19 This is a schematic diagram of the ECG signal of patient 10 provided in one embodiment of the present invention;

[0058] Figure 20 This is a schematic diagram of the ECG signal of patient 11 provided in one embodiment of the present invention;

[0059] Figure 21 This is a schematic diagram of the ECG signal of patient 12 provided in one embodiment of the present invention;

[0060] Figure 22 This is a schematic diagram of a loss function provided in one embodiment of the present invention;

[0061] Figure 23This is a schematic diagram illustrating the changes in test accuracy and training accuracy provided in one embodiment of the present invention. Detailed Implementation

[0062] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0063] Unless otherwise defined, 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 invention pertains. All data and information collected in this application have been obtained with full consent, and the collection, use, and processing of such information must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0064] This embodiment provides a method for arrhythmia identification based on a lightweight neural network model, such as Figure 1 As shown, it includes the following steps:

[0065] Step 1: Divide the electrocardiogram fluctuation curve of the subject to be monitored into multiple heartbeat cycle waveforms, and cluster them to obtain multiple clusters, from which the abnormal heartbeat fluctuation cluster is obtained;

[0066] Step 2: Obtain the body indicator fluctuation curves at the same time as the heart rate cycle waveforms in the abnormal heart rate fluctuation cluster;

[0067] Step 3: Based on the waveform differences of adjacent body indicator fluctuation curves, the target heart rate cycle waveform is obtained. The target heart rate cycle waveform represents the heart rate cycle waveform affected by exercise.

[0068] Step 4: Combine the temporally continuous heartbeat cycle waveforms in the abnormal heartbeat fluctuation cluster into a continuous heartbeat cycle waveform, and obtain the true motion abnormality probability of the continuous heartbeat cycle waveform based on the number and density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform.

[0069] Step 5: Based on the actual motion abnormality probability, obtain the weights of the waveform features of the continuous heartbeat cycle waveform.

[0070] The following detailed explanation of each step, in conjunction with the accompanying drawings, is provided.

[0071] Step 1: Divide the electrocardiogram fluctuation curve of the subject to be monitored into multiple heartbeat cycle waveforms, and cluster them to obtain multiple clusters, from which the abnormal heartbeat fluctuation cluster is obtained.

[0072] The subject to be monitored is typically a patient. In a specific application scenario, the patient wears a wearable device (such as a smartwatch or smart bracelet with ECG acquisition function, or a professional ECG acquisition device) in a hospital or at home to collect ECG fluctuation curves within a monitoring period. The length of the monitoring period is set according to actual needs. Since this embodiment uses the patient's ECG fluctuation curves as training samples to train a lightweight neural network model and obtain an arrhythmia recognition model, the monitoring period can be set longer to improve the accuracy of the training samples. Furthermore, this embodiment can acquire one or multiple ECG fluctuation curves. The following example uses one ECG fluctuation curve.

[0073] Simultaneously, wearable devices also collect relevant physical indicators from the patient. These indicators specifically fluctuate during patient movement, such as respiratory rate, body temperature, and blood pressure. The patient's physical indicators are acquired within a pre-set sampling frequency over a monitoring period, thus obtaining a fluctuation curve for these indicators. In this embodiment, respiratory rate is used as an example; therefore, the respiratory rate is acquired within the pre-set sampling frequency over a monitoring period, resulting in a respiratory rate fluctuation curve.

[0074] Since the electrocardiogram (ECG) waveform includes multiple heartbeat cycles, it is divided into waveforms representing multiple heartbeat cycles. Each heartbeat cycle waveform consists of the P wave, QRS complex, and T wave, etc. Figure 2 As shown. Figure 2 The waveform also includes the PR interval, ST segment, and QT interval. Accordingly, in this embodiment, the types of waves included in the heartbeat cycle waveform include the P wave, QRS complex, and T wave. In an exemplary embodiment, existing P wave detection methods, QRS complex detection methods, and T wave detection methods can be used to detect the P wave, QRS complex, and T wave in each heartbeat cycle waveform, and the corresponding feature values ​​for each wave can be obtained. The feature values ​​for each wave are shown in Table 1.

[0075] Table 1

[0076]

[0077] Then, multiple heart rate cycle waveforms are clustered to obtain multiple clusters. In an exemplary embodiment, the K-means clustering method is used to cluster the heart rate cycle waveforms. The DTW (Dynamic Time Warping) distance between pairwise heart rate cycle waveforms is obtained, and the K-means clustering method is used to cluster all heart rate cycle waveforms based on the DTW distance to obtain multiple clusters. The number of clusters is determined by the silhouette coefficient. The more clusters obtained, the more times the patient's heart rate fluctuates in the electrocardiogram fluctuation curve.

[0078] Please refer to Table 2, which shows the characteristic values ​​for each wave in multiple patients.

[0079] Table 2

[0080]

[0081] Please see Figure 10 , Figure 10 This is a schematic diagram of the ECG signal for patient 1; please refer to [link / reference]. Figure 11 , Figure 11 A schematic diagram of the ECG signal for patient 2; please refer to [link / reference]. Figure 12 , Figure 12 A schematic diagram of the ECG signal for patient 3; please refer to [link / reference]. Figure 13 , Figure 13 A schematic diagram of the ECG signal for patient 4; please refer to [link / reference]. Figure 14 , Figure 14 A schematic diagram of the ECG signal for patient 5; please refer to [link / reference]. Figure 15 , Figure 15 A schematic diagram of the ECG signal for patient 6; please refer to [link / reference]. Figure 16 , Figure 16 A schematic diagram of the ECG signal for patient 7; please refer to [link / reference]. Figure 17 , Figure 17 A schematic diagram of the ECG signal for patient 8; please refer to [link / reference]. Figure 18 , Figure 18 A schematic diagram of the ECG signal for patient 9; please refer to [link / reference]. Figure 19 , Figure 19 A schematic diagram of the ECG signal for patient 10; please refer to [link / reference]. Figure 20 , Figure 20 A schematic diagram of the ECG signal for patient 11; please refer to [link / reference]. Figure 21 , Figure 21 A schematic diagram of the ECG signal for patient 12; please refer to [link / reference]. Figure 22 , Figure 22 This is a schematic diagram of the loss function; please refer to [link / reference]. Figure 23 , Figure 23 This is a diagram illustrating the changes in test accuracy and training accuracy.

[0082] Finally, based on the normality of the characteristic values ​​of each wave in the cluster, an abnormal heartbeat fluctuation cluster is determined from the resulting multiple clusters. In an exemplary embodiment, such as Figure 3 As shown, the following is a specific process for obtaining clusters of abnormal heartbeat fluctuations:

[0083] Step 1-1: Obtain the degree of fluctuation anomaly for each wave in the cluster.

[0084] Taking each wave as the analysis object, the greater the difference between the eigenvalue of each wave in the cluster and the corresponding preset normal eigenvalue range, the greater the probability of abnormal fluctuation of that wave. Correspondingly, such as... Figure 4 As shown, the following is a specific process for obtaining the degree of fluctuation anomaly of each wave in the cluster:

[0085] Step 1-1-1: Obtain the feature values ​​of various waves in the waveform diagram of each heartbeat cycle in the cluster.

[0086] For ease of explanation, let's define the first cluster as any cluster, the first heartbeat cycle waveform as any heartbeat cycle waveform within the first cluster, and the first wave as any wave within the first heartbeat cycle waveform. We then obtain the eigenvalues ​​of the first wave in the first heartbeat cycle waveform within the first cluster.

[0087] Step 1-1-2: Calculate the average value of the characteristic values ​​of the same wave in all heartbeat cycle waveforms in the cluster to obtain the cluster single wave characteristic value of each wave in the cluster.

[0088] Since the first cluster includes multiple heartbeat cycle waveforms, the feature values ​​of the first wave in each heartbeat cycle waveform in the first cluster are obtained. The average of these feature values ​​is then calculated as the cluster single-wave feature value for the first wave. This yields the cluster single-wave feature value for each wave type in the first cluster.

[0089] Step 1-1-3: Based on the difference between the cluster single-wave characteristic value of each wave in the cluster and the corresponding preset normal characteristic value range, the fluctuation anomaly degree of each wave in the cluster is obtained.

[0090] It should be understood that each type of wave has a normal characteristic value range. Therefore, a normal characteristic value range is preset for each type of wave. Taking the first type of wave as an example, when it is within the normal characteristic value range corresponding to the first type of wave, it indicates that the first type of wave is normal. When it is not within the normal characteristic value range corresponding to the first type of wave, it indicates that the first type of wave is abnormal. Moreover, the greater the difference between the first type of wave and the normal characteristic value range, the greater the degree of abnormality of the first type of wave.

[0091] Therefore, based on the difference between the cluster single-wave characteristic value of each wave in the first cluster and the corresponding preset normal characteristic value range, the fluctuation anomaly degree of each wave in the first cluster is obtained.

[0092] For the first wave of the first cluster, if the cluster single-wave characteristic value of the first wave is within the normal characteristic value range corresponding to the first wave, then the fluctuation anomaly degree of the first wave of the first cluster is set to 0; otherwise, the difference characteristic of the fluctuation anomaly degree of the first wave is obtained. The process of obtaining the difference characteristic is as follows: if the cluster single-wave characteristic value of the first wave is greater than the upper limit of the normal characteristic value range corresponding to the first wave, the difference between the cluster single-wave characteristic value of the first wave and the upper limit of the normal characteristic value range corresponding to the first wave is calculated, and this difference is the difference characteristic; if the cluster single-wave characteristic value of the first wave is less than the lower limit of the normal characteristic value range corresponding to the first wave, the difference between the lower limit of the normal characteristic value range corresponding to the first wave and the cluster single-wave characteristic value of the first wave is calculated, and this difference is the difference characteristic.

[0093] Then, the difference feature of the first wave in the first cluster is normalized. The normalized result is the fluctuation anomaly degree of the first wave in the first cluster. The normalization method here is as follows: obtain the maximum and minimum values ​​of the difference features of the first wave in all clusters, and then use the maximum and minimum value normalization method to normalize the difference feature of the first wave in the first cluster.

[0094] It should be understood that when arrhythmias occur, abnormalities such as atrial fibrillation, ventricular fibrillation, and conduction block may occur, and different abnormalities will manifest differently on the electrocardiogram fluctuation curve. For example, atrial fibrillation is characterized by the disappearance of the P wave. Therefore, the above method is used to obtain the fluctuation abnormality sub-degree of each wave in each cluster.

[0095] Step 1-2: Merge the fluctuation anomaly sub-degrees of all wave types in the cluster to obtain the fluctuation anomaly degree of the cluster.

[0096] In an exemplary embodiment, the fluctuation anomaly sub-degrees of all wave types in the first cluster are merged. The average fluctuation anomaly sub-degree of all wave types in the first cluster is calculated as the fluctuation anomaly degree of the first cluster. The fluctuation anomaly degree of each cluster is obtained using the above method.

[0097] Steps 1-3: Select clusters with fluctuation abnormality levels greater than or equal to the preset fluctuation abnormality level threshold as heartbeat fluctuation abnormality clusters.

[0098] The degree of fluctuation anomaly reflects the anomalousness of each wave in a cluster; the higher the degree of fluctuation anomaly, the more severe the anomalousness of each wave in the cluster. In an exemplary embodiment, a preset fluctuation anomaly degree threshold is established. The value range of this preset fluctuation anomaly degree threshold is 0-1, and the specific value is set according to the actual situation. This preset fluctuation anomaly degree threshold is used to determine whether the fluctuation anomaly degree is low or high. Therefore, if a more reliable fluctuation anomaly degree judgment logic is required, the preset fluctuation anomaly degree threshold can be set slightly smaller, so that more clusters meet the judgment logic. This embodiment uses 0.3 as an example.

[0099] The degree of fluctuation abnormality in each cluster is compared with a preset fluctuation abnormality threshold. Clusters with fluctuation abnormality levels greater than or equal to the preset threshold are selected as the heartbeat fluctuation abnormality clusters. Abnormal fluctuations in the heartbeat cycle waveform within the heartbeat fluctuation abnormality clusters may indicate arrhythmia.

[0100] Step 2: Obtain the body indicator fluctuation curves at the same time as the heart rate cycle waveforms in the abnormal heart rate fluctuation cluster.

[0101] It should be understood that there may be only one or more clusters of abnormal heartbeat fluctuations. The following example uses any one cluster of abnormal heartbeat fluctuations. If there are multiple clusters of abnormal heartbeat fluctuations, the data processing procedure for each cluster is the same as that for a single cluster.

[0102] Abnormal heart rate fluctuations may be caused by factors such as exercise. Therefore, it is necessary to combine the patient's physical indicator fluctuation curves to determine whether the abnormal fluctuations are due to exercise or other factors. Since the electrocardiogram fluctuation curves and physical indicator fluctuation curves are acquired synchronously, obtaining the abnormal heart rate fluctuation clusters yields the acquisition time of each heart rate cycle waveform within the cluster. Therefore, the physical indicator fluctuation curves acquired at the same time as each heart rate cycle waveform within the abnormal heart rate fluctuation clusters are obtained. The number of heart rate cycle waveforms in the abnormal heart rate fluctuation clusters is the same as the number of physical indicator fluctuation curves, and they correspond one-to-one. Each heart rate cycle waveform has a physical indicator fluctuation curve acquired at the same time. Assuming that the abnormal heart rate fluctuation clusters include U heart rate cycle waveforms, then there are also U physical indicator fluctuation curves, and the u-th heart rate cycle waveform in the abnormal heart rate fluctuation cluster is acquired at the same time as the u-th physical indicator fluctuation curve, where u equals [1, U].

[0103] Step 3: Based on the waveform differences of adjacent body indicator fluctuation curves, the target heart rate cycle waveform is obtained. The target heart rate cycle waveform represents the heart rate cycle waveform affected by exercise.

[0104] Based on the waveform differences between adjacent bodily indicator fluctuation curves, it is determined whether the fluctuation of the bodily indicator is a sudden or continuous fluctuation, thereby determining the heart rate cycle waveform affected by exercise, i.e., the target heart rate cycle waveform. In an exemplary embodiment, such as... Figure 5 As shown, the following is a specific screening process for the target heart rate cycle waveform:

[0105] Step 3-1: Based on the waveform difference between the first physical indicator fluctuation curve and the second physical indicator fluctuation curve, obtain the degree of sudden fluctuation of the first physical indicator fluctuation curve.

[0106] For ease of explanation, the first body index fluctuation curve is defined as the body index fluctuation curve at the same time as any heartbeat cycle waveform in the abnormal heartbeat fluctuation cluster.

[0107] Because patients' physical indicators fluctuate during exercise, the segment of the physical indicator fluctuation curve that is the left adjacent to the first physical indicator fluctuation curve and of the same length is defined as the second physical indicator fluctuation curve. Since "left adjacent" means earlier than the first physical indicator fluctuation curve in time, the second physical indicator fluctuation curve is collected earlier than the first physical indicator fluctuation curve, and the start time of the first physical indicator fluctuation curve is the end time of the second physical indicator fluctuation curve. Furthermore, the first and second physical indicator fluctuation curves have the same length. For example, the start and end times of the second physical indicator fluctuation curve are 10:01:00 and 10:02:00, respectively, and the start and end times of the first physical indicator fluctuation curve are 10:02:00 and 10:03:00, respectively. The length of both the first and second physical indicator fluctuation curves is 1 minute. It should be understood that if the first physical indicator fluctuation curve does not have a preceding physical indicator fluctuation curve in time (i.e., it does not have a left adjacent physical indicator fluctuation curve), it will not participate in the acquisition of subsequent fluctuation suddenness; only the fluctuation suddenness of other physical indicator fluctuation curves with left adjacent values ​​will be acquired.

[0108] Based on the waveform difference between the first and second bodily index fluctuation curves, the degree of sudden fluctuation of the first bodily index fluctuation curve is obtained. The greater the waveform difference between the two, the greater the degree of sudden fluctuation of the first bodily index fluctuation curve.

[0109] In one exemplary embodiment, such as Figure 6 As shown, the following is a specific process for obtaining the degree of sudden fluctuation in the fluctuation curve of the first physical indicator:

[0110] Step 3-1-1: Obtain the difference between the mean data of the first body indicator fluctuation curve and the mean data of the second body indicator fluctuation curve.

[0111] Since both the first and second physical indicator fluctuation curves are composed of multiple data points, and taking respiratory rate as an example, both are composed of respiratory rate values ​​at various sampling times. Therefore, we calculate the average of the data points included in the first physical indicator fluctuation curve to obtain the first data mean, and then calculate the average of the data points included in the second physical indicator fluctuation curve to obtain the second data mean. We then obtain the difference between the first and second data means, specifically the absolute value of the difference between them. A larger difference in data means indicates a greater difference between the first and second physical indicator fluctuation curves.

[0112] Step 3-1-2: Obtain the Euclidean distance between the first data set and the second data set.

[0113] Obtain the fluctuation range of the first physical indicator fluctuation curve, that is, obtain the maximum and minimum data values ​​in the first physical indicator fluctuation curve, and form the first data group with the maximum and minimum data values ​​in the first physical indicator fluctuation curve; obtain the fluctuation range of the second physical indicator fluctuation curve, that is, obtain the maximum and minimum data values ​​in the second physical indicator fluctuation curve, and form the second data group with the maximum and minimum data values ​​in the second physical indicator fluctuation curve.

[0114] To obtain the Euclidean distance between the first data set and the second data set, specifically: calculate the difference between the minimum data value in the first data set and the minimum data value in the second data set, calculate the difference between the maximum data value in the first data set and the maximum data value in the second data set, and then calculate the square root of the sum of the squares of these two differences. The result is the Euclidean distance between the first data set and the second data set.

[0115] The greater the Euclidean distance between the first and second data sets, the greater the difference between the first and second body index fluctuation curves.

[0116] Step 3-1-3: Based on the difference in the mean of the data and the Euclidean distance, obtain the degree of sudden fluctuation of the first physical indicator fluctuation curve.

[0117] As shown in the above steps, the greater the difference between the mean values ​​of the first and second data groups, the greater the difference between the fluctuation curves of the first and second physical indicators, and the greater the suddenness of fluctuation in the first physical indicator's fluctuation curve. The difference in mean values ​​is directly proportional to the degree of sudden fluctuation. Similarly, the greater the Euclidean distance between the first and second data groups, the greater the difference between the fluctuation curves of the first and second physical indicators, and the greater the suddenness of fluctuation in the first physical indicator's fluctuation curve. The Euclidean distance is also directly proportional to the degree of sudden fluctuation. Therefore, based on the difference in mean values ​​and the Euclidean distance between the first and second data groups, the degree of sudden fluctuation in the first physical indicator's fluctuation curve can be determined.

[0118] In an exemplary embodiment, the product of the difference between the data means of the first data group and the second data group and their Euclidean distance is calculated, and then the product is normalized. The normalized result is the degree of sudden fluctuation of the first physical indicator fluctuation curve. The normalization method here can be: obtaining the maximum and minimum values ​​of the products of the difference between the data means of the physical indicator fluctuation curves corresponding to all heart rate cycle waveforms in the abnormal heart rate fluctuation cluster, and then using the maximum and minimum value normalization method to normalize the product corresponding to the first physical indicator fluctuation curve.

[0119] Using the above process, the degree of sudden fluctuation of the body index fluctuation curve corresponding to the heartbeat cycle waveform in each heartbeat cycle abnormal cluster is obtained, that is, the degree of sudden fluctuation of the heartbeat cycle waveform in each heartbeat cycle abnormal cluster is obtained.

[0120] It should be understood that if multiple types of body indicator fluctuation curves are collected, such as respiratory rate and body temperature, the degree of fluctuation suddenness of each body indicator fluctuation curve is obtained separately, and then the average value is calculated to obtain the final degree of fluctuation suddenness.

[0121] Step 3-2: Use the heartbeat cycle waveform corresponding to the fluctuation suddenness level that is greater than or equal to the preset fluctuation suddenness level threshold as the target heartbeat cycle waveform in the heartbeat fluctuation abnormal cluster.

[0122] The greater the degree of fluctuation suddenness, the more likely it is an abnormality caused by movement. A preset threshold for the degree of fluctuation suddenness is established, with a value ranging from 0 to 1. The specific value of this preset threshold is set according to the actual situation. This preset threshold is used to filter out heart rate cycle waveforms of movement abnormalities. If a more stringent judgment logic is required, the preset threshold can be set slightly lower to make it easier to obtain heart rate cycle waveforms that meet the requirements and avoid omissions. This embodiment uses 0.5 as an example.

[0123] The fluctuation abruptness of each heartbeat cycle waveform in the abnormal heartbeat fluctuation cluster is compared with a preset fluctuation abruptness threshold. Heartbeat cycle waveforms with fluctuation abruptness greater than or equal to the preset threshold are selected and used as target heartbeat cycle waveforms in the abnormal heartbeat fluctuation cluster. The target heartbeat cycle waveform represents the fluctuation anomaly caused by motion.

[0124] Step 4: Combine the temporally continuous heartbeat cycle waveforms in the abnormal heartbeat fluctuation cluster into a continuous heartbeat cycle waveform. Based on the number and density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform, obtain the true motion abnormality probability of the continuous heartbeat cycle waveform.

[0125] Generally, exercise is a continuous process. Therefore, heart rate cycle waveforms that are sequentially continuous within the abnormal heart rate fluctuation cluster are combined into a continuous heart rate cycle waveform diagram. In an exemplary embodiment, all heart rate cycle waveforms obtained from the electrocardiogram fluctuation curve can be sorted sequentially to obtain the sequence number of each heart rate cycle waveform. Then, heart rate cycle waveforms with consecutive sequence numbers within the abnormal heart rate fluctuation cluster are combined into a continuous heart rate cycle waveform diagram. For example, if the sequence numbers of all heart rate cycle waveforms are 1-10, and the sequence numbers of each heart rate cycle waveform in the abnormal heart rate fluctuation cluster are 2, 3, 6, 7, and 9, then the second and third heart rate cycle waveforms are combined into a continuous heart rate cycle waveform diagram, and the sixth and seventh heart rate cycle waveforms are combined into a continuous heart rate cycle waveform diagram.

[0126] It should be understood that for isolated heartbeat cycle waveforms in a cluster of abnormal heartbeat fluctuations, that is, when there are no adjacent heartbeat cycle waveforms before and after it in the cluster of abnormal heartbeat fluctuations, then we do not divide them into continuous heartbeat cycle waveforms, nor do we obtain the true motion abnormality probability of continuous heartbeat cycle waveforms, such as the 9th heartbeat cycle waveform in the example above.

[0127] Then, based on the number and density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform fluctuations, the true motion abnormality probability of the continuous heartbeat cycle waveform is obtained.

[0128] The density characterizes the clustering of target heartbeat cycle waveforms in a continuous heartbeat cycle waveform graph. In an exemplary embodiment, the distance between any two adjacent target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graph is obtained. This distance can be the difference in the index of any two adjacent target heartbeat cycle waveforms. For example, if the indexes of two adjacent target heartbeat cycle waveforms are 2 and 4, then the distance between these two adjacent target heartbeat cycle waveforms is 4-2=2. Then, the average distance between all pairs of adjacent target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graph is calculated as the density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graph.

[0129] In one exemplary embodiment, such as Figure 7 As shown, the following is a specific process for obtaining the probability of true motion anomalies:

[0130] Step 4-1: Obtain the percentage of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graphs.

[0131] For any continuous heartbeat cycle waveform, obtain the number of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform and the total number of heartbeat cycle waveforms in the continuous heartbeat cycle waveform. Calculate the ratio of the number of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform to the total number of heartbeat cycle waveforms in the continuous heartbeat cycle waveform to obtain the proportion of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform.

[0132] Step 4-2: Based on the proportion and density of the number of waves, obtain the true probability of motion abnormalities in the continuous heartbeat cycle waveform.

[0133] The larger the proportion, the more target heartbeat cycle waveforms there are in the continuous heartbeat cycle waveform diagram. Since motion is usually a continuous process, the possibility of motion abnormalities in the continuous heartbeat cycle waveform diagram is greater, that is, the probability of actual motion abnormality is greater. Therefore, the probability of actual motion abnormality is directly proportional to the proportion. The smaller the density, the smaller the distance between the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform diagram. Since motion is usually a continuous process, the possibility of motion abnormalities in the continuous heartbeat cycle waveform diagram is greater, that is, the probability of actual motion abnormality is greater. Therefore, the probability of actual motion abnormality is inversely proportional to the density.

[0134] In one exemplary embodiment, the density is negatively correlated and normalized, for example, by exp(-x), where x is the object to be negatively correlated and normalized, and exp is an exponential function with the natural constant e as its base. Then, the product of the quantity proportion and the negatively correlated and normalized density is calculated, which is the true probability of motion abnormality in the continuous heartbeat cycle waveform.

[0135] Using the above process, the true motion abnormality probability of each continuous heartbeat cycle waveform is obtained.

[0136] It should be understood that if there is no target heartbeat cycle waveform in the continuous heartbeat cycle waveform, the actual motion abnormality probability of the continuous heartbeat cycle waveform is 0. Furthermore, if there is only one target heartbeat cycle waveform in the continuous heartbeat cycle waveform, the actual motion abnormality probability of the continuous heartbeat cycle waveform is also set to 0.

[0137] Step 5: Based on the actual motion abnormality probability, obtain the weights of the waveform features of the continuous heartbeat cycle waveform.

[0138] Since a higher probability of actual motion anomalies increases the likelihood of motion influence on the continuous heartbeat cycle waveform, a smaller weight needs to be assigned to it to reduce its impact on the accuracy of lightweight neural network model training results. Therefore, the probability of actual motion anomalies in the continuous heartbeat cycle waveform is inversely proportional to the weight of its waveform features.

[0139] If only one electrocardiogram (ECG) fluctuation curve is obtained, the weights of the waveform features of each consecutive heartbeat cycle waveform are derived based on the true motion abnormality probability of each consecutive heartbeat cycle waveform within that ECG fluctuation curve. In an exemplary embodiment, such as... Figure 8 As shown, the following is a specific acquisition process:

[0140] Step 5-1: Determine the importance of the continuous heartbeat cycle waveform based on the actual motion abnormality probability of the continuous heartbeat cycle waveform.

[0141] For ease of explanation, the first continuous heartbeat cycle waveform is set to any single continuous heartbeat cycle waveform. The importance of the first continuous heartbeat cycle waveform is determined based on the actual motion anomaly probability of that waveform. The importance is inversely proportional to the actual motion anomaly probability. In an exemplary embodiment, the difference between the value 1 and the actual motion anomaly probability of the first continuous heartbeat cycle waveform is calculated; this difference represents the importance of the first continuous heartbeat cycle waveform.

[0142] Step 5-2: Obtain the weights of the waveform features of the continuous heartbeat cycle waveform based on their importance.

[0143] The weights of the waveform features of the continuous heartbeat cycle waveform are determined based on their importance. The higher the importance, the higher the weight. In an exemplary embodiment, the importance of the obtained continuous heartbeat cycle waveform is used as the weight of the waveform features of the continuous heartbeat cycle waveform.

[0144] If at least two electrocardiogram (ECG) fluctuation curves are obtained, then according to steps 1-4, the true motion abnormality probability of each continuous heartbeat cycle waveform in each ECG fluctuation curve is obtained. Then, based on the true motion abnormality probability of each continuous heartbeat cycle waveform in each ECG fluctuation curve, the weight of the waveform features of each continuous heartbeat cycle waveform in each ECG fluctuation curve is obtained. In an exemplary embodiment, such as... Figure 9 As shown, the following is a specific acquisition process:

[0145] Step 5-3: Obtain the number of clusters in each ECG fluctuation curve, and combine the actual motion abnormality probability of the continuous heartbeat cycle waveform in each ECG fluctuation curve to obtain the importance of the continuous heartbeat cycle waveform in each ECG fluctuation curve.

[0146] For ease of explanation, the first electrocardiogram fluctuation curve is used to represent any electrocardiogram fluctuation curve below.

[0147] The number of clusters in the first electrocardiogram fluctuation curve is obtained, and the importance of each continuous heartbeat cycle waveform in the first electrocardiogram fluctuation curve is obtained by combining the actual motion abnormality probability of each continuous heartbeat cycle waveform in the first electrocardiogram fluctuation curve.

[0148] Since a larger number of clusters indicates more changes in heart rate fluctuations in the first electrocardiogram (ECG) waveform, resulting in richer waveform features, and thus greater importance for identifying arrhythmias, the importance of the first ECG waveform is directly proportional to the number of clusters. The importance of each continuous heartbeat cycle waveform within the first ECG waveform is inversely proportional to the actual probability of motion abnormalities in that waveform.

[0149] In one exemplary embodiment, the total number of clusters in all electrocardiogram (ECG) fluctuation curves is obtained. The ratio of the number of clusters in the first ECG fluctuation curve to the total number of clusters is calculated as a cluster count feature in the first ECG fluctuation curve.

[0150] The product of the number of clusters in the first electrocardiogram fluctuation curve and the true motion abnormality probability of each continuous heartbeat cycle waveform in the first electrocardiogram fluctuation curve is calculated as the importance of each continuous heartbeat cycle waveform in the first electrocardiogram fluctuation curve.

[0151] Using the above process, the importance of each continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve is obtained.

[0152] Step 5-4: Based on the importance of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve, obtain the weight of the waveform characteristics of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve.

[0153] Based on the importance of each continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve, the weight of the waveform feature of each continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve is obtained. In an exemplary embodiment, the importance of each continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve is used as the weight of the waveform feature of each continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve.

[0154] The weights of the waveform features in the continuous heartbeat cycle waveform are used as the weights of the waveform features in each individual heartbeat cycle waveform within the continuous heartbeat cycle waveform. For each individual heartbeat cycle waveform in the electrocardiogram fluctuation curve other than the continuous heartbeat cycle waveforms, the weight is set to 1. This yields the weights of the waveform features for each heartbeat cycle waveform in the electrocardiogram fluctuation curve.

[0155] In one exemplary embodiment, in subsequent specific applications, 80% of all heartbeat cycle waveforms in all electrocardiogram fluctuation curves are used as the training set, and 20% are used as the test set.

[0156] The waveform features of each heartbeat cycle waveform in the training set are determined. Specifically, waveform features refer to the feature values ​​of each wave in the heartbeat cycle waveform. The waveform features of each heartbeat cycle waveform in the training set are multiplied by their corresponding weights to achieve a weighted sum. The weighted waveform features of each heartbeat cycle waveform are then used as the training set input into a lightweight neural network model for training. The loss function during training is the cross-entropy loss function. The model training process is a well-known technique and will not be detailed here. The trained model is denoted as the arrhythmia recognition model based on the lightweight neural network model. Please refer to [link to relevant documentation]. Figure 22 , Figure 22 This is a schematic diagram of the loss function; please refer to [link / reference]. Figure 23 , Figure 23 This diagram illustrates the changes in test accuracy and training accuracy. Figure 22 and Figure 23 The horizontal axis of the coordinate system represents the period, also known as eopch. One period is the process of all training samples undergoing a complete forward and backward propagation once.

[0157] Please refer to Table 3, which is the training data table;

[0158] Table 3

[0159]

[0160] When it is necessary to identify arrhythmias in patients in the future, the waveform characteristics of each wave in the patient's electrocardiogram fluctuation curve can be input into the arrhythmia identification model, and the arrhythmia identification results can be output.

[0161] This embodiment also provides a cardiac arrhythmia recognition system based on a lightweight neural network model, including: a memory and a processor; the memory is connected to the processor, and the memory is used to store program instructions; the processor is used to implement the steps in the above-described embodiment of the cardiac arrhythmia recognition method based on a lightweight neural network model when the program instructions are executed.

[0162] In one exemplary embodiment, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the embodiment of the arrhythmia identification method based on a lightweight neural network model.

[0163] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0164] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for identifying cardiac arrhythmias based on a lightweight neural network model, characterized in that, include: The electrocardiogram (ECG) fluctuation curve of the subject under monitoring is divided into multiple heartbeat cycle waveforms and clustered to obtain multiple clusters, from which the abnormal heartbeat fluctuation cluster is obtained. The process of obtaining the abnormal heartbeat fluctuation cluster is as follows: the fluctuation abnormality sub-degree of each wave in the cluster is obtained; the wave types include P wave, QRS complex, and T wave; the fluctuation abnormality sub-degrees of all waves in the cluster are merged to obtain the fluctuation abnormality degree of the cluster; the clusters corresponding to the fluctuation abnormality degree greater than or equal to the preset fluctuation abnormality degree threshold are taken as the abnormal heartbeat fluctuation clusters. Obtain the body indicator fluctuation curves at the same time as the heartbeat cycle waveforms in the abnormal heartbeat fluctuation cluster; Based on the waveform differences of adjacent body indicator fluctuation curves, a target heart rate cycle waveform is obtained by screening. The target heart rate cycle waveform represents the heart rate cycle waveform affected by exercise. The heartbeat cycle waveforms that are sequentially continuous in the heartbeat fluctuation abnormal cluster are combined into a continuous heartbeat cycle waveform. The true motion abnormality probability of the continuous heartbeat cycle waveform is obtained based on the number and density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform. The weights of waveform features in the continuous heartbeat cycle waveform are obtained based on the actual motion abnormality probability.

2. The arrhythmia identification method based on a lightweight neural network model as described in claim 1, characterized in that, The process of obtaining the wave anomaly degree for each wave in the cluster includes: Obtain the feature values ​​of various waves in the waveform diagram of each heartbeat cycle in the cluster; Calculate the average value of the characteristic values ​​of the same wave in all heartbeat cycle waveforms in the cluster to obtain the cluster single wave characteristic value of each wave in the cluster. Based on the difference between the cluster single-wave characteristic value of each wave in the cluster and the corresponding preset normal characteristic value range, the fluctuation anomaly degree of each wave in the cluster is obtained.

3. The arrhythmia identification method based on a lightweight neural network model as described in claim 1, characterized in that, The screening process for the target heart rate cycle waveform includes: The degree of sudden fluctuation of the first physical indicator fluctuation curve is obtained based on the waveform difference between the first physical indicator fluctuation curve and the second physical indicator fluctuation curve; the first physical indicator fluctuation curve is the physical indicator fluctuation curve at the same time as any heartbeat cycle waveform in the abnormal heartbeat fluctuation cluster, and the second physical indicator fluctuation curve is the physical indicator fluctuation curve segment that is adjacent to the first physical indicator fluctuation curve on the left and of the same length. The heartbeat cycle waveform corresponding to a fluctuation suddenness level greater than or equal to a preset fluctuation suddenness level threshold is used as the target heartbeat cycle waveform in the heartbeat fluctuation anomaly cluster.

4. The arrhythmia identification method based on a lightweight neural network model as described in claim 3, characterized in that, The process of obtaining the degree of sudden fluctuation in the primary physical indicator fluctuation curve includes: Obtain the difference between the average data of the first body indicator fluctuation curve and the average data of the second body indicator fluctuation curve; Obtain the Euclidean distance between the first data set and the second data set. The first data set consists of the maximum and minimum data in the first body index fluctuation curve, and the second data set consists of the maximum and minimum data in the second body index fluctuation curve. Based on the difference in the mean of the data and the Euclidean distance, the degree of sudden fluctuation of the first body indicator fluctuation curve is obtained, and the degree of sudden fluctuation is proportional to both the difference in the mean of the data and the Euclidean distance.

5. The arrhythmia identification method based on a lightweight neural network model as described in claim 1, characterized in that, The process of obtaining the density includes: obtaining the distance between any two adjacent target heartbeat cycle waveforms in the continuous heartbeat cycle waveform diagram, and calculating the average value of the distance to obtain the density of the target heartbeat cycle waveforms in the continuous heartbeat cycle waveform diagram.

6. The arrhythmia identification method based on a lightweight neural network model as described in claim 5, characterized in that, The process of obtaining the probability of true motion anomalies includes: Obtain the percentage of target heartbeat cycle waveforms in the continuous heartbeat cycle waveform graphs; Based on the quantity ratio and density, the true motion abnormality probability of the continuous heartbeat cycle waveform is obtained. The true motion abnormality probability is directly proportional to the quantity ratio and inversely proportional to the density.

7. The arrhythmia identification method based on a lightweight neural network model as described in claim 1, characterized in that, The process of obtaining the weights of waveform features in a continuous heartbeat cycle waveform includes: The importance of the continuous heartbeat cycle waveform is obtained from the actual motion abnormality probability of the continuous heartbeat cycle waveform, and the importance is inversely proportional to the actual motion abnormality probability. The weights of waveform features in a continuous heartbeat cycle waveform are determined based on their importance.

8. The arrhythmia identification method based on a lightweight neural network model as described in claim 1, characterized in that, If multiple electrocardiogram (ECG) fluctuation curves of the subject to be monitored are obtained, the process of obtaining the weights of the waveform features of the continuous heartbeat cycle waveforms in each ECG fluctuation curve includes: The number of clusters in each electrocardiogram fluctuation curve is obtained, and the importance of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve is obtained by combining the actual motion abnormality probability of the continuous heartbeat cycle waveform in each electrocardiogram fluctuation curve. The importance is directly proportional to the number of clusters and inversely proportional to the actual motion abnormality probability. The weights of the waveform characteristics of the continuous heartbeat cycle waveforms in each electrocardiogram fluctuation curve are obtained based on the importance of the waveforms in each electrocardiogram fluctuation curve.

9. A cardiac arrhythmia recognition system based on a lightweight neural network model, characterized in that it includes: Memory and processor; The memory is connected to the processor; The memory is used to store program instructions; The processor is configured to implement the arrhythmia identification method based on a lightweight neural network model as described in any one of claims 1-8 when program instructions are executed.

Images