阅读说明：本技术 一种有损压缩心电信号房颤筛查系统 (Lossy compression electrocardiosignal atrial fibrillation screening system ) 是由 张宏坡 董忠仁 孙梦雅 王震 于 2020-06-29 设计创作，主要内容包括：本发明提供了一种有损压缩心电信号房颤筛查系统。本发明技术方案首先使用稀疏二进制观测矩阵对原始心电信号进行随机投影,将原始心电数据从高维空间转换到低维空间完成压缩,然后,使用卷积神经网络对压缩后的心电信号进行房颤分类。本发明提出一种压缩感知技术与卷积神经网络结合的新颖房颤筛查方法(CS-CNN),直接对有损压缩心电信号房颤筛查,有效且高效地提高了基于心电压缩感知数据的房颤检出率。(The invention provides a lossy compression electrocardiosignal atrial fibrillation screening system. According to the technical scheme, firstly, a sparse binary observation matrix is used for randomly projecting original electrocardiosignals, original electrocardio data are converted into a low-dimensional space from a high-dimensional space to be compressed, and then a convolutional neural network is used for carrying out atrial fibrillation classification on the compressed electrocardiosignals. The invention provides a novel atrial fibrillation screening method (CS-CNN) combining a compression sensing technology and a convolutional neural network, which is used for directly screening the atrial fibrillation of a lossy compression electrocardiosignal and effectively and efficiently improving the atrial fibrillation detection rate based on electrocardio compression sensing data.)

1. The lossy compression electrocardiosignal atrial fibrillation screening system comprises a wearable device end and a remote server end, and is characterized in that: the wearable equipment end carries out random projection on the original electrocardiosignals according to a compression perception theory, and converts the original electrocardio data from a high-dimensional space to a low-dimensional space to complete compression; the remote server side firstly performs normalization processing on a compression signal sent by the wearable equipment side, then inputs the processed data into a pre-established atrial fibrillation screening model, and outputs a screening result of the atrial fibrillation signal; the screening method comprises the following specific steps:

step 1: the wearable equipment end carries out random projection on the original electrocardiosignal to obtain a compressed signal Y; the process is as follows:

wherein the content of the first and second substances,

y⁽ⁱ⁾is the cardiac electrical signal after compression and is,

x⁽ⁱ⁾is the i-th raw ECG signal,

is an observation matrix with dimensions n × m;

step 2: the wearable device side transmits the compressed data to a remote server side;

and step 3: the server side carries out preprocessing operations such as data enhancement, standardization and the like on the compressed signal; the process is as follows:

wherein the content of the first and second substances,

s⁽ⁱ⁾is the data after the enhancement of the data,

is the data after normalization;

and 4, step 4: the server inputs the preprocessed data into a pre-established atrial fibrillation signal screening model and outputs a screening result of the atrial fibrillation signals; the process is as follows:

wherein the content of the first and second substances,

theta is a relevant parameter in the convolutional neural network,

d is the number of beat classes (D ═ 2, output atrial fibrillation or non-atrial fibrillation),

2. The system for screening atrial fibrillation according to claim 1, wherein the system further comprises: the whole one-dimensional convolutional neural network model consists of four convolutional layers and two pooling layers, wherein the two convolutional layers are in one group, and the groups are separated by the pooling layers. The first group uses 32 convolution kernels, the convolution kernel size is 3, and the step size is 1; the second group uses 64 convolution kernels, with a convolution kernel size of 3 and a step size of 1. In addition, the two groups adopt the maximum pooling layer with the size of 2 and the step length of 2. After pooling, through Flatten, the two fully-connected layers used 512 and 256 neurons respectively, and finally exported through the softmax layer.

3. The system for screening atrial fibrillation according to claim 1, wherein the system further comprises: in the training process of the atrial fibrillation signal screening model, a Drop out layer is connected to the back of the pooling layer and the back of the Flatten layer, and the ratio of the Drop out layer to the pooling layer is 0.2 and 0.3 respectively; adam is used for optimization, the initial learning rate is set to 0.0009, and the model parameters are updated in each loop iteration.

4. The system for screening atrial fibrillation according to claim 1, wherein the system further comprises: and the wearable equipment end transmits the compressed data subjected to random projection to a remote server end through a wireless transmitting module.

5. The system for screening atrial fibrillation according to claim 1, wherein the system further comprises: the wearable equipment terminal collects human body electrocardio data.

Technical Field

The invention belongs to the technical field of electrocardio monitoring, and particularly relates to a lossy compression electrocardiosignal atrial fibrillation screening system.

Background

Atrial Fibrillation (AF) is the most common persistent arrhythmic disease and is investigated to have an AF probability of 37.1% in people older than 55 years of age and subject to genetic and clinical risk factor burden. According to clinical analysis, AF greatly increases the probability of stroke, and the proportion also gradually increases with age, and also increases the economic burden on the medical system.

Electrocardiography is a non-invasive technique for observing the heart condition of a patient, and has become a standard medical examination means for clinical application of patients with cardiovascular diseases. However, the conventional electrocardiographic detection method has some limitations. The patient must be under the care of a medical professional for up to 24 hours of electrocardiographic collection. In recent years, learners prove that the automatic diagnosis of the ECG is not limited to clinical application, and the whole process from electrocardio acquisition to electrocardio analysis can be completed in the home of a user through wearable equipment, so that the problem of electrocardio dynamic monitoring is greatly relieved. At present, wearable equipment needs to face the problem of power consumption for acquiring electrocardiosignals. It was investigated that the charge of a common lithium battery would be exhausted if the transmission was continued for several hours at a transmission rate of 250 kb/s. The most direct mode for reducing power consumption compresses the electrocardio data and then transmits the compressed electrocardio data, so that the task of monitoring the electrocardio for a longer time is completed under the condition of not increasing the battery capacity of the wearable equipment.

In the field of signal processing, as a breakthrough technology for breaking the nyquist sampling theorem, Compressed Sensing (CS) attracts attention of scholars from birth and is applied to a remote electrocardio monitoring framework for electrocardio signal acquisition, transmission and compression. Studies in the literature indicate that Compressive Sensing (CS) is attractive for long-term Electrocardiogram (ECG) remote monitoring, and can extend the useful life of resource-limited wireless wearable sensors.

On one hand, conventional atrial fibrillation signal detection algorithms are analyzed on raw signals, and are not combined with low-power consumption research on wearable devices. On the other hand, classification of the electrocardio compressed sensing data is based on a machine learning method for extracting features, and when long-term electrocardio data is processed, the features still need to be extracted manually to reduce data dimensionality. Meanwhile, the machine learning method is time-consuming when classifying long-term electrocardiogram data. The long-term real-time monitoring of the electrocardiogram data and the atrial fibrillation screening task based on the low-power-consumption wearable device are still a challenge.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a lossy compression electrocardiosignal atrial fibrillation screening system. The specific scheme is as follows:

step 1: the wearable equipment end carries out random projection on the original electrocardiosignal to obtain a compressed signal Y; the process is as follows:

wherein the content of the first and second substances,

y⁽ⁱ⁾is the cardiac electrical signal after compression and is,

x⁽ⁱ⁾is the i-th raw ECG signal,

is an observation matrix with dimensions n × m;

step 2: the wearable device side transmits the compressed data to a remote server side;

and step 3: the server side carries out preprocessing operations such as data enhancement, standardization and the like on the compressed signal; the process is as follows:

wherein the content of the first and second substances,

s⁽ⁱ⁾is the data after the enhancement of the data,

is the average value of the average of the values,

is the standard deviation of the measured data to be measured,

is the data after normalization;

and 4, step 4: the server inputs the preprocessed data into a pre-established atrial fibrillation signal screening model and outputs a screening result of the atrial fibrillation signals; the process is as follows:

wherein the content of the first and second substances,

theta is a relevant parameter in the convolutional neural network,

d is the number of beat classes (D ═ 2, output atrial fibrillation or non-atrial fibrillation),

is the probability of the ith electrocardiogram record output by the CNN model.

Compared with the prior art, the atrial fibrillation screening method has outstanding substantive characteristics and remarkable progress, and particularly provides a novel atrial fibrillation screening model (CS-CNN) based on a compressive sensing technology, which is combined with a sparse binary observation matrix and a convolutional neural network, researches the atrial fibrillation screening performance of the model under different compression ratios, expands the application of compression learning by utilizing the convolutional neural network, directly screens the electrocardio compression sensing data for atrial fibrillation, ensures the prediction accuracy and simultaneously improves the calculation efficiency.

Drawings

Fig. 1 is the overall architecture of the atrial fibrillation signal screening system of the present invention.

Detailed Description

The technical solution of the present invention is further described in detail by the following embodiments.

As shown in fig. 1, a system for screening atrial fibrillation of a lossy compressed cardiac signal comprises:

compressing the original electrocardiosignals according to different compression ratios to obtain compressed data;

enhancing compressed sensing data by adopting an SMOTE method and normalizing the enhanced data;

and inputting the processed data into a CNN network model for atrial fibrillation screening.

Given a dataset X { (X)⁽¹⁾,z⁽¹⁾),…,(x⁽ⁱ⁾,z⁽ⁱ⁾),…,(x⁽ⁿ⁾,z⁽ⁿ⁾) -said raw signal compression is done by:

x⁽ⁱ⁾the duration of the ith electrocardiosignal is 1 s;

is the observation matrix fixed in the experiment, the dimension is n × m;

y⁽ⁱ⁾compressed sensing data obtained by observing matrix random projection.

The data enhancement of the compressed sensing data by adopting an SMOTE method and the normalization operation of the enhanced data are completed by the following steps:

s⁽ⁱ⁾is data after SMOTE method processing;

is the mean value;

is the standard deviation;

is the data after normalization.

The step of inputting the processed data into the CNN network model for atrial fibrillation screening is completed by the following steps:

θ is a relevant parameter in the convolutional neural network;

is the probability of the ith electrocardiogram record output by the CNN model.

The convolutional neural network CS-CNN network structure parameters provided by the invention are shown in table 1, and the network consists of four convolutional layers and two pooling layers, wherein the two convolutional layers are in a group, and the groups are separated by the pooling layers. The first group uses 32 convolution kernels, the convolution kernel size is 3, and the step size is 1; the second group uses 64 convolution kernels, with a convolution kernel size of 3 and a step size of 1. In addition, the two groups adopt the maximum pooling layer with the size of 2 and the step length of 2. The Drop out layer was attached behind both the pooling layer and the Flatten layer at ratios of 0.2 and 0.3, respectively. Then, 512 and 256 neurons are used by the two full-connection layers respectively, and finally, the output is obtained by applying the softmax layer. The optimizer used Adam with a learning rate of 0.0009.

TABLE 1 convolutional neural network parameters

It is not necessary to find the true global minimum during deep network model learning. Thus, when accuracy is lost during learning or enters a relatively flat region, Early stopping (Early stopping) techniques may be selected for use. Early stopping may be considered an unobtrusive form of normalization without affecting learning motivation, and this strategy may be used in conjunction with other normalization techniques. The technology can prevent the model from being over-fitted and accelerate the learning speed. By default, the maximum number of early stop periods may be set to 100 in view of efficiency and efficiency. The default settings for the model parameters are given in table 2.

TABLE 2 model parameter settings

Verification experiment

The data set used for the experiment was the MIT-BIH atrial fibrillation data set. MIT-BIH atrial fibrillation dataset was originally made by the Boston Israel hospital using a dynamic electrocardiography, with a main range of signal bandwidth of 0.1Hz-40 Hz. A total of 25 patients with atrial fibrillation (mainly paroxysmal fibrillation) had long-term cardiac electrical recordings. A total of 23 records included the 2-lead signal, record numbers 00735 and 03665 having only beat files and unrated qrs wave annotation files. Each recording lasted 10 hours, including two lead signals, with a sampling rate of 250 Hz. After expert labeling, four beat types are mainly included, AFIB (atrial fibrillation), AFL (atrial flutter), J (AV boundary rhythm) and N (for all other beats). We classify MIT-BIH atrial fibrillation datasets into two categories: and (3) recording non-atrial fibrillation (comprising N, AFL and J) and atrial fibrillation, and screening patients with atrial fibrillation by performing classification operation on the compressed data.

The method aims to detect the atrial fibrillation rhythm in the electrocardio compressed sensing signal of the wearable device. Since the selection of the observation matrix is crucial for compressed sensing, and at the same time, the observation matrix may affect the final screening result and the screening efficiency.

The Gaussian random observation matrix (GRM) and the sparse binary observation matrix (BSM) which are most commonly used in the electrocardio compression sensing direction are evaluated from the aspect of the execution time of a compression task. The detailed results are shown in Table 3. From the results, it can be seen that both observation matrix compression tasks are performed in a short time, but the BSM is faster. When CR is 90%, 1s of original ecg signals only needs 1.65ms to complete compression, which means that only 5.94s is needed to complete one hour of original signal compression process, and only 10% of original data needs to be transmitted.

TABLE 3 different observation matrixes are used for compressing the 1s electrocardiogram data

In addition to comparing GRM and BSM compression task performance times, we also investigated their impact on atrial fibrillation screening outcome performance. Table 4 is the screening results on the test set at different CRs using GRM and BSM compressed sensing data as model inputs.

It can be seen from table 4 that BSM is more suitable for cardiac signal compression than GRM. As CR increases, the values of the four indicators will slowly decrease. This indicates that the larger the CR, the worse the CNN model performs. This is because as the compression ratio increases, the original signal loses a lot of features, and CR 90% means that 90% of the features are lost. The BSM can ensure that the Acc value can still reach 88.08% under the condition that the original electrocardiosignal loses 90% of characteristics.

TABLE 4 screening results on test sets at different CRs using GRM and BSM compressed sensory data as model inputs

After the observation matrix is determined, the CS-CNN is compared to the overall screening results of the classical convolutional neural networks AlexNet, VGG-16 and GoogleNet. Specific experimental results are shown in table 5. The experimental results show that the F1 value of the CS-CNN model is 0.25% lower than the best AlexNet when CR is 80%, and the F1 value of the CS-CNN model is 1.6% lower than the best VGG-16 when CR is 90%, but the evaluation indexes of Sen, Spe, Acc, F1, etc. of CS-CNN are all best when CR is seven compression ratios such as 10% to 70%. The effectiveness of CS-CNN in compressed sensing electrocardiogram data screening is proved.

The real-time nature is worth discussing for atrial fibrillation screening problem, and in order to verify the screening efficiency of the CS-CNN proposed by us, the AF screening analysis is respectively carried out on clinical electrocardiogram data of 24-hour duration and sample (1s) by AlexNet, VGG-16 and GoogleLeNet. The results of the experiments are shown in Table 6. From the results it can be seen that CS-CNN was tested for AF screening at different CRs, with minimal time usage overall. For example, when the CR is 10%, 1.13s can complete AF screening and analysis of 24-hour cardiac electrical signals, and only 0.02s is needed to complete AF screening of one sample (1 s); when CR is 50%, AF screening and analysis of 24-hour electrocardiosignals can be completed within 0.91s, and AF screening of sample (1s) can be completed within only 0.02 s; when the CR is 90%, the AF screening and analysis of 24-hour electrocardiosignals can be completed within 0.71s, and the AF screening of one sample (1s) can be completed within only 0.01 s. Therefore, the CS-CNN meets the requirement of real-time atrial fibrillation screening on the whole.

TABLE 5 atrial fibrillation screening Performance under different CR for different models

TABLE 6 atrial fibrillation time (units: seconds) for different models at different CR screens

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention and not to limit it; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will understand that: modifications to the specific embodiments of the invention or equivalent substitutions for parts of the technical features may be made; without departing from the spirit of the present invention, it is intended to cover all aspects of the invention as defined by the appended claims.