Arrhythmia classification method and device based on wrist watch PPG signal

Abstract

This invention provides a method and device for classifying cardiac arrhythmias based on PPG signals from a wristwatch. The method includes: Step 1, collecting PPG signals from the human body using a wristwatch; Step 2, obtaining an MGE Block module with improved ResNet residual blocks by fusing MGC and an attention-based ECA module; Step 3, stacking multiple MGE Block modules to extract multi-scale spatial features of the PPG signals, including feature vectors; Step 4, transmitting the feature vectors to a BiGRU module to extract temporal features of the PPG signals; Step 5, fusing the multi-scale spatial features and temporal features to identify and classify various cardiac arrhythmias. This invention can identify various cardiac arrhythmias based on PPG signals collected from a wristwatch with high accuracy, enabling the monitoring of cardiac health using a wristwatch.

Description

Technical Field

[0001] The present invention relates to the technical field of human arrhythmia detection, and in particular to a method and device for classifying arrhythmia based on wristwatch PPG signals. Background Art

[0002] Cardiovascular disease is a series of diseases that affect the function of the heart and vascular system. Patients with mild cardiovascular disease may experience symptoms such as chest tightness and weakness, which affect normal work and daily life; while patients with severe cardiovascular disease may suffer from insufficient blood supply to the heart, leading to fainting and even life-threatening. Generally speaking, the key physiological parameters monitored for heart health include heart rate, blood pressure, blood oxygen saturation, the type of arrhythmia and the frequency of its occurrence. Among them, heart rhythm is the normal rhythm of the heart's electrical activity and is also an important indicator for assessing the health of the heart. Arrhythmia refers to an abnormality in the rhythm of the heart or the regularity of the heartbeat. In patients with cardiovascular disease, the incidence of arrhythmias has increased significantly and the types are diverse. For example, atrial fibrillation, ventricular tachycardia, premature beats, atrioventricular block, etc., will affect the effective contraction and discharge of the heart, leading to the serious consequence of thrombosis.

[0003] Most studies at home and abroad diagnose arrhythmias based on electrocardiogram (ECG) signals. Ahmed proposed a novel one-dimensional convolutional neural network (CNN) architecture and conducted experiments on the MIT-BIH dataset. The overall accuracy of the classification of four types of arrhythmias reached 99%. Chen fused CNN and LSTM networks to achieve arrhythmia classification on a public dataset, achieving an accuracy of 99.32%. Hannun proposed a deep convolutional neural network with 16 residual blocks, which classified 12 types of heart rhythms and achieved an average F1 score of 83.7%, which exceeded the average score of cardiologists.

[0004] With the popularity of wearable devices, especially the increasing application of smart watches, the diagnosis of arrhythmias based on photoplethysmography (PPG) signals has gradually become one of the research hotspots, mainly in two major directions: machine learning and deep learning.

[0005] In terms of traditional machine learning, Sung et al. used a logistic regression model to identify atrial fibrillation by extracting six features such as the waveform of the PPG signal as input. Fallet et al. used a decision tree classification model. They calculated a series of parameters such as mean and standard deviation from the PPG data, which were used to identify and analyze atrial fibrillation. However, these methods require manual feature extraction, and the extracted features are limited and may not capture the complex relationships in the data, resulting in limited recognition results.

[0006] In recent years, deep learning methods that can perform automatic feature extraction have begun to be applied to arrhythmia detection. For example, Shen et al. established a data set containing more than 4,000 records of PPG signals and used a 49-layer CNN network model to identify atrial fibrillation. The results showed that the area under the operating characteristic curve was 0.9480. Zaen and his team combined acceleration data to develop a classification model based on a recurrent neural network (RNN). The model has a detection accuracy of 0.986 on the type of atrial fibrillation, and the sensitivity and specificity are 1.0 and 0.978 respectively. Kwon et al. used a joint model of a one-dimensional CNN network and an RNN network to identify atrial fibrillation with an accuracy of 97.58%. However, these studies have focused on the identification of a single type of arrhythmia, with an emphasis on the identification of atrial fibrillation.

[0007] In order to study the feasibility of multi-type arrhythmia classification, Liu et al. constructed a PPG signal dataset of six heart rhythms, including sinus rhythm (SR), premature ventricular contraction (PVC), premature atrial contraction (PAC), ventricular tachycardia (VT), supraventricular tachycardia (SVT) and atrial fibrillation (AF), and proposed a deep convolutional neural network DCNN for identification, with an overall accuracy of 85.0%. However, from the literature reviewed, there are currently few research papers on multi-type arrhythmias based on PPG signals. Summary of the invention

[0008] The purpose of the present invention is to provide a method and device for classifying arrhythmias based on PPG signals of a wristwatch, so as to solve the problems that there are few studies on arrhythmias identified based on PPG signals in the prior art and the recognition accuracy is not high. The method and device can identify a variety of arrhythmias based on the PPG signals collected on the wristwatch with high recognition accuracy, and the purpose of using the wristwatch to monitor heart health status can be achieved.

[0009] In order to achieve the above objectives, in a first aspect, the present invention provides a method for classifying arrhythmias based on wrist watch PPG signals, comprising:

[0010] Step 1: Use the wristwatch to collect the human body's PPG signal;

[0011] Step 2: The MGEBlock module of the improved ResNet residual block is obtained by fusing the MGC and the ECA module based on the attention mechanism;

[0012] Step 3: stack multiple layers of MGE Block modules to extract multi-scale spatial features of the PPG signal, where the multi-scale spatial features include feature vectors;

[0013] Step 4: Transfer the feature vector to the BiGRU module to extract the temporal features of the PPG signal;

[0014] Step 5: Fuse multi-scale spatial features and temporal features to identify and classify various arrhythmias.

[0015] According to a method for classifying arrhythmias based on wristwatch PPG signals provided by the present invention, the MGE Block module includes an MGE unit; the MGE unit is used to use MGC and an ECA module based on an attention mechanism to capture multi-scale spatial features of the PPG signal through convolution kernels of different sizes.

[0016] According to a method for classifying arrhythmias based on wristwatch PPG signals provided by the present invention, the MGE unit is specifically used for:

[0017] The multi-scale feature map of the PPG signal in the channel direction is obtained through MGC;

[0018] Use the ECA module based on the attention mechanism to filter out key attention data from the multi-scale feature map and obtain the channel attention vector of the multi-scale feature map;

[0019] Use the Softmax function to normalize the channel attention vector;

[0020] Multiply the normalized channel attention vector and the multi-scale feature map to obtain the multi-scale spatial features.

[0021] According to a method for classifying arrhythmias based on wristwatch PPG signals provided by the present invention, an ECA module based on an attention mechanism is used to screen out key attention data from a multi-scale feature map, and obtain a channel attention vector of the multi-scale feature map, including:

[0022] Perform global average pooling on the feature map of each scale in the multi-scale feature map, and compress the original size [H, W, C] of the feature map of this scale into [1, 1, C], where H is the height of the feature map of this scale, W is the width of the feature map of this scale, and C is the number of channels;

[0023] The convolution kernel is calculated according to the set number of channels and the preset formula. The preset formula is:

[0024]

[0025] In the formula, k is the convolution kernel;

[0026] The feature map of this scale is convolved with the convolution kernel, and the channel attention vector in the direction of this scale is obtained through the sigmoid function.

[0027] According to a method for classifying arrhythmias based on wristwatch PPG signals provided by the present invention, the expression of the Softmax function is:

[0028]

[0029] In the formula, is the i-th component of the channel attention vector, and n is the dimension of the channel attention vector.

[0030] According to the present invention, a method for classifying arrhythmias based on wristwatch PPG signals is provided for identifying and classifying normal heart rhythm, atrial fibrillation, ventricular premature beats, supraventricular tachycardia, ventricular tachycardia, and atrial premature beats.

[0031] In a second aspect, the present invention further provides an arrhythmia classification device based on a wristwatch PPG signal, comprising:

[0032] A unit is used to collect PPG signals of a human body using a wristwatch;

[0033] A construction unit for obtaining an MGE Block module of the residual block of the improved ResNet by fusing the MGC and the ECA module based on the attention mechanism;

[0034] A spatial extraction unit is used to stack multiple layers of MGE Block modules to extract multi-scale spatial features of PPG signals, where the multi-scale spatial features include feature vectors;

[0035] A time extraction unit, used to transmit the feature vector to the BiGRU module to extract the time features of the PPG signal;

[0036] The recognition and classification unit is used to fuse multi-scale spatial features and temporal features to identify and classify various arrhythmias.

[0037] According to an arrhythmia classification device based on a wristwatch PPG signal provided by the present invention, the MGE Block module includes an MGE unit; the MGE unit is used to use MGC and an ECA module based on an attention mechanism to capture multi-scale spatial features of the PPG signal through convolution kernels of different sizes.

[0038] According to the arrhythmia classification device based on wristwatch PPG signal provided by the present invention, the MGE unit is specifically used for:

[0039] The multi-scale feature map of the PPG signal in the channel direction is obtained through MGC;

[0040] Use the ECA module based on the attention mechanism to filter out key attention data from the multi-scale feature map and obtain the channel attention vector of the multi-scale feature map;

[0041] Use the Softmax function to normalize the channel attention vector;

[0042] Multiply the normalized channel attention vector and the multi-scale feature map to obtain the multi-scale spatial features.

[0043] The present invention provides a method and device for classifying arrhythmias based on wristwatch PPG signals. The method is an arrhythmia classification method based on the MGERB model, and can identify the characteristics of six different arrhythmias. The method uses the PPG signal as input, uses the MGE Block module stacking of multiple improved ResNet residual blocks to extract multi-scale spatial features, uses the BiGRU module to extract time features, and fuses the multi-scale spatial features and time features to identify and classify multiple arrhythmias. The present invention can identify multiple arrhythmias based on the PPG signal collected on the wristwatch, with a high recognition accuracy, and can achieve the purpose of using the wristwatch to monitor heart health status. Experiments show that the method proposed in the present invention achieves a classification effect with an overall accuracy of 91.11% on the public data set, and achieves a classification effect with an overall accuracy of 95.57% on the self-collected data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] In the attached picture:

[0046] Figure 1 It is a structural diagram of the MGERB model of the present invention;

[0047] Figure 2 It is a structural schematic diagram of the MGE unit of the present invention;

[0048] Figure 3 It is a schematic diagram of the principle of group convolution of the present invention;

[0049] Figure 4 It is a structural schematic diagram of the ECA module of the present invention;

[0050] Figure 5 It is a structural schematic diagram of the BiGRU module of the present invention;

[0051] Figure 6 The ROC curve diagram of the six heart rhythms in the test set of the present invention;

[0052] Figure 7 The figure is a flow chart of the arrhythmia classification method based on wrist watch PPG signal of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Some embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0055] See also Figure 7 The embodiment of the present invention provides a method for classifying arrhythmias based on wrist watch PPG signals, comprising the following steps:

[0056] Step 1: Use the wristwatch to collect the human body's PPG signal;

[0057] Step 2: The MGEBlock module of the improved ResNet residual block is obtained by fusing the MGC and the ECA module based on the attention mechanism;

[0058] Step 3: stack multiple layers of MGE Block modules to extract multi-scale spatial features of the PPG signal, where the multi-scale spatial features include feature vectors;

[0059] Step 4: Transfer the feature vector to the BiGRU module to extract the temporal features of the PPG signal;

[0060] Step 5: Fuse multi-scale spatial features and temporal features to identify and classify various arrhythmias.

[0061] In order to facilitate a clearer understanding of the technical solution of the present invention, a detailed description is given below.

[0062] The method of the embodiment of the present invention is implemented based on the MGERB (Multi-scale Group convolution Efficientchannel attention Residual network Bidirectional gated recurrent unit) model. The MGERB model extracts spatial features by stacking multiple layers of MGE Block and extracts temporal features by using the BiGRU model to construct an overall neural network architecture.

[0063] The MGERB model extracts spatial and temporal features of PPG signals to achieve the purpose of identifying multiple arrhythmias. First, a new module named MGE Block is obtained by fusing multi-scale grouped convolution (MGC) and ECA module based on attention mechanism to improve the residual block of residual network (ResNet); then, the MGE Block module is stacked in multiple layers to extract the multi-scale spatial features of PPG signals. After that, the feature vector of the extracted multi-scale spatial features is transferred to the gated bidirectional recurrent unit (BiGRU) module to extract the temporal features of PPG signals. Finally, the multi-scale spatial features from the MGE Block module and the temporal features from the BiGRU module are fused to identify and classify multiple arrhythmias.

[0064] It should be noted that the MGERB model can accurately identify a variety of arrhythmias, mainly because it effectively extracts spatial and temporal features from the PPG signal by combining the MGE Block module and the BiGRU module of the improved residual block. Combining the two types of spatial and temporal features for comprehensive analysis enables the MGERB model to distinguish the subtle differences between different types of arrhythmias, thereby achieving accurate identification and classification of a variety of arrhythmias.

[0065] Figure 1 The architecture diagram of the MGERB model shows four stacked MGE Block modules. The MGERB model includes a one-dimensional convolution layer (1D Conv), a batch normalization layer (BN), a rectified liner unit (ReLu), a maximum pooling layer (Maxpooling), four MGE Block modules, a BiGRU module, a batch normalization layer, a rectified linear unit layer, a random dropout layer (DropOut), a fully connected layer (Dense Layer), and a Softmax function. The MGE Block module includes a one-dimensional convolution layer, a batch normalization layer, an MGE unit, a batch normalization layer, a one-dimensional convolution layer, a batch normalization layer, and a rectified linear unit layer connected in sequence.

[0066] It is understandable that 1D Conv is mainly used to process one-dimensional data. BN normalizes the input of each training batch to speed up convergence, alleviate the problems of gradient vanishing and explosion, and enhance the generalization ability of the model. DropOut is a commonly used regularization method. During the training process, some neuron outputs are randomly set to 0 to prevent overfitting. ReLu is a nonlinear activation function that is often used in the hidden layer of convolutional neural networks and can well extract the nonlinear features of data. Maxpooling reduces the dimension of data by selecting the maximum value of each local area while retaining important feature information. Dense Layer is usually used in the last layer of a neural network to map input features to output results to output the final prediction results.

[0067] Among them, in terms of spatial feature extraction, firstly, the MGE Block module of the improved ResNet residual block can avoid the gradient vanishing problem in deep network training through residual connection, so that the MGERB model can mine more detailed spatial features at a deep level, such as the shape and amplitude of the PPG signal waveform. Secondly, the MGE Block module replaces the 3×3 convolution in the residual block of ResNet with the MGE (Multi-scale Groupconv ECA, MGE) unit. In the MGE unit, multi-scale group convolution is used, which captures the spatial features of the PPG signal at multiple scales through convolution kernels of different sizes, enhancing the MGERB model's ability to understand and extract different waveform features of the PPG signal. In addition, the MGE unit also incorporates the ECA module based on the attention mechanism, so that the MGERB model focuses on the features that are more critical to arrhythmia identification, while discarding unimportant features. Therefore, when processing PPG signals, the MGERB model can focus on the key information that is more useful for diagnosing arrhythmias, thereby improving the accuracy and efficiency of classification.

[0068] The structure of the MGE unit is as follows Figure 2 As shown in the figure, the functions of the MGE unit are mainly implemented in four steps. First, the multi-scale feature map of the PPG signal in the channel direction is obtained through MGC. Then, the ECA module based on the attention mechanism is used. Its core function is to filter out the key attention data from the multi-scale feature map obtained by MGC, and then obtain the channel attention vector of the multi-scale feature map. Then, in order to promote the attention interaction on different channels of the multi-scale feature map, the channel attention vector is normalized using the Softmax function. Finally, the normalized channel attention vector is multiplied with the multi-scale feature map to obtain the multi-scale spatial feature.

[0069] The following is a specific embodiment of the functions of the MGE unit.

[0070] First, the input PPG signal feature map is divided into four parts, so that the number of channels in each part is the same. , feature extraction of different scales is performed on these four parts, and the convolution kernel sizes are 3×3, 5×5, 7×7, and 9×9 respectively. Multi-scale convolution is used to capture the multi-level spatial information of the PPG signal feature map, significantly enhancing the feature expression ability. This method can also enable the network to better understand the semantic content of the input image, thereby further improving the recognition accuracy of the model.

[0071] However, in order to reduce the amount of calculation, group convolution is adopted. After experiments, it is determined that the number of groups for each part is 1, 4, 8, and 16 respectively.

[0072] As a variant of the convolutional network structure, grouped convolution innovatively optimizes the structure of the traditional convolutional network by cleverly dividing the input feature channels into several independent subgroups and applying convolution operations to each subgroup. Figure 3 Here is an illustration of how it works, taking a grouped convolution divided into two groups as an example:

[0073] (a) The structure of the input feature map is , divided into 2 groups for convolution on the channel. Among them, is the height of the feature map, is the width of the feature map, C 1 is the number of channels in the feature map.

[0074] (b) There are also 2 filter groups. The input data processed by each group is the input feature map with half the number of channels. Therefore, the filter size of each group is , the number of each group is .in, C 2 is the number of filters.

[0075] (c) Convolve the filter with the input feature map to obtain the output feature map. The structure of the output feature map is H 2× W 2× , since the number of output channels should be consistent with the number of convolution kernels, the number of output channels of each group of filters is .

[0076] (d) Obtained by stacking the output channel maps of all groups in the channel dimension The output of the channel, the output structure size is H 2× W 2×C2.

[0077] In standard convolution, the parameter quantity is shown in formula (1):

[0078] (1)

[0079] In grouped convolution, the number of parameters is reduced as shown in formula (2):

[0080] (2)

[0081] Through this parallel processing method, the number of parameters in the grouped convolution becomes 1 / 2 of the standard convolution. , effectively improving the computational efficiency. is the number of groups in the grouped convolution. In the above example, 2.

[0082] The function of the ECA module based on the attention mechanism is to extract the channel attention weight of each scale for the feature map of each scale, which is an optimization of the SE attention mechanism. The two major operations of SE are compression and excitation. However, the dimensionality reduction of the input parameters in the compression step will cause some of the original feature information to be lost, thus affecting the output result. In addition, SE uses two fully connected layers, which increases the computational complexity of the model. In comparison, the ECA module uses a 1×1 convolutional layer after the global average pooling layer, avoiding the dimensionality reduction operation of the fully connected layer in SE, and performing local cross-channel interactions through the 1×1 convolutional layer. The specific structure is as follows: Figure 4 Through this design, the computational complexity is reduced and the performance of the model can be effectively improved.

[0083] The implementation process of the ECA module is as follows:

[0084] (a) First, global average pooling is performed on the feature maps of each scale in the multi-scale feature map, and the original size [H, W, C] of the feature map of this scale is compressed to [1, 1, C], where H is the height of the feature map of this scale, W is the width of the feature map of this scale, and C is the number of channels.

[0085] (b) Next, the convolution kernel is calculated based on the set number of channels and the preset formula. The preset formula is:

[0086]

[0087] In the formula, k is the convolution kernel;

[0088] (c) Then, the feature map of this scale is convolved using the convolution kernel, and the channel attention vector in the direction of this scale is obtained through the sigmoid function.

[0089] In order to promote the effective interaction of multi-scale channel attention, the Softmax function is used to recalibrate the channel attention. The expression of the Softmax function is shown in formula (3):

[0090] (3)

[0091] In the formula, is the i-th component of the channel attention vector, and n is the dimension of the channel attention vector.

[0092] The Softmax function normalizes the channel attention at different scales, which not only effectively reduces the sensitivity of the MGERB model, but also realizes the interaction between local and global channel attention. Finally, the calibrated attention vectors are concatenated in a cascade manner to form a complete channel attention vector, further improving the MGERB model's efficiency in capturing and recognizing details and global information.

[0093] In terms of temporal feature extraction, the BiGRU module can fully capture dynamic changes in time series, such as key temporal features such as heartbeat rhythm and frequency, by simultaneously processing the forward and reverse sequences of the PPG signal. This in-depth understanding of the temporal dimension is particularly important for the identification of arrhythmias, because different types of arrhythmias often show different characteristics in time series.

[0094] GRU optimizes the three-gate mechanism of the LSTM network by removing the cell state and integrating the forget gate and input gate into a unified update gate, making the update gate parameters fewer and more efficient. GRU can retain the properties of long time series spanning tens to hundreds of time steps in its internal state. When applying GRU to identify arrhythmias, GRU can effectively capture the long-term and short-term dependencies in the PPG signal through its gating mechanism, analyze its dynamic changes over time, and identify the irregularity and abnormal patterns of the heart rhythm, thereby achieving the purpose of identifying multiple arrhythmias.

[0095] The neuron state of the GRU neural network model is always transmitted from front to back, without considering the input information at the future moment. Therefore, in order to comprehensively use the past and future information of the PPG signal, the present invention uses a BiGRU module including a forward GRU layer and a backward GRU layer.

[0096] BiGRU is a bidirectional gated recurrent unit that combines the advantages of bidirectionality and gating mechanism. It is mainly used to process sequence data. The structure of the BiGRU module is as follows: Figure 5 As shown, Figure 5 The structural schematic diagram at time t=1, 2, and 3 is shown in FIG.

[0097] The specific calculation principle of the BiGRU module is shown in equations (4), (5) and (6).

[0098] (4)

[0099] (5)

[0100] (6)

[0101] in, Indicates the current input. is the forward GRU hidden state output, is the reverse GRU hidden state output, and They are and The weight vector of is the bias vector at the current moment.

[0102] Based on the same inventive concept, another embodiment of the present invention provides an arrhythmia classification device based on a wristwatch PPG signal, the device corresponding to the method of the aforementioned embodiment, the device comprising:

[0103] A unit is used to collect PPG signals of a human body using a wristwatch;

[0104] A construction unit for obtaining an MGE Block module of the residual block of the improved ResNet by fusing the MGC and the ECA module based on the attention mechanism;

[0105] A spatial extraction unit is used to stack multiple layers of MGE Block modules to extract multi-scale spatial features of PPG signals, where the multi-scale spatial features include feature vectors;

[0106] A time extraction unit, used to transmit the feature vector to the BiGRU module to extract the time features of the PPG signal;

[0107] The recognition and classification unit is used to fuse multi-scale spatial features and temporal features to identify and classify various arrhythmias.

[0108] The Receiver Operating Characteristic (ROC) curve represents the relationship between the true positive rate and the false positive rate of the model at different thresholds. The area under the ROC curve (AUC) is used to evaluate the classification performance of the model. The larger the AUC, the better the performance of the model.

[0109] The confusion matrix shows the relationship between the model's prediction results on the test set and the true label in a tabular form. Through the confusion matrix, a series of evaluation indicators can be further calculated, such as accuracy, recall, specificity, precision, F1 score, etc., to comprehensively evaluate the classification performance of the model.

[0110] Based on the arrhythmia classification method based on wristwatch PPG signals provided by the present invention, the PPG signal data disclosed by 91 patients were further studied. By analyzing these PPG signal data, a total of 46,827 PPG signal segments lasting 10 seconds were obtained, which contained normal heart rhythm (SR) and five arrhythmia rhythms: atrial fibrillation (AF), premature ventricular contraction (PVC), supraventricular tachycardia (SVT), ventricular tachycardia (VT), and premature atrial contraction (PAC).

[0111] Figure 6The ROC curves of the six heart rhythms on the test set are plotted. The horizontal axis is the false positive rate, and the vertical axis is the true positive rate. The closer the curve area is to 1 and the closer the curve is to the upper left corner, the better the classification effect of the model. Figure 6 It can be seen that the blue curve (SR) is closest to the upper left corner and has the largest area, followed by the yellow curve (AF), and the worst effect is the red curve (PAC).

[0112] Table 1 Experimental results of the MGERB model proposed in this paper on public datasets

[0113]

[0114] As can be seen from Table 1, if several indicators are considered comprehensively, the classification effect of the six rhythms is SR>AF>SVT>PVC>VT>PAC. The model has the best effect on SR and AF classification, with a precision of 0.9838, a recall of 0.9875, a specificity of 0.9925, and an F1 score of 0.9857 for SR; a precision of 0.9370, a recall of 0.9604, a specificity of 0.9671, and an F1 score of 0.9486 for AF. The classification effect of PAC is the worst, with a precision, recall, and F1 score of no more than 75%, which may be due to the fact that the waveforms of PAC and PVC are too similar, and the amount of data for PAC is too small. The overall accuracy of the six classifications is 91.11%.

[0115] The inventor of the present invention used the EVAL-HCRWATCH4Z wristwatch as a tool for collecting data samples.

[0116] The collection of data sets in the study faces the problem of sample quantity. Relatively speaking, there are more patients with atrial fibrillation and data is easy to collect. However, in other types of arrhythmias, such as ventricular tachycardia (VT), supraventricular tachycardia (SVT), premature atrial contractions (PAC) and premature ventricular contractions (PVC), the number of patients is relatively small, which makes it difficult to independently collect a large amount of data for each type of arrhythmia. Considering that PAC and PVC, although they have different origins, both belong to the broader category of arrhythmias, namely premature contractions. Similarly, VT and SVT are both manifestations of tachycardia, it is considered to merge PVC and PAC into one category, and VT and SVT into one category for data collection of PPG signals.

[0117] Table 2 Experimental results of the MGERB model proposed in this paper on the self-collected data set

[0118]

[0119] As can be seen from Table 2, the overall accuracy of the model proposed in the present invention reaches 95.57%.

[0120] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the embodiments disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for classifying arrhythmias based on wrist watch PPG signals, characterized in that: include: Step 1: Use the wristwatch to collect the human body's PPG signal; Step 2: The MGEBlock module of the improved ResNet residual block is obtained by fusing the MGC and the ECA module based on the attention mechanism; Step 3, stacking multiple layers of MGE Block modules to extract multi-scale spatial features of the PPG signal, wherein the multi-scale spatial features include feature vectors; Step 4: Transmit the feature vector to the BiGRU module to extract the temporal features of the PPG signal; Step 5: Fusing the multi-scale spatial features and temporal features to identify and classify various arrhythmias; The MGE Block module includes an MGE unit; the MGE unit is used to use MGC and an ECA module based on an attention mechanism to capture multi-scale spatial features of the PPG signal through convolution kernels of different sizes; The MGE unit is specifically used for: Obtaining a multi-scale feature map of the PPG signal in the channel direction through MGC; Use the ECA module based on the attention mechanism to filter out key attention data from the multi-scale feature map and obtain the channel attention vector of the multi-scale feature map; Use the Softmax function to normalize the channel attention vector; Multiply the normalized channel attention vector and the multi-scale feature map to obtain the multi-scale spatial features.

2. The arrhythmia classification method based on wrist watch PPG signal according to claim 1, characterized in that: The ECA module based on the attention mechanism is used to filter out key attention data from the multi-scale feature map to obtain the channel attention vector of the multi-scale feature map, including: Perform global average pooling on the feature map of each scale in the multi-scale feature map, and compress the original size [H, W, C] of the feature map of this scale into [1, 1, C], where H is the height of the feature map of this scale, W is the width of the feature map of this scale, and C is the number of channels; The convolution kernel is calculated according to the set number of channels and the preset formula, and the preset formula is: In the formula, k is the convolution kernel; The feature map of this scale is convolved with the convolution kernel, and the channel attention vector in the direction of this scale is obtained through the sigmoid function.

3. The arrhythmia classification method based on wrist watch PPG signal according to claim 1, characterized in that: The expression of the Softmax function is: In the formula, is the i-th component of the channel attention vector, and n is the dimension of the channel attention vector.

4. The arrhythmia classification method based on wrist watch PPG signal according to claim 1, characterized in that: Used to identify and classify normal heart rhythm, atrial fibrillation, ventricular premature beats, supraventricular tachycardia, ventricular tachycardia, and atrial premature beats.

5. A device for classifying arrhythmias based on wristwatch PPG signals, characterized in that: include: A unit is used to collect PPG signals of a human body using a wristwatch; A construction unit for obtaining an MGE Block module of the residual block of the improved ResNet by fusing the MGC and the ECA module based on the attention mechanism; A spatial extraction unit, used for stacking multiple layers of MGE Block modules to extract multi-scale spatial features of the PPG signal, wherein the multi-scale spatial features include feature vectors; A time extraction unit, used to transmit the feature vector to a BiGRU module to extract the time feature of the PPG signal; An identification and classification unit, used for fusing the multi-scale spatial features and temporal features to identify and classify various arrhythmias; The MGE Block module includes an MGE unit; the MGE unit is used to use MGC and an ECA module based on an attention mechanism to capture multi-scale spatial features of the PPG signal through convolution kernels of different sizes; The MGE unit is specifically used for: Obtaining a multi-scale feature map of the PPG signal in the channel direction through MGC; Use the ECA module based on the attention mechanism to filter out key attention data from the multi-scale feature map and obtain the channel attention vector of the multi-scale feature map; Use the Softmax function to normalize the channel attention vector; Multiply the normalized channel attention vector and the multi-scale feature map to obtain the multi-scale spatial features.

Images