阅读说明：本技术 一种基于心电压缩感知转置投影信号的房颤检测方法 (Atrial fibrillation detection method based on electrocardio compression sensing transposed projection signal ) 是由 张宏坡 谷红壮 董忠仁 孙梦雅 于 2021-05-20 设计创作，主要内容包括：本发明提供了一种基于心电压缩感知转置投影信号的房颤检测方法。本发明技术方案首先使用确定性观测矩阵对心电信号进行随机投影完成压缩,然后在云端利用确定性观测矩阵的转置对压缩信号进行转置投影操作得到近似信号,并将近似信号与原始信号进行相似性验证,最后将近似信号输入转置投影CNN(TP-CNN)进行房颤检测。本发明提出一种压缩感知与残差神经网络结合的房颤筛查方法,直接对压缩心电的转置投影信号进行房颤检测,与最先进的方法进行比较,超过面向压缩信号的方法的结果,与原始信号的分类结果相比也是具有竞争性的。(The invention provides an atrial fibrillation detection method based on an electrocardio compression sensing transposed projection signal. According to the technical scheme, firstly, a deterministic observation matrix is used for randomly projecting a electrocardiosignal to complete compression, then transposition projection operation is carried out on the compressed signal by utilizing transposition of the deterministic observation matrix at the cloud end to obtain an approximate signal, similarity verification is carried out on the approximate signal and an original signal, and finally the approximate signal is input into transposition projection CNN (TP-CNN) to carry out atrial fibrillation detection. The invention provides an atrial fibrillation screening method combining compressed sensing and a residual neural network, which directly detects atrial fibrillation of a transposed projection signal of compressed electrocardio, compares the result with the most advanced method, exceeds the result of a method facing a compressed signal, and is competitive with the classification result of an original signal.)

1. An atrial fibrillation detection method based on an electrocardio compression sensing transposed projection signal comprises a wearable device end and a remote server end, and is characterized in that: the wearable equipment end randomly projects the original electrocardiosignals according to a compression perception theory, and the original electrocardiosignals are compressed; the remote server side performs projection transposition and normalization processing on a compressed signal sent by the wearable equipment side, inputs the processed data into a pre-established atrial fibrillation detection 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∈R^Mis a compressed signal, is a compressed representation of the original signal,

x is a raw ECG signal and X is a raw ECG signal,

is an observation matrix to control the degree of compression by adjusting the size of M;

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

and step 3: the server side performs transposition projection operation on the compressed signal; the process is as follows:

wherein the content of the first and second substances,

X^∈R^Nis an approximation signal, has the same size as the original signal,

is the transpose of the observation matrix;

and 4, step 4: the server side performs normalization operation on the transposed projection signal; the process is as follows:

wherein the content of the first and second substances,

is the average value of the average of the values,

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

z is data after normalization;

and 4, step 4: inputting the preprocessed data into a pre-established atrial fibrillation detection model by the server side, and outputting a detection result of an atrial fibrillation signal; the process is as follows:

R＝g(Z,θ)；

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),

o is the probability that a compressed signal of the model output belongs to a certain class.

2. The atrial fibrillation detection method based on the electrocardio-compressed sensing transposed projection signal as claimed in claim 1, wherein the atrial fibrillation detection method comprises the following steps: the whole model mainly comprises four modules: the specific parameters of the rolling block, the residual block 1, the residual block 2 and the detection block are shown in the table 1. The convolutional block includes Conv, BN, and ReLU. The residual block 1 includes Conv, BN, ReLU, Dropout, and Conv. The residual block 2 includes BN, ReLU, Dropout, Conv, BN, ReLU, Dropout, and Conv. The detection module includes GAP and Softmax. GAP reduces feature dimensionality and provides a more specific internal representation. Softmax is used to output the probability of each compressed ECG slice being atrial fibrillation or non-atrial fibrillation.

3. The atrial fibrillation detection method based on the electrocardio-compressed sensing transposed projection signal as claimed in claim 1, wherein the atrial fibrillation detection method comprises the following steps: during the training of the atrial fibrillation detection algorithm, the batch size is set to 256 and the epoch is set to 50. The loss function used for training is the classification cross entropy, Adam is adopted for optimization, the initial learning rate is set to be 0.0009, and the model parameters are updated in each cycle iteration.

4. The atrial fibrillation detection method based on the electrocardio-compressed sensing transposed projection signal as claimed in claim 1, wherein the atrial fibrillation detection method comprises the following steps: 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 atrial fibrillation detection method based on the electrocardio-compressed sensing transposed projection signal as claimed in claim 1, wherein the atrial fibrillation detection method comprises the following steps: 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 an atrial fibrillation detection method based on an electrocardio compression sensing transposed projection signal.

Background

Atrial fibrillation is the most common persistent arrhythmia and is not life threatening in itself, but it greatly increases the incidence of the various complications involved (stroke, thrombosis, etc.) and can even lead to death. As reported by the world heart association, the incidence of atrial fibrillation and related mortality has increased exponentially over the past decade, and has become a problem of global health concern. In 2010, 3,350 million people worldwide had atrial fibrillation. Atrial fibrillation requires long-term pharmacological treatment and hospitalization, placing a large, ever-increasing economic burden on the healthcare system. Therefore, it is of great importance to accurately detect atrial fibrillation and take effective treatment measures.

The electrocardiogram is one of the most common noninvasive tools for recording the physiological activities of the heart, and the electrocardiogram data can provide clinicians with important information on the health and pathology of the heart, and is a typical atrial fibrillation diagnostic tool. Some 12-lead ecg monitoring devices are designed to ensure accurate atrial fibrillation detection. Unfortunately, for patients with atrial fibrillation, wearing multi-lead dynamic electrocardiograms for long-term monitoring is cumbersome and has a great influence on daily activities of the patients. Although atrial fibrillation is diagnosed by a 12-lead electrocardiogram, screening for atrial fibrillation by using a single-lead electrocardiogram signal is a cost-effective method. Wearable devices have found numerous applications in beyond-diagnostic health care because they can continuously monitor a patient's heartbeat activity on a daily basis. Battery capacity and storage capacity may be limited based on considerations of the size and portability of the wearable device. The wearable equipment is used for signal acquisition and processing in the process of carrying out daily electrocardiogram monitoring tasks, and most of the wearable equipment is used for transmission. For example, when the sampling rate is 400Hz and the resolution is 12 bits, 26MB of data needs to be stored or transmitted in a single lead electrocardiosignal; when the resolution is 16 bits, two leads are required to transmit 138MB of data. Therefore, how to reduce the signal acquisition of the wearable device as much as possible is a crucial issue to improve the data transmission efficiency.

The electrocardiographic compression algorithm has been studied by researchers for a long Time, such as an algorithm based on Time, like Amplitude-Zone-Time-Epoch-coding (aztec) and coding-Reduction-Time-Encoding System (CORTES), and an algorithm based on a transform domain, like discrete wavelet transform, discrete cosine transform. For low power devices, compression algorithms can be power hungry due to complex transformations. Compressed sensing is a lossy compression method, because operations that guarantee sub-nyquist sampling and low complexity have great attraction to remote monitoring systems, computational resources that can reduce the power consumption of measurement devices when electrocardiosignals are compressed. Currently, related work of electrocardio compression sensing is mainly divided into two frames: (a) reconstructing an analysis framework: compressing the signal at the acquisition end, transmitting the compressed signal to the cloud end, reconstructing the compressed signal at the cloud end, and finally analyzing the reconstructed signal in the next step, which is shown in fig. 2 (a); (b) a compression learning framework: and (3) signal compression is carried out at the acquisition end, the compressed signals are transmitted to the cloud end, and the compressed signals are directly classified at the cloud end, which is shown in a figure 2 (b). The problem with framework (a) is that reconstruction of the compressed signal at the cloud requires significant computational complexity. The framework (b) is directed to the problem of the framework (a), and the compressed signal is directly classified by skipping the reconstruction process by using the compression learning theory, but because the compressed signal loses much information, the result of the framework (b) on classifying the electrocardio-compressed signal is not good enough.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides an atrial fibrillation detection method based on an electrocardio compression sensing transposed projection signal. 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∈R^Mis a compressed signal, is a compressed representation of the original signal,

x is a raw ECG signal and X is a raw ECG signal,

is an observation matrix to control the degree of compression by adjusting the size of M;

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

and step 3: the server side performs transposition projection operation on the compressed signal; the process is as follows:

wherein the content of the first and second substances,

X^∈R^Nis an approximation signal, has the same size as the original signal,

is the transpose of the observation matrix;

and 4, step 4: the server side performs normalization operation on the transposed projection signal; the process is as follows:

wherein the content of the first and second substances,

is the average value of the average of the values,

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

z is data after normalization;

and 4, step 4: inputting the preprocessed data into a pre-established atrial fibrillation detection model by the server side, and outputting a detection result of an atrial fibrillation signal; the process is as follows:

R＝g(Z，θ)；

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),

o is the probability that a compressed signal output by the model belongs to a certain class;

compared with the prior art, the atrial fibrillation detection method has outstanding substantive characteristics and remarkable progress, and particularly provides an atrial fibrillation detection model (TP-CNN) combining compressed sensing and a residual error network. Compared with the most advanced method, the detection result of the atrial fibrillation is far higher than that of the compressed signal, and in addition, compared with the detection result of the atrial fibrillation of the original signal, the detection method also has competitive performance. Under the wearable application scene, the high-compression-ratio and high-precision atrial fibrillation detection can be realized, and the method is a promising method.

Drawings

Fig. 1 is an overall architecture of the atrial fibrillation detection method of the present invention.

Fig. 2 shows different frames of electrocardiographic compression sensing.

Fig. 3 shows the results of atrial fibrillation detection at different CR for different observation matrices.

Fig. 4 is a pearson correlation coefficient between an original signal and an approximated signal.

Fig. 5 shows cosine similarities of four features between an original signal and an approximated signal.

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;

performing transposition projection on the compressed signal and normalizing transposition projection data;

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

Giving an electrocardiosignal X epsilon R^NThe original signal compression is completed by the following steps:

Y∈R^Mis a compressed signal, is a compressed representation of the original signal,

x is a raw ECG signal and X is a raw ECG signal,

is an observation matrix to control the degree of compression by adjusting the size of M.

The transpose projection operation of the compressed data is completed by the following steps:

X^∈R^Nis an approximation signal, has the same size as the original signal,

is the transpose of the observation matrix.

The normalization operation of the inverted projection signal is completed by the following steps:

is the average value of the average of the values,

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

z is the data after normalization.

Inputting the processed data into the TP-CNN network model for atrial fibrillation detection is completed by the following steps:

R＝g(Z，θ)；

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),

o is the probability that a compressed signal of the model output belongs to a certain class.

The TP-CNN network structure parameters provided by the invention are shown in Table 1, and the whole model mainly comprises four modules: the specific parameters of the rolling block, the residual block 1, the residual block 2 and the detection block are shown in the table 1. The convolutional block includes Conv, BN, and ReLU. The residual block 1 includes Conv, BN, ReLU, Dropout, and Conv. The residual block 2 includes BN, ReLU, Dropout, Conv, BN, ReLU, Dropout, and Conv. The detection module includes GAP and Softmax. During training, the batch size is set to 256 and epoch to 50. The loss function used for training is the classification cross entropy, Adam is adopted for optimization, the initial learning rate is set to be 0.0009, and the model parameters are updated in each cycle iteration.

TABLE 1 detailed parameters of the model

Verification experiment

The data set used for the experiment was the MIT-BIH atrial fibrillation data set. The MIT-BIH atrial fibrillation dataset consists of long-term cardiac electrical recordings of 25 patients with atrial fibrillation (mainly paroxysmal fibrillation). Of these, record numbers 00735 and 03665 are only rhythm files and beat note files that have not been reviewed, so we use the remaining 23 records. Each recording lasted about 10 hours, with a sampling rate of 250Hz, including two lead signals (we chose the first lead). 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). The present invention separates datasets into two categories: the recording of atrial fibrillation and non-atrial fibrillation recording (including N, AFL and J) and atrial fibrillation recording, all electrocardiosignals are divided into 8s electrocardio segments, and approximate signals are subjected to classification operation to detect atrial fibrillation.

The method aims to detect the atrial fibrillation rhythm in the electrocardio compressed sensing signal of the wearable device. The observation matrix is an important part of compressed sensing, and its characteristics directly determine the physical feasibility of compressed sensing theory and the accuracy of compressed sensing signal reconstruction (although reconstruction is not involved here). The domain of electrocardiographic compressed sensing often uses gaussian observation matrices (GRSM) in the early days, but is limited in practical application because of floating point data. To solve this problem, random sparse matrices like Bernoulli distributed observation matrices (BDRM) and sparse binary observation matrices (SBSM) have been proposed, which can greatly reduce the computational complexity of the compression process and can also perform similarly to dense Gaussian observation matrices because they have most zeros themselves.

In order to analyze the influence of the observation matrixes on the atrial fibrillation detection result in detail, GRSM, BDRM, SBSM and SDMM are used for compressing original electrocardiosignals, the compressed signals are transposed and projected to obtain respective approximate signals, the approximate signals obtained correspondingly by each observation matrix are input into a neural network for atrial fibrillation detection, in order to simplify the process, experiments are only carried out on four compression ratios, and the specific result is shown in figure 3.

As can be seen from FIG. 3, CR ranges from 2 to 10, and the MCC (Acc), Recall (Rec), precision (Pre), specificity (Specification), F1 score (F1) and MCC correspond to GRSM, BDRM and SBSM, respectively, which are reduced in steps, especially the MCC performs remarkably (MCC is more suitable than Acc and F1 in the evaluation of the second classification task), and the MCC is reduced from 94.51%, 94.17% and 95.20% to 79.72%, 79.08% and 76.74% in the evaluation of the second classification task. However, the SDMM is 98.79%, 98.67%, 98.60% and 98.49% at the four compression ratios of MCC, respectively, indicating that the use of SDMM is highly suitable.

The input of the neural network is that an observation matrix compresses an original signal, and an approximate signal is obtained by transposition projection and normalization, wherein the approximate signal is not a reconstructed signal obtained by a reconstruction algorithm with high computation complexity and is different from the original signal. Why can the approximate signal obtained by SDMM give a better atrial fibrillation detection performance? In the following, the invention will analyze the relation between the approximated signal and the original signal.

We perform the analysis from both the signal bulk and the signal local features. First, we use the pearson correlation coefficient to measure the degree of similarity between the original signal and the approximated signal. Fig. 4 shows the pearson correlation coefficient between the original signal and the approximated signal, which we calculated for 23 records in the AFDB at nine compression ratios. As can be seen from fig. 4, even when CR is 10, the median of the pearson correlation coefficients of 23 records is 0.8391. In other words, even if the original signal is compressed to 1/10, its approximate signal is similar to the original signal. In addition, a singular point-08405 record appears in the figure, and the reason why the record becomes the singular point is probably that the record is extremely noisy per se and even has part of unreadable.

Then, we perform analysis from the local features of the cardiac signal. Four features of meanRR, RMSSD, SDNN and R wave density are respectively extracted from an original signal and an approximate signal. Because cosine similarity is widely applied to evaluating feature similarity, for the four features, the cosine similarity is calculated under different compression ratios for the 23 recorded original signals and the approximate signals respectively, and it is shown by the calculation result of fig. 5 that CR is from 2 to 10, cosine similarity between the original signals and the approximate signals of all recorded meanRR and rdence is higher than 0.9855, median of cosine similarity between the original signals and the approximate signals of all recorded RMSSD and SDNN is higher than 0.8798, which indicates that local features between the approximate signals and the original signals are also relatively similar. In summary, the data from the signal as a whole to the local features can lead to the conclusion that: the approximate signal has the characteristics of extremely approximate original signals, so that a neural network can be provided for effective learning.

After verifying the validity of the approximation signal processed by SDMM, the part analyzes the detection result of atrial fibrillation under different compression ratios (CR is 0,2,3,4,5,6,7,8,9,10) of TP-CNN. CR-0 indicates raw electrocardiographic data that is not compressed.

Table 2 lists the results of TP-CNN and our previous working CS-CNN at different compression ratios. The experimental result shows that, from the aspect of the method, no matter TP-CNN or CS-CNN, the classification result of the original signal is very close to the result of the approximate data under different compression ratios, and no obvious reduction occurs, wherein the performances of six evaluation indexes of CS-CNN under different compression ratios are different by about 2% at most, and the performances of six evaluation indexes of TP-CNN under different compression ratios are different by about 1% at most, which can explain the effectiveness of processing the compressed signal into the approximate signal. From the expression level, six indexes of TP-CNN are higher than that of CS-CNN, for example, Acc (%), Rec (%), Pre (%), Spec (%), F1 (%) and MCC (%) of TP-CNN are respectively 4.39%, 6.09%, 4.81%, 3.24%, 5.45% and 9.12% higher than that of CS-CNN at CR of 10, which shows the effectiveness of TP-CNN proposed by us.

TABLE 2 AF detection results at different compression ratios

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.