Granted Patent B2

US 12,251,232 · App. 18/379,182 · Granted Mar 18, 2025

Inventors: Yinglong Wang (Jinan, CN); Guoxuan Xu (Jinan, CN); Minglei Shu (Jinan, CN); Zhaoyang Liu (Jinan, CN); Pengyao Xu (Jinan, CN)

Assignees: QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES); SHANDONG COMPUTER SCIENCE CENTER (NATIONAL SUPERCOMPUTING CENTER IN JINAN)

A61B5/367A61B5/7264

Abstract

Claims (48)

1. A multi-label electrocardiogram (ECG) signal classification method based on an improved attention mechanism, comprising the following steps:

a) preprocessing a multi-label ECG signal to acquire a preprocessed multi-label ECG signal X;

b) establishing a multi-scale feature extraction module, and inputting the preprocessed multi-label ECG signal X into the multi-scale feature extraction module to acquire an attention feature map X s ;

c) establishing a deep attention feature fusion (DAFF) network, and inputting the attention feature map X s into the DAFF network to acquire a fused feature X′ s ; and

d) establishing a classification module, and inputting the feature X′ s into the classification module to acquire an ECG signal classification result;

wherein step b) comprises the following sub-steps:

b-1) forming the multi-scale feature extraction module with a residual module A and a residual module B, wherein the residual module A comprises a batch normalization (BN) layer, a rectified linear unit (ReLU) activation function layer, a convolutional layer, a maximum pooling layer, and an attention fusion module; and the residual module B comprises a BN layer, a ReLU activation function layer, a convolutional layer, a maximum pooling layer, and an attention fusion module;

b-2) inputting the preprocessed multi-label ECG signal X into the BN layer, the ReLU activation function layer, and the convolutional layer of the residual module A sequentially to acquire a feature map X sc 1 ;

b-3) inputting the preprocessed multi-label ECG signal X into the maximum pooling layer of the residual module A to acquire a feature map X sm 1 ;

b-4) forming the attention fusion module of the residual module A with a local attention block and a global attention block, wherein the local attention block is formed sequentially with a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; and inputting the feature map X sc 1 into the local attention block to acquire a local attention feature map X sc_l 1 ;

b-5) forming the global attention block of the attention fusion module in the residual module A sequentially with an average pooling layer, a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; and inputting the feature map X sm 1 into the global attention block to acquire a global attention feature map X sc_g 1 ;

b-6) adding the local attention feature map X sc_l 1 to the global attention feature map X sc_g 1 to acquire an attention feature map X sa 1 ;

b-7) inputting the preprocessed multi-label ECG signal X into the BN layer, the ReLU activation function layer, and the convolutional layer of the residual module B sequentially to acquire a feature map X sc 2 ;

b-8) inputting the preprocessed multi-label ECG signal X into the maximum pooling layer of the residual module B to acquire a feature map X sm 2 ;

b-9) forming the attention fusion module of the residual module B with a local attention block and a global attention block, wherein the local attention block is formed sequentially with a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; and inputting the feature map X sc 2 into the local attention block to acquire a local attention feature map X sc_l 2 ;

b-10) forming the global attention block of the attention fusion module in the residual module B sequentially with an average pooling layer, a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; and inputting the feature map X sm 2 into the global attention block to acquire a global attention feature map X sc_g 2 ;

b-11) adding the local attention feature map X sc_l 1 to the global attention feature map X sc_g 2 to acquire an attention feature map X sa 2 ; and

b-12) adding the attention feature map X sa 1 to the attention feature map X sa 2 to acquire a feature map X s .

2. The multi-label ECG signal classification method based on the improved attention mechanism according to claim 1 , wherein step a) comprises the following sub-steps:

a-1) uniformly adjusting, by downsampling, a data frequency of the multi-label ECG signal to 2,048 sample points to acquire a downsampled multi-label ECG signal; decomposing, by resonance-based sparse decomposition, the downsampled multi-label ECG signal into three parts; and retaining a low-resonance component x L with an instantaneous oscillation part;

a-2) cutting all segments in the low-resonance component x L to a uniform length; and

a-3) converting, by an unsqueue( ) function or a reshape( ) function, the segments with the uniform length from one-dimensional to two-dimensional to acquire the preprocessed multi-label ECG signal X.

3. The multi-label ECG signal classification method based on the improved attention mechanism according to claim 1 , wherein

in step b-1), the convolutional layer of the residual module A is provided with a 1×25 convolution kernel, with a scale of 32;

in step b-1), the maximum pooling layer of the residual module A is provided with a 1×2 convolution kernel, with a scale of 32;

in step b-4), the first convolutional layer is provided with a 1×1 convolution kernel, with a scale of 16, and the second convolutional layer is provided with a 1×1 convolution kernel, with a scale of 32;

in step b-5), the first convolutional layer is provided with a 1×1 convolution kernel, with a scale of 16, and the second convolutional layer is provided with a 1×1 convolution kernel, with a scale of 32;

in step b-1), the convolutional layer of the residual module B is provided with a 1×15 convolution kernel, with a scale of 32;

in step b-1), the maximum pooling layer of the residual module B is provided with a 1×2 convolution kernel, with a scale of 32;

in step b-9), the first convolutional layer is provided with a 1×1 convolution kernel, with a scale of 16, and the second convolutional layer is provided with a 1×1 convolution kernel, with a scale of 32; and

in step b-10), the first convolutional layer is provided with a 1×1 convolution kernel, with a scale of 16, and the second convolutional layer is provided with a 1×1 convolution kernel, with a scale of 32.

4. The multi-label ECG signal classification method based on the improved attention mechanism according to claim 1 , wherein step c) comprises the following sub-steps:

c-1) forming the DAFF network with a first DAFF module and a second DAFF module in parallel, wherein the first DAFF module and the second DAFF module each comprise a convolutional layer, a fully connected layer, h parallel self-attention mechanisms, and an attention fusion module;

c-2) inputting the attention feature map X s into the convolutional layer of the first DAFF module to acquire a feature X sc 1 ; inputting the feature X sc 1 into the fully connected layer of the first DAFF module to acquire a new vector X s v , a new vector X s k , and a new vector X s q ; inputting the new vector X s v , the new vector X s k , and the new vector X s q into the h parallel self-attention mechanisms of the first DAFF module to acquire h features X s_A1 1 , X s_A1 2 , . . . , X s_A1 i , . . . . , X s_A1 h , wherein X s_A1 i denotes a feature output by an i-th self-attention mechanism, i∈{1, . . . ,h};

c-3) forming the attention fusion module of the first DAFF module sequentially with a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; inputting the h features X s_A1 1 , X s_A1 2 , . . . , X s_A1 i , . . . . , X s_A1 h into the first convolutional layer to acquire features X s_A1_c 1 , X s_A1_c 2 , . . . , X s_A1_c i , . . . . , X s_A1_c h ; and inputting the features X s_A1_c 1 , X s_A1_c 2 , . . . , X s_A1_c i , . . . . , X s_A1_c h into the BN layer, the ReLU activation function layer, and the second convolutional layer to acquire new features X s_A1 1′ , X s_A1 2′ , . . . , X s_A1 i′ , . . . . , X s_A1 h′ ;

c-4) superposing the new features X s_A1 1′ , X s_A1 2′ , . . . , X s_A1 i′ , . . . . , X s_A1 h′ to acquire a feature X s_A1 output by the first DAFF module;

c-5) inputting the attention feature map X s into the convolutional layer of the second DAFF module to acquire a feature X sc 2 ; inputting the feature X sc 2 into the fully connected layer of the second DAFF module to acquire a new vector X s v′ , a new vector X s k′ , and a new vector X s q′ ; inputting the new vector X s v′ , the new vector X s k′ , and the new vector X s q′ into the h parallel self-attention mechanisms of the second DAFF module to acquire h features X s_B1 1 , X s_B1 2 , . . . , X s_B1 i , . . . . , X s_B1 h , wherein X s_B1 i denotes a feature output by an i-th self-attention mechanism, i∈{1, . . . , h};

c-6) forming the attention fusion module of the second DAFF module sequentially with a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; inputting the h features X s_B1 1 , X s_B1 2 , . . . , X s_B1 i , . . . . , X s_B1 h into the first convolutional layer to acquire features X s_B1_c 1 , X s_B1_c 2 , . . . , X s_B1_c i , . . . . , X s_B1_c h ; and inputting the features X s_B1_c 1 , X s_B1_c 2 , . . . , X s_B1_c i , . . . . , X s_B1_c h into the BN layer, the ReLU activation function layer, and the second convolutional layer to acquire new features X s_B1 1′ , X s_B1 2′ , . . . , X s_B1 i′ , . . . . , X s_B1 h′ ;

c-7) superposing the new features X s_B1 1′ , X s_B1 2′ , . . . , X s_B1 i′ , . . . . , X s_B1 h′ to acquire a feature X s_B1 output by the second DAFF module;

c-8) setting a multi-scale feature attention fusion module, wherein the multi-scale feature attention fusion module sequentially comprises an average pooling layer, a first convolutional layer, a BN layer, a ReLU activation function layer, and a second convolutional layer; inputting the feature X s_A1 into the multi-scale feature attention fusion module to acquire a global attention feature map X′ s_A1 ; and inputting the feature X s_B1 into the multi-scale feature attention fusion module to acquire a global attention feature map X′ s_B1 ; and

c-9) fusing the global attention feature map X′ s_A1 with the global attention feature map X′ s_B1 to acquire the feature X′ s .

5. The multi-label ECG signal classification method based on the improved attention mechanism according to claim 4 , wherein

in step c-2), the convolutional layer of the first DAFF module is provided with a 3×3 convolution kernel;

in step c-3), in the attention fusion module of the first DAFF module, the first convolutional layer is provided with a 1×1 convolution kernel, and the second convolutional layer is provided with a 1×1 convolution kernel;

in step c-5), the convolutional layer of the second DAFF module is provided with a 3×3 convolution kernel;

in step c-6), in the attention fusion module of the second DAFF module, the first convolutional layer is provided with a 1×1 convolution kernel, and the second convolutional layer is provided with a 1×1 convolution kernel; and

in step c-8), the multi-scale feature attention fusion module is provided with a 1×1 average pooling layer, the first convolutional layer is provided with a 1×1 convolution kernel, and the second convolutional layer is provided with a 1×1 convolution kernel.

6. The multi-label ECG signal classification method based on the improved attention mechanism according to claim 1 , wherein step d) comprises: forming the classification module with a multi-layer perceptron (MLP) and a softmax function, wherein the MLP sequentially comprises a layer normalization (LN) layer and a fully connected layer; inputting the feature X′ s into the MLP; and normalizing, by the softmax function, an output result of the MLP into probabilities of different categories in an interval [0,1].

Assignments (3)

Priority Claims (1)

Continuity (1)

References Cited (14)