Electrocardiogram (ECG) signal detection and positioning method based on weakly supervised learning
An electrocardiograph (ECG) signal detection and positioning method based on weakly supervised learning is provided. A deep learning model mainly includes a multi-scale feature extraction module, a self-attention encoding module, and a classification and positioning module. An extracted original ECG signal is denoised and segmented to obtain a fixed-length pure ECG signal segment. In the convolutionally-connected multi-scale feature extraction module, a channel local attention (CLA) layer is introduced, and a PReLU activation function is used to achieve a better local information extraction capability. The self-attention encoding module is introduced to establish an association between a local feature and a global feature. The classification and positioning module is introduced to output a general location of an abnormal signal. A fusion module enables the model to map a local predicted value onto a global predicted value, and model parameters are trained on a weakly annotated dataset.
1. An electrocardiograph (ECG) detection and positioning method based on weakly supervised learning, comprising the following steps:
a-1) operating a plurality of ECG devices to obtain n pieces of ECG signal data of different lengths;
a-2) collecting the n pieces of ECG signal data of different lengths from the plurality of ECG devices to obtain an original ECG dataset X 0 , wherein X 0 ={x 1 , x 2 , . . . , x j , . . . , x n }, x j represents a j th original ECG signal record, and j∈{1, 2, . . . , n};
b) preprocessing the original ECG dataset X 0 to obtain a preprocessed ECG dataset X 0 ′, wherein X 0 ′={x 1 ′, x 2 ′, . . . , x j ′, . . . , x n ′}, and x j ′ represents a j th preprocessed original ECG signal record;
c) establishing a deep learning model, and inputting the preprocessed ECG dataset into the deep learning model to obtain a globally recorded category prediction value y pre and a predicted abnormal signal location set L; and
d) training the deep learning model;
wherein the step c) comprises the following substeps:
c-1) constituting the deep learning model by a multi-scale feature extraction module, a self-attention encoding module, and a classification and positioning module;
c-2) constituting the multi-scale feature extraction module of the deep learning model by a start convolutional block, a first multi-scale feature extraction block, and a second multi-scale feature extraction block, wherein the start convolutional block of the multi-scale feature extraction module is sequentially constituted by a first start convolutional layer, a batch normalization (BN) layer, a PRELU activation function layer, and a second start convolutional layer; and inputting the j th preprocessed original ECG signal record x j ′ into the start convolutional block of the multi-scale feature extraction module to obtain a feature map X 1 ;
c-3) constituting the first multi-scale feature extraction block of the multi-scale feature extraction module by a first BN layer, a first PRELU activation function layer, a channel local attention (CLA) layer, a first convolutional layer, a second BN layer, a second PRELU activation function layer, a dropout layer, a second convolutional layer, and a maximum pooling layer, wherein the CLA layer is constituted by a first branch and a second branch, the first branch of the CLA layer is sequentially constituted by a global average pooling layer, a first fully connected layer, a Swish activation function layer, a second fully connected layer, and a Sigmoid activation function layer, and the second branch of the CLA layer is sequentially constituted by a global maximum pooling layer, a maximum pooling layer, a first fully connected layer, a Swish activation function layer, a second fully connected layer, a Sigmoid activation function layer, and an upsampling layer; sequentially inputting the feature map X 1 into the first BN layer and the first PRELU activation function layer of the first multi-scale feature extraction block to obtain a feature map X 1_1 ; inputting the feature map X 1_1 into the first branch of the CLA layer of the first multi-scale feature extraction block to obtain a feature map X 1_1 1_2 ; inputting the feature map X 1_1 into the second branch of the CLA layer of the first multi-scale feature extraction block to obtain a feature map X 1_1 2_3 ; multiplying the feature map X 1_1 and the feature map X 1_1 1_2 to obtain a first product, and then multiplying the first product and the feature map X 1_1 2_3 to obtain a feature map X 1_2 ; sequentially inputting the feature map X 1_2 into the first convolutional layer, the second BN layer, the second PRELU activation function layer, the dropout layer, and the second convolutional layer of the first multi-scale feature extraction block to obtain a feature map X 1_3 ; and adding up the feature map X 1_3 and the feature map X 1 to obtain a first addition result, and then inputting the first addition result into the maximum pooling layer of the first multi-scale feature extraction block to obtain a feature map X 2 ;
c-4) constituting the second multi-scale feature extraction block of the multi-scale feature extraction module by a first BN layer, a first PRELU activation function layer, a CLA layer, a first convolutional layer, a second BN layer, a second PRELU activation function layer, a dropout layer, a second convolutional layer, and a maximum pooling layer; sequentially inputting the feature map X 2 into the first BN layer and the first PRELU activation function layer of the second multi-scale feature extraction block to obtain a feature map X 2_1 ; inputting the feature map X 2_1 into a first branch of the CLA layer of the second multi-scale feature extraction block to obtain a feature map X 2_1 1_2 ; inputting the feature map X 2_1 into a second branch of the CLA layer of the second multi-scale feature extraction block to obtain a feature map X 2_1 2_3 ; multiplying the feature map X 2_1 and the feature map X 2_1 1_2 to obtain a second product, and then multiplying the second product and the feature map X 2_1 2_3 to obtain a feature map X 2_2 ; sequentially inputting the feature map X 2_2 into the first convolutional layer, the second BN layer, the second PRELU activation function layer, the dropout layer, and the second convolutional layer of the second multi-scale feature extraction block to obtain a feature map X 2_3 ; and adding up the feature map X 2_3 and the feature map X 2 to obtain a second addition result, and then inputting the second addition result into the maximum pooling layer of the second multi-scale feature extraction block to obtain a feature map X 3 ;
c-5) encoding the feature map X 3 based on a predefined location to obtain a location information graph X 3 P , and concatenating the feature map X 3 and the location information graph X 3 P to obtain a feature map X 4 ;
c-6) constituting the self-attention encoding module of the deep learning model by a multi-head self-attention layer, a layer norm layer, and a multi-layer perceptron (MLP) layer, and inputting the feature map X 4 into the self-attention encoding module to obtain a feature map X 5 ;
c-7) constituting the classification and positioning module by a local classification module, a fusion module, and an abnormal location output module;
c-8) constituting the local classification module of the classification and positioning module sequentially by a convolutional layer, a flatten layer, a fully connected layer, and a Softmax function; and inputting the feature map X 5 into the local classification module to obtain a predicted value y c pre of a local location, wherein y c pre ={y 1 pre , y 2 pre , . . . , y l pre , . . . y l pre }, y l pre represents a predicted value output by the fully connected layer for an l th location, l∈{1, 2, . . . , m}, m represents a pre quantity of predicted values output by the fully connected layer, y l pre =0 indicates a normal ECG signal, and y l pre =1 indicates an abnormal ECG signal;
c-9) inputting the predicted value y c pre of the local location into the fusion module of the classification and positioning module to obtain the globally recorded category prediction value y pre ; and
c-10) inputting the predicted value y c pre of the local location into the abnormal location output module of the classification and positioning module, wherein the abnormal location output module determines whether each location in the predicted value y c pre of the local location has an abnormal signal, wherein the abnormal signal reflects an activity state of a heart, and outputs the abnormal signal location set L, wherein L={l, l 2 , . . . , l, . . . , l k }, l i represents coordinates of a location of an i th abnormal signal in the predicted value y c pre of the local location, i∈{1, 2, . . . , k}, and k represents a quantity of predicted abnormal signals;
wherein in the step c-9), the globally recorded category prediction value y pre is calculated according to a formula y pre =(1/b 2 )log((1/m)Σ l=1 m exp(by l pre )), wherein b is a constant.
2. The ECG signal detection and positioning method based on weakly supervised learning according to claim 1 , wherein the step b) comprises the following substeps:
b-1) removing noise of another wave band from the original ECG dataset X 0 by using a bandpass filter; and
b-2) slicing the original ECG dataset X 0 without the noise of the another wave band into fixed-length signals by using a sliding window, to obtain the preprocessed ECG dataset X 0 ′.
3. The ECG signal detection and positioning method based on weakly supervised learning according to claim 2 , wherein in the step b-1), the noise of the another wave band is removed by using a 3-45 Hz bandpass filter.
4. The ECG signal detection and positioning method based on weakly supervised learning according to claim 1 , wherein
in the step c-2), the first start convolutional layer of the start convolutional block has a 1*1 convolution kernel, with a step of 1, and the second start convolutional layer has a 3*3 convolution kernel, with a step of 2;
in the step c-3), the first convolutional layer and the second convolutional layer of the first multi-scale feature extraction block each are a one-dimensional convolution with an 8*8 convolution kernel, a parameter size of the dropout layer of the first multi-scale feature extraction block is 0.2, and the maximum pooling layer of the first multi-scale feature extraction block has a convolution kernel whose size is 2, with a step of 2;
in the step c-3), an expansion coefficient of the first fully connected layer in the first branch of the CLA layer is 4, a parameter of the global maximum pooling layer in the second branch of the CLA layer is set to 1, the maximum pooling layer in the second branch of the CLA layer has a convolution kernel whose size is 32, with a step of 32, an expansion coefficient of the first fully connected layer in the second branch of the CLA layer is 4, and a parameter of the upsampling layer in the CLA layer is 32;
in the step c-4), an expansion coefficient of the first fully connected layer in the first branch of the CLA layer is 4, a parameter of the global maximum pooling layer in the second branch of the CLA layer is set to 1, the maximum pooling layer in the second branch of the CLA layer has a convolution kernel whose size is 32, with a step of 32, an expansion coefficient of the first fully connected layer in the second branch of the CLA layer is 4, and a parameter of the upsampling layer in the CLA layer is 32; and
in the step c-6), a multi-head self-attention in the MLP layer has 6 heads.
5. The ECG signal detection and positioning method based on weakly supervised learning according to claim 1 , wherein in the step d), a deep learning module is optimized by using an Adam optimizer based on a cross-entropy loss function, to obtain a trained deep learning model.
6. The ECG signal detection and positioning method based on weakly supervised learning according to claim 1 , wherein b=3.