# US 2025/0143647 A1 · Multi-lead electrocardiogram (ecg) signal classification method based on self-supervised learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0143647 A1
- Document ID: US20250143647A1
- Status: Published application
- Kind: A1
- Application number: 18/664,341
- Filed: 2024-05-15
- Published: 2025-05-08
- Earliest claimed: 2023-11-08
- Assignees: Qilu University of Technology (Shandong Academy of Sciences); Shandong Computer Science Center (National Supercomputing Center In Jinan)
- Applicants: Qilu University of Technology (Shandong Academy of Sciences); Shandong Computer Science Center (National Supercomputing Center In Jinan)
- Inventors: Yinglong WANG; Wei LIU; Minglei SHU; Pengyao XU; Shuwang ZHOU; Zhaoyang LIU
- Web page: https://patentgrep.com/patent/US20250143647A1

## CPC

- A61B 5/7267: {involving training the classification device}
- A61B 5/346: Analysis of electrocardiograms
- G06N 3/045: Combinations of networks
- G06N 3/048: Activation functions
- G06N 3/08: Learning methods
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G16H 50/20: for computer-aided diagnosis, e.g. based on medical expert systems
- G16H 50/70: for mining of medical data, e.g. analysing previous cases of other patients
- Y02A 90/10: Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
- A61B: DIAGNOSIS; SURGERY; IDENTIFICATION (usefulness limited to only animals A61D)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G16H: HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- Y02A: TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE

## Abstract

A multi-lead electrocardiogram (ECG) signal classification method based on self-supervised learning relates to the technical field of ECG signal classification. The method includes: processing an original signal through different data augmentation methods, designing an appropriate encoder module, extracting a feature of an ECG signal through a large amount of easily available unlabeled data such that an encoder learns more class information of the ECG signal, fine-tuning the model encoder with a small amount of labeled data for feature optimization, and continuously optimizing a parameter of a feature extractor by training a model such that a generated feature well reflects a structure and information of input data. Through self-supervised learning, the method reduces obstacles caused by performing ECG signal classification through a large amount of expensive manually labeled data, improving the generalization ability of the model.

## Claims

### Claim at position 1

1-10. (canceled)

### Claim 11

A multi-lead electrocardiogram (ECG) signal classification method based on self-supervised learning, comprising the following steps:
a) acquiring original ECG signal data X;
b) preprocessing the original ECG signal data X to acquire a preprocessed ECG signal X′;
c) constructing a data augmentation unit, inputting the preprocessed ECG signal X′ into the data augmentation module, and outputting an augmented positive sample Xa and an augmented negative sample Ya;
d) constructing an encoder module; inputting the augmented positive sample Xa into the encoder module, and outputting an encoded sequence Xc; and inputting the augmented negative sample Ya into the encoder module, and outputting an encoded sequence Yc;
e) constructing a feature dimensionality reconstruction module; inputting the encoded sequence Xc into the feature dimensionality reconstruction module, and outputting a dimensionality-reconstructed lead feature Xg; and inputting the encoded sequence Yc into the feature dimensionality reconstruction module, and outputting a dimensionality-reconstructed lead feature Yg;
f) training, by an adaptive moment estimation (Adam) optimizer, the encoder module through a contrastive learning loss to acquire a trained encoder module;
g) replacing the original ECG signal data X in the step b) with ECG signal data Z that needs classification, and repeating the step b) to acquire a preprocessed ECG signal Z′;
h) inputting the preprocessed ECG signal Z′ into the trained encoder module, and outputting an encoded sequence Zc;
i) replacing the encoded sequence Xc in the step e) with the encoded sequence Zc, and repeating the step e) to acquire a dimensionality-reconstructed lead feature Zg; and
j) acquiring a final class based on the dimensionality-reconstructed lead feature Zg;
wherein the step d) comprises:
d-1) constructing the encoder module, comprising a one-dimensional convolutional layer, a time-domain fusion module temporal convolutional fusion network (TCFN), an average pooling layer, and a Dropout layer;
d-2) inputting the augmented positive sample Xa into the one-dimensional convolutional layer, and outputting a feature Xc1; and inputting the augmented negative sample Ya into the one-dimensional convolutional layer, and outputting a feature Yc1;
d-3) dividing the time-domain fusion module TCFN into a first branch, a second branch, and a third branch;
d-4) constructing the first branch, comprising a first convolutional block, a second convolutional block, a third convolutional block, and a fourth convolutional block, wherein each of the first convolutional block, the second convolutional block, the third convolutional block, and the fourth convolutional block comprises a first convolutional layer, a second convolutional layer, a BatchNormal1d layer, a rectified linear unit (ReLU) layer, a MaxPool1d layer, and a Dropout layer in sequence; inputting the feature Xc1 into the first convolutional block in the first branch, and outputting a feature Xc11; inputting the feature Xc11 into the second convolutional block in the first branch, and outputting a feature Xc12; inputting the feature Xc12 into the third convolutional block in the first branch, and outputting a feature Xc13; inputting the feature Xc13 into the fourth convolutional block in the first branch, and outputting a feature X′c1; inputting the feature Yc1 into the first convolutional block in the first branch, and outputting a feature Yc11; inputting the feature Yc11 into the second convolutional block in the first branch, and outputting a feature Yc12; inputting the feature Yc12 into the third convolutional block in the first branch, and outputting a feature Yc13; and inputting the feature Yc13 into the fourth convolutional block in the first branch, and outputting a feature Y′c1;
d-5) constructing the second branch, comprising a lead memory module LM_Block and a skip connection module, wherein the lead memory module LM_Block comprises a global pooling layer, a long short-term memory (LSTM) layer, a one-dimensional convolutional layer, and a Relu activation function in sequence; inputting the feature Xc1 into the lead memory module LM_Block, and outputting a feature Xc11; inputting the feature Xc11 and the feature Xc1 into the skip connection module for a first addition operation to acquire a feature X″c1; inputting the feature Yc1 into the lead memory module LM_Block, and outputting a feature Yc11; and inputting the feature Yc11 and the feature Yc1 into the skip connection module for a second addition operation to acquire a feature Y″c1;
d-6) constructing the third branch, comprising an attention module, a first one-dimensional convolutional layer, a second one-dimensional convolutional layer, a third one-dimensional convolutional layer, and a maximum pooling layer; inputting the feature Xc1 into the attention module, and performing, by a torch.nn.MultiheadAttention function in python, a first attention operation on the feature Xc1 to obtain a first resulting feature; inputting the first resulting feature into the first one-dimensional convolutional layer, the second one-dimensional convolutional layer, the third one-dimensional convolutional layer, and the maximum pooling layer in sequence, and outputting an encoded sequence Y′″c1; inputting the feature Yc1 into the attention module, and performing, by the torch.nn.MultiheadAttention function in python, a second attention operation on the feature Yc1 to obtain a second resulting feature; and inputting the second resulting feature into the first one-dimensional convolutional layer, the second one-dimensional convolutional layer, the third one-dimensional convolutional layer, and the maximum pooling layer in sequence, and outputting an encoded sequence Y′″c1; and
d-7) subjecting the feature X′c1, the feature X″c1, and the encoded sequence X′″c1 to addition and averaging to acquire a feature Xc2; subjecting the feature Y′c1, the feature Y″c1, and the encoded sequence Y′″c1 to addition and averaging to acquire a feature Yc2; inputting the feature Xc2 into the average pooling layer and the Dropout layer of the encoder module in sequence, and outputting the encoded sequence Xc; and inputting the feature Yc2 into the average pooling layer and the Dropout layer of the encoder module in sequence, and outputting the encoded sequence Yc.

### Claim 12 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step a) comprises: acquiring an ECG signal from a PhysioNet2020 dataset, resampling the ECG signal by a resample( ) function in a Python library sklearn.utils, and adjusting a length of the ECG signal to L sample points to acquire the original ECG signal data X, X∈RC×L, wherein R denotes a real number space, and C denotes a number of channels.

### Claim 13 (depends on 12)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 12, wherein C=12, and L=1000.

### Claim 14 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step b) comprises:
b-1) calling, by python, a signal.butter method from a scipy module; and performing, by a Butterworth bandpass filter with a cutoff frequency comprising a low frequency of 0.5 Hz and a high frequency of 49 Hz, denoising and baseline wander removal on the original ECG signal data X; and
b-2) resampling, by a resample( ) function in a Python library sklearn.utils, an acquired ECG signal for each lead in the original ECG signal data X after denoising and baseline wander removal; and adjusting a length of the acquired ECG signal to L to acquire the preprocessed ECG signal X′.

### Claim 15 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step c) comprises:
c-1) constructing the data augmentation unit, comprising a first augmentation module and a second augmentation module;
c-2) inputting the preprocessed ECG signal X′ into the first augmentation module, and multiplying a value of each lead signal in the preprocessed ECG signal X′ by −1 to acquire the augmented positive sample Xa; and
c-3) inputting the preprocessed ECG signal X′ into the second augmentation module, and performing data augmentation on the preprocessed ECG signal X′ through Gaussian noise to acquire the augmented negative sample Ya.

### Claim 16 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein in the step d-1), the one-dimensional convolutional layer of the encoder module comprises a convolutional kernel with a size of 3; in the step d-4), each of the first convolutional layer and the second convolutional layer in the first convolutional block comprises a convolutional kernel with a size of 64, each of the first convolutional layer and the second convolutional layer in the second convolutional block comprises a convolutional kernel with a size of 32, each of the first convolutional layer and the second convolutional layer in the third convolutional block comprises a convolutional kernel with a size of 16, and each of the first convolutional layer and the second convolutional layer in the fourth convolutional block comprises a convolutional kernel with a size of 3; in the step d-5), in the lead memory module LM_Block, the LSTM layer has a hidden_size=256, and the one-dimensional convolutional layer comprises a convolutional kernel with a size of 3; and in the step d-6), in the third branch, the first one-dimensional convolutional layer comprises a convolutional kernel with a size of 32, the second one-dimensional convolutional layer comprises a convolutional kernel with a size of 16, and the third one-dimensional convolutional layer comprises a convolutional kernel with a size of 3.

### Claim 17 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step e) comprises:
e-1) constructing the feature dimensionality reconstruction module, comprising a convolutional layer, a linear layer, and a Relu activation function in sequence; and
e-2) inputting the encoded sequence Xc into the feature dimensionality reconstruction module, and outputting the dimensionality-reconstructed lead feature Xg; and inputting the encoded sequence Yc into the feature dimensionality reconstruction module, and outputting the dimensionality-reconstructed lead feature Yg.

### Claim 18 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step h) comprises:
h-1) inputting the preprocessed ECG signal Z′ into the one-dimensional convolutional layer of the trained encoder module, and outputting a feature Zc1;
h-2) inputting the feature Zc1 into the first convolutional block in the first branch of the time-domain fusion module TCFN of the trained encoder module, and outputting a feature Zc11; inputting the feature Zc11 into the second convolutional block in the first branch of the time-domain fusion module TCFN of the trained encoder module, and outputting a feature Zc12; inputting the feature Zc12 into the third convolutional block in the first branch of the time-domain fusion module TCFN of the trained encoder module, and outputting a feature Zc13; and inputting the feature Zc13 into the fourth convolutional block in the first branch of the time-domain fusion module TCFN of the trained encoder module, and outputting a feature Z′c1, h-3) inputting the feature Zc1 into the lead memory module LM_Block in the second branch of the time-domain fusion module TCFN of the trained encoder module, and outputting a feature Zc11; and inputting the feature Zc11 and the feature Zc1 into the skip connection module in the second branch of the time-domain fusion module TCFN of the trained encoder module for a third addition operation to acquire a feature Z″c1;
h-4) inputting the feature Zc1 into the attention module in the third branch of the time-domain fusion module TCFN of the trained encoder module, and performing, by the torch.nn.MultiheadAttention function in python, a third attention operation on the feature Zc1 to obtain a third resulting feature; and inputting the third resulting feature into the first one-dimensional convolutional layer, the second one-dimensional convolutional layer, the third one-dimensional convolutional layer, and the maximum pooling layer in the third branch of the time-domain fusion module TCFN of the trained encoder module in sequence, and outputting an encoded sequence Z′″c1; and
h-5) subjecting the feature Z′c1, the feature Z″c1, and the encoded sequence Z′″c1 to addition and averaging to acquire a feature Zc2; and inputting the feature Zc2 into the average pooling layer and the Dropout layer of the trained encoder module in sequence, and outputting the encoded sequence Zc.

### Claim 19 (depends on 11)

The multi-lead ECG signal classification method based on self-supervised learning according to claim 11, wherein the step j) comprises:
j-1) inputting the dimensionality-reconstructed lead feature Zg into a fully connected layer and a softmax activation function in sequence, and outputting a probability distribution from 0 to 1; and
j-2) processing, by an argmax function in a python library numpy, the probability distribution, and outputting the final class.
