WO2021114704A1 - 高阶多项式激活函数的st段分类神经网络及其应用 - Google Patents

高阶多项式激活函数的st段分类神经网络及其应用 Download PDF

Info

Publication number
WO2021114704A1
WO2021114704A1 PCT/CN2020/109292 CN2020109292W WO2021114704A1 WO 2021114704 A1 WO2021114704 A1 WO 2021114704A1 CN 2020109292 W CN2020109292 W CN 2020109292W WO 2021114704 A1 WO2021114704 A1 WO 2021114704A1
Authority
WO
WIPO (PCT)
Prior art keywords
output
layer
dimensional
neural network
size
Prior art date
Application number
PCT/CN2020/109292
Other languages
English (en)
French (fr)
Inventor
朱俊江
陈红岩
黄浩
范婵娇
Original Assignee
上海数创医疗科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海数创医疗科技有限公司 filed Critical 上海数创医疗科技有限公司
Publication of WO2021114704A1 publication Critical patent/WO2021114704A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7225—Details of analog processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device

Definitions

  • This application belongs to the technical field of electrocardiogram processing, and in particular relates to a high-order polynomial activation function ST segment classification neural network and its application.
  • the ST segment is the segment from the J point at the end of the QRS complex to the beginning of the T wave. It is an important part of the ECG.
  • the normal ST segment floats upright and shallowly with the T wave.
  • the changes of ST segment include: ST segment elevation, ST segment depression, ST segment extension and ST segment shortening. Combined with clinical symptoms, ST segment changes can be used as an important basis for the diagnosis of myocardial infarction.
  • the importance of the ST segment is whether it is depressed or elevated.
  • the ST segment has the characteristics of low frequency and small amplitude. During the ECG detection process, its shape is easily affected by external noise and changes, which increases the difficulty of analysis. . Therefore, accurate positioning of the ST segment and quantitative analysis of the waveform are the keys to the rapid diagnosis of corresponding heart diseases.
  • the current ECG detection algorithms are mainly used in the classification and recognition of ECG signals, and the automatic diagnosis algorithm for ST-segment waveforms has a low maturity.
  • the mainstream activation function is selected as the activation function of the model, it is often difficult to select regularization parameters, and the generalization ability of the model is poor.
  • the technical problem to be solved by the present invention is to provide a high-order polynomial activation function ST segment classification neural network and its application in order to solve the deficiencies in the prior art.
  • a high-order polynomial activation function ST segment classification neural network including:
  • the multi-lead electrocardiograms include normal electrocardiograms and multiple ST-segment abnormal types, a total of t types.
  • the sampling frequency of the multi-lead electrocardiograms is the same or is preprocessed into Same, and the multi-lead ECG is filtered by the same mean filter, and then filtered by the wavelet filtering method, and the ECG signals of different leads in the multi-lead ECG signal are spliced to form a new long-chain ECG signal as
  • the element of the ECG signal training set; the ECG signal is calibrated according to the type of label vector. Different types of label vectors are different and all are (a1, a2, ..., at), only in a1, a2, ..., at One is 1, the others are 0;
  • the output of the fully connected layer is a label vector (a1, a2,..., at), only one of a1, a2,..., at is 1, and the rest are 0.
  • the ST segment classification neural network of the high-order polynomial activation function of the present invention Preferably, the ST segment classification neural network of the high-order polynomial activation function of the present invention
  • the convolutional layer and the pooling layer share a total of 18 layers of Layer1-Layer18;
  • Layer1 is a one-dimensional convolutional layer, in which there are 5 convolution kernels with a size of 37, and the output is 5 one-dimensional vectors;
  • Layer2 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer3 is a one-dimensional convolutional layer, in which there are 5 convolution kernels with a size of 31, and the output is 5 one-dimensional vectors;
  • Layer4 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer5 is a one-dimensional convolution layer, in which there are 5 convolution kernels with a size of 29, and the output is 5 one-dimensional vectors of 14948;
  • Layer6 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer7 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 23, and the output is 10 one-dimensional vectors;
  • Layer8 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer9 is a one-dimensional convolutional layer, in which there are 10 convolution kernels each with a size of 15, and the output is 10 one-dimensional vectors;
  • Layer10 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer11 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 9, and the output is 10 one-dimensional vectors;
  • Layer12 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer13 is a one-dimensional convolution layer, in which there are 10 convolution kernels of size 5, and the output is 10 one-dimensional vectors;
  • Layer14 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer15 is a one-dimensional convolutional layer, in which there are 10 convolution kernels with a size of 1, and the output is 10 one-dimensional vectors;
  • Layer16 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer17 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 1, and the output is 10 one-dimensional vectors;
  • Layer18 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors.
  • the convolutional layer and the pooling layer use the ReLU function as the activation function.
  • the ST segment classification neural network of the high-order polynomial activation function of the present invention splices the ECG signals of different leads in the multi-lead ECG signal to form a new long chain ECG signal, and the ECG signal is based on the different leads. Perform permutation and combination in order to obtain the long-chain ECG signals in all permutations, and use the long-chain ECG signals in all permutations as the training set.
  • the ST segment classification neural network of the high-order polynomial activation function of the present invention splices the ECG signals of different leads in the multi-lead ECG signals to form a new long chain ECG signal, and the ECG signals are determined according to different leads. Permutation and combination in order to obtain all the long-chain ECG signals in the permutation order, and the long-chain ECG signals in each different order form a training set separately, and several different convolutional neural networks are trained through different training sets.
  • the ST segment abnormal types include ST segment level elevation, ST segment level depression and ST arch back elevation, which together form 4 types with normal electrocardiogram.
  • the training algorithms used in training are: stochastic gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm, and Adamax algorithm.
  • the present invention also provides an application method of ST segment classification neural network of high-order polynomial activation function, which includes the following steps:
  • S3 Determine the type of multi-lead ECG according to the label vector output by the ST segment classification neural network of the high-order polynomial activation function.
  • the present invention also provides an application method of the ST segment classification neural network of high-order polynomial activation function.
  • the ST-segment classification neural network using the above-mentioned high-order polynomial activation function includes the following steps:
  • S1 Obtain a multi-lead ECG signal, the sampling frequency of the multi-lead ECG is the same or preprocessed to be the same, and the multi-lead ECG is filtered by the same mean filter, and then filtered by the wavelet filtering method , And splicing the ECG signals of different leads in the multi-lead ECG signal to form a new long-chain ECG signal, and arrange and combine according to the order of the different leads to obtain the long-chain ECG signals in all permutations;
  • S2 Input the long-chain ECG signals in different orders into the convolutional neural network trained by the training set in the corresponding order;
  • the ST segment classification neural network of the high-order polynomial activation function of the present application has a good effect of removing random noise through mean filtering, and wavelet filtering removes the baseline, and finally removes the noise in the signal. And through the combination based on convolutional neural network + high-order polynomial activation function, the great divergence of high-order polynomial function can directly increase the complexity of the model, avoid the problem of hyperparameter selection in the regularization process, and significantly improve The generalization ability of neural networks.
  • FIG. 1 is a schematic diagram of the ST segment classification neural network structure of a high-order polynomial activation function according to an embodiment of the present application
  • Figure 2 is a comparison diagram of the effect when using a higher-order polynomial function and a Sigmoid function.
  • This embodiment provides a high-order polynomial activation function ST segment classification neural network, as shown in FIG. 1, including:
  • ST-segment abnormality types include ST-segment levels
  • Elevation ST-segment depression
  • the output of the fully connected layer is a label vector (a1, a2,..., at), only one of a1, a2,..., at is 1, and the rest are 0.
  • the convolutional layer and the pooling layer share a total of 18 layers of Layer1-Layer18;
  • Layer1 is a one-dimensional convolutional layer, in which there are 5 convolution kernels with a size of 37, and the output is 5 one-dimensional vectors;
  • Layer2 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer3 is a one-dimensional convolutional layer, in which there are 5 convolution kernels with a size of 31, and the output is 5 one-dimensional vectors;
  • Layer4 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer5 is a one-dimensional convolution layer, in which there are 5 convolution kernels with a size of 29, and the output is 5 one-dimensional vectors of 14948;
  • Layer6 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 5 one-dimensional vectors;
  • Layer7 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 23, and the output is 10 one-dimensional vectors;
  • Layer8 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer9 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 15, and the output is 10 one-dimensional vectors;
  • Layer10 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer11 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 9, and the output is 10 one-dimensional vectors;
  • Layer12 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer13 is a one-dimensional convolution layer, in which there are 10 convolution kernels of size 5, and the output is 10 one-dimensional vectors;
  • Layer14 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer15 is a one-dimensional convolutional layer, in which there are 10 convolution kernels with a size of 1, and the output is 10 one-dimensional vectors;
  • Layer16 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors;
  • Layer17 is a one-dimensional convolution layer, in which there are 10 convolution kernels with a size of 1, and the output is 10 one-dimensional vectors;
  • Layer18 is a one-dimensional maximum pooling layer, in which the size and step length of the core are both 2, and the output is 10 one-dimensional vectors.
  • the convolutional layer and the pooling layer preferably use the ReLU function as the activation function.
  • the ECG signals of different leads in the multi-lead ECG signals are spliced to form a new long-chain ECG signal.
  • the ECG signals are arranged and combined in the order of the different leads to obtain the long-chain hearts in all permutations.
  • For electrical signals use all the long-chain ECG signals in the sequence as the training set. That is to say, the final training set can get 12! (Factorial) times the long-chain ECG signal data, and part of the long-chain ECG signal in the sequence can also be used as the training set.
  • This embodiment provides a high-order polynomial activation function ST segment classification neural network, including: consisting of multiple neural networks; the number of neural networks is factorial times the number of leads;
  • composition of each neural network is the same as in Example 1;
  • the training set of each neural network comes from a long chain of ECG signals in a sort order.
  • the link sequence of the long-chain ECG signal can be:
  • This embodiment provides an application method of a high-order polynomial activation function ST segment classification neural network, which includes the following steps:
  • S1 Obtain a multi-lead ECG signal, the sampling frequency of the multi-lead ECG is the same or preprocessed to be the same, and the multi-lead ECG is filtered by the same mean filter, and then filtered by the wavelet filtering method (The filtering method is the same as the filtering method used in the training of the ST segment classification neural network of the high-order polynomial activation function), and splicing the ECG signals of different leads in the multi-lead ECG signal to form a new long-chain ECG signal;
  • S2 Input the new long-chain ECG signal into the ST segment classification neural network of the high-order polynomial activation function as described in Example 1;
  • S3 Determine the type of multi-lead ECG according to the type represented by the label vector.
  • the corresponding normal ECG, ST-segment horizontal elevation, ST-segment horizontal depression, and ST-segment arched back elevation are labeled as (1,0,0,0), (0,1, 0,0),(0,0,1,0),(0,0,0,1)
  • the output is (1,0,0,0),(0,1,0,0),(0 ,0,1,0), (0,0,0,1) represent normal ECG, ST-segment elevation, ST-segment horizontal depression, and ST-segment arched back elevation, respectively).
  • This embodiment provides an application method of the ST-segment classification neural network of high-order polynomial activation function.
  • the ST-segment classification neural network using the high-order polynomial activation function as described in Embodiment 2 includes the following steps:
  • S1 Obtain a multi-lead ECG signal, the sampling frequency of the multi-lead ECG is the same or preprocessed to be the same, and the multi-lead ECG is filtered by the same mean filter, and then filtered by the wavelet filtering method , And splicing the ECG signals of different leads in the multi-lead ECG signal to form a new long-chain ECG signal, arrange and combine according to the order of the different leads to obtain all the long-chain ECG signals in the arrangement order (to be identified);
  • 12 leads have 12! (Here! is the meaning of factorial) A neural network, when the ECG signal is preprocessed, especially when splicing in the order of different leads, it will form 12! A long-chain ECG signal, input all the long-chain ECG signals into the corresponding neural network, you can get 12! A label vector.
  • the label vector it is necessary to judge whether a certain label vector exceeds 80% of the total number, and the label vector exceeding 80% is used as the type result of the multi-lead ECG , If there is no more than 80% of the label vector, it will prompt an alarm and cannot be accurately identified.
  • this application can be provided as methods, systems, or computer program products. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
  • computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment.
  • the instructions provide steps for implementing the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Signal Processing (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Pathology (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Cardiology (AREA)
  • Power Engineering (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

一种高阶多项式激活函数的ST段分类神经网络，通过均值滤波去除随机噪声效果好，小波滤波去除基线，最终去除信号中的噪声。并通过基于卷积神经网络与高阶多项式型激活函数的组合，以高阶多项式函数的极大的发散性可以直接增加模型的复杂度，避开正则化过程中超参数选择问题，进而显著提高了神经网络的泛化能力。

Description

高阶多项式激活函数的ST段分类神经网络及其应用 技术领域
本申请属于心电图处理技术领域，尤其是涉及一种高阶多项式激活函数的ST段分类神经网络及其应用。
背景技术
ST段是自QRS波群终了的J点开始至T波开始的一段，它是心电图中重要组成部分。正常的ST段是随T波的直立而浅浅的上飘。ST段的改变包括：ST段抬高，ST段压低，ST段延长和ST段缩短。结合临床症状，ST段的改变可以作为诊断心肌梗塞的重要依据。ST段的重要性在于它是否存在压低或抬高，而ST段存在频率低、幅值较小的特点，在心电检测的过程中其形态易受外界噪声干扰而产生变化，从而增加了分析难度。因此，ST段的准确定位和波形定量分析是快速诊断相应心脏疾病的关键。当前的心电检测算法主要应用于心电信号的分类识别，对于ST段波形的自动诊断算法成熟度较低。选取主流的激活函数作为模型的激活函数的时候，往往出现正则化参数选取困难，模型泛化能力差。
发明内容
本发明要解决的技术问题是：为解决现有技术中的不足，从而提供一种高阶多项式激活函数的ST段分类神经网络及其应用。
本发明解决其技术问题所采用的技术方案是：
一种高阶多项式激活函数的ST段分类神经网络，包括：
若干组卷积层和池化层以及一个扁平层和一个全连接层；
收集充分多条已知类型的临床静息多导联心电图，多导联心电图包括正常心电图及多个ST段异常类型共t个类型，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号作为心电信号训练集的元素；将心电信号依据类型标定标签向量，不同种类型的标签向量不同且均为(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0；
所述全连接层包括一个隐藏层和输出层，其中隐藏层中的激励函数为高阶多项式型激活函数f(x)＝a nx n+a n-1x n-1+…+a 1x+a 0，n为自然数，输出层中的激励函数采用softmax函数；所述全连接层中的损失函数为categorical_crossentropy；
采用训练集的长链心电信号数据作为输入及其对应的标签向量作为输出对卷积神经网络进行训练，确定每个卷积层、池化层、扁平层和全连接层的参数；
全连接层的输出为标签向量(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，高阶多项式型激活函数中n＝5，且a 5＝0.001，a 3＝-0.0060，a 1＝0.2003，a 0＝0.5000。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，
所述卷积层和池化层共有Layer1-Layer18共18层；
Layer1为一维卷积层，其中共有5个大小均为37的卷积核，输出为5个一 维向量；
Layer2为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer3为一维卷积层，其中共有5个大小均为31的卷积核，输出为5个一维向量；
Layer4为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer5为一维卷积层，其中共有5个大小均为29的卷积核，输出为14948的5个一维向量；
Layer6为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer7为一维卷积层，其中共有10个大小均为23的卷积核，输出为10个一维向量；
Layer8为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer9为一维卷积层，其中共有10个大小均为15的卷积核，输出为10个一维向量；
Layer10为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer11为一维卷积层，其中共有10个大小均为9的卷积核，输出为10个一维向量；
Layer12为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer13为一维卷积层，其中共有10个大小均为5的卷积核，输出为10个一维向量；
Layer14为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer15为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
Layer16为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer17为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
Layer18为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，所述卷积层和池化层以ReLU函数作为激活函数。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，对心电信号按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号，将所有排列顺序下的长链心电信号一同作为训练集。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，拼接多导联心电信号中不同导联心电信号形成新的长链心电信号时，对心电信号按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号，并且每种不同顺序 下的长链心电信号分别形成训练集，通过不同的训练集训练得到若干不同的卷积神经网络。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，ST段异常类型包括ST段水平抬高、ST段水平压低和ST弓背抬高，与正常心电图共同组成4种类型。
优选地，本发明的高阶多项式激活函数的ST段分类神经网络，训练时采用的训练算法为：随机梯度下降算法、Adam算法、RMSProp算法、Adagrad算法、Adadelta算法、Adamax算法。
本发明还提供一种高阶多项式激活函数的ST段分类神经网络的应用方法，包括以下步骤：
S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号；
S2：将新的长链心电信号输入到上述的高阶多项式激活函数的ST段分类神经网络中；
S3：根据高阶多项式激活函数的ST段分类神经网络输出的标签向量确定多导联心电图的类型。
本发明还提供一种高阶多项式激活函数的ST段分类神经网络的应用方法，使用上述的高阶多项式激活函数的ST段分类神经网络，包括以下步骤：
S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号；
S2：将不同顺序下的长链心电信号，分别输入到上述对应顺序的训练集所训练的卷积神经网络中；
S3：若所有卷积神经网络中输出的某个标签向量占卷积神经网络输出总数的80％以上时，则以该标签向量确定多导联心电图的类型。本发明的有益效果是：
本申请的高阶多项式激活函数的ST段分类神经网络，通过均值滤波去除随机噪声效果好，小波滤波去除基线，最终去除信号中的噪声。并通过基于卷积神经网络+高阶多项式型激活函数的组合，以高阶多项式函数的极大的发散性可以直接增加模型的复杂度，避开正则化过程中超参数选择问题，进而显著提高了神经网络的泛化能力。
附图说明
下面结合附图和实施例对本申请的技术方案进一步说明。
图1是本申请实施例的高阶多项式激活函数的ST段分类神经网络结构示意图；
图2是采用高阶多项式函数与Sigmoid函数时的效果对比图。
具体实施方式
需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
下面将参考附图并结合实施例来详细说明本申请的技术方案。
实施例1
本实施例提供一种高阶多项式激活函数的ST段分类神经网络，如图1所示， 包括：
若干组卷积层和池化层以及一个扁平层和一个全连接层；
收集充分多条已知类型的临床静息多导联心电图(比如12导联)，多导联心电图包括正常心电图及多个ST段异常类型共t个类型(比如ST段异常类型包括ST段水平抬高、ST段水平压低和ST弓背抬高，与正常心电图共同组成t＝4种类型，为达到充分多条，通常每种类型的数据都包括了至少15000条)，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波(均值滤波器的参数可以为5，小波滤波可以为‘db4’、level＝8)，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号作为心电信号训练集的元素；将心电信号依据类型标定标签向量，不同种类型的标签向量不同且均为(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0；(比如将对应的正常心电图、ST段水平抬高、ST段水平压低和ST段弓背抬高这四种类型分别标签化为(1,0,0,0),(0,1,0,0),(0,0,1,0),(0,0,0,1))
所述全连接层包括一个隐藏层和输出层，其中隐藏层中的激励函数为高阶多项式型激活函数f(x)＝a nx n+a n-1x n-1+…+a 1x+a 0，n为自然数，输出层中的激励函数采用softmax函数；所述全连接层中的损失函数为categorical_crossentropy；
采用训练集的长链心电信号数据作为输入及其对应的标签向量作为输出对卷积神经网络进行训练，确定每个卷积层、池化层、扁平层和全连接层的参数；
全连接层的输出为标签向量(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0。
高阶多项式型激活函数中优选n＝5，且a 5＝0.001，a 3＝-0.0060，a 1＝0.2003，a 0＝0.5000，a 4和a 2为0。
所述卷积层和池化层共有Layer1-Layer18共18层；
Layer1为一维卷积层，其中共有5个大小均为37的卷积核，输出为5个一维向量；
Layer2为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer3为一维卷积层，其中共有5个大小均为31的卷积核，输出为5个一维向量；
Layer4为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer5为一维卷积层，其中共有5个大小均为29的卷积核，输出为14948的5个一维向量；
Layer6为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
Layer7为一维卷积层，其中共有10个大小均为23的卷积核，输出为10个一维向量；
Layer8为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer9为一维卷积层，其中共有10个大小均为15的卷积核，输出为10个 一维向量；
Layer10为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer11为一维卷积层，其中共有10个大小均为9的卷积核，输出为10个一维向量；
Layer12为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer13为一维卷积层，其中共有10个大小均为5的卷积核，输出为10个一维向量；
Layer14为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer15为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
Layer16为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
Layer17为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
Layer18为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量。所述卷积层和池化层优选以ReLU函数作为激活函数。
作为一种改进，拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，对心电信号按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号，将所有排列顺序下的长链心电信号一同作为训练集。也即最后训练集能够得到12！(阶乘)倍的长链心电信号数据，也可以取其中部分排列顺序的长链心电信号作为训练集。
实施例2
本实施例提供一种高阶多项式激活函数的ST段分类神经网络，包括：由多个神经网络构成；神经网络的数量为导联数量的阶乘倍；
每个神经网络的构成与实施例1中的相同；
每个神经网络的训练集来自一种排列顺序下的长链心电信号。
十二导联，即I、II、III、avL、avF、avR、v1、v2、v3、v4、v5、v6；则长链心电信号的链接顺序可以是：
1.I、II、III、avL、avF、avR、v1、v2、v3、v4、v5、v6
2.II、I、III、avL、avF、avR、v1、v2、v3、v4、v5、v6
3.III、II、I、avL、avF、avR、v1、v2、v3、v4、v5、v6
以此类推最终能够12！(此处的！为阶乘的意思)种顺序方法，也就得到12！种长链心电信号训练集，以此每种训练集训练得到12！种神经网络。
实施例3
本实施例提供一种高阶多项式激活函数的ST段分类神经网络的应用方法，包括以下步骤：
S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波(滤波方法与高阶多项式激活函数的ST段分类神经网络训练时所采用的滤波方法相同)，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号；
S2：将新的长链心电信号输入到如实施例1所述的高阶多项式激活函数的ST段分类神经网络中；
S3：根据标签向量所代表的类型确定多导联心电图的类型。(比如训练时将对应的正常心电图、ST段水平抬高、ST段水平压低和ST段弓背抬高这四种类型分别标签化为(1,0,0,0),(0,1,0,0),(0,0,1,0),(0,0,0,1)，则输出为(1,0,0,0),(0,1,0,0),(0,0,1,0),(0,0,0,1)分别代表正常心电图、ST段水平抬高、ST段水平压低和ST段弓背抬高)。
实施例4
本实施例提供一种高阶多项式激活函数的ST段分类神经网络的应用方法，使用如实施例2所述的高阶多项式激活函数的ST段分类神经网络，包括以下步骤：
S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号(待识别)；
S2：将不同顺序下的长链心电信号(待识别)，分别输入到如实施例2中对应顺序的训练集所训练的卷积神经网络中；也即按照不同的顺序输入不同的卷积神经网络中，训练时采用了何种导联排列顺序的训练集，此处就输入相应顺序得到的长链心电信号(待识别)；
S3：若所有卷积神经网络中输出的某个标签向量占卷积神经网络输出总数的80％以上时，则以该标签向量确定多导联心电图的类型。
比如：12导联具有12！(此处的！为阶乘的意思)个神经网络，心电信号进行预处理时，尤其是在以不同导联的顺序进行拼接时会形成12！个长链心电信号，将所有长链心电信号输入到对应的神经网络中，就能够得到12！个标签向量，此时，要通过标签向量判断该多导联心电图的类型，则需要判断某个标签向量是否超过总数的80％以上，以超过80％的标签向量作为多导联心电图的类型结果，如果没有超过80％的标签向量，则提示报警，无法准确识别。
以上述依据本申请的理想实施例为启示，通过上述的说明内容，相关工作人员完全可以在不偏离本项申请技术思想的范围内，进行多样的变更以及修改。本项申请的技术性范围并不局限于说明书上的内容，必须要根据权利要求范围来确定其技术性范围。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。

Claims (10)

  1. 一种高阶多项式激活函数的ST段分类神经网络，其特征在于，包括：
    若干组卷积层和池化层以及一个扁平层和一个全连接层；
    收集充分多条已知类型的临床静息多导联心电图，多导联心电图包括正常心电图及多个ST段异常类型共t个类型，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号作为心电信号训练集的元素；将心电信号依据类型标定标签向量，不同种类型的标签向量不同且均为(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0；
    所述全连接层包括一个隐藏层和输出层，其中隐藏层中的激励函数为高阶多项式型激活函数f(x)＝a nx n+a n-1x n-1+…+a 1x+a 0，n为自然数，输出层中的激励函数采用softmax函数；所述全连接层中的损失函数为categorical_crossentropy；
    采用训练集的长链心电信号数据作为输入及其对应的标签向量作为输出对卷积神经网络进行训练，确定每个卷积层、池化层、扁平层和全连接层的参数；
    全连接层的输出为标签向量(a1,a2,……,at)，a1,a2,……,at中仅有一个为1，其余为0。
  2. 根据权利要求1所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，高阶多项式型激活函数中n＝5，且a 5＝0.001，a 3＝-0.0060，a 1＝0.2003，a 0＝0.5000。
  3. 根据权利要求1或2所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，
    所述卷积层和池化层共有Layer1-Layer18共18层；
    Layer1为一维卷积层，其中共有5个大小均为37的卷积核，输出为5个一维向量；
    Layer2为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
    Layer3为一维卷积层，其中共有5个大小均为31的卷积核，输出为5个一维向量；
    Layer4为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
    Layer5为一维卷积层，其中共有5个大小均为29的卷积核，输出为14948的5个一维向量；
    Layer6为一维最大池化层，其中核的大小和步长均为2，输出为5个一维向量；
    Layer7为一维卷积层，其中共有10个大小均为23的卷积核，输出为10个一维向量；
    Layer8为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
    Layer9为一维卷积层，其中共有10个大小均为15的卷积核，输出为10个一维向量；
    Layer10为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
    Layer11为一维卷积层，其中共有10个大小均为9的卷积核，输出为10个一维向量；
    Layer12为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
    Layer13为一维卷积层，其中共有10个大小均为5的卷积核，输出为10个一维向量；
    Layer14为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
    Layer15为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
    Layer16为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量；
    Layer17为一维卷积层，其中共有10个大小均为1的卷积核，输出为10个一维向量；
    Layer18为一维最大池化层，其中核的大小和步长均为2，输出为10个一维向量。
  4. 根据权利要求3所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，所述卷积层和池化层以ReLU函数作为激活函数。
  5. 根据权利要求1-4任一项所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，对心电信号按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号，将所有排列顺序下的长链心电信号一同作为训练集。
  6. 根据权利要求1-5任一项所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，拼接多导联心电信号中不同导联心电信号形成新的长链心电信号时，对心电信号按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号，并且每种不同顺序下的长链心电信号分别形成训练集，通过不同的训练集训练得到若干不同的卷积神经网络。
  7. 根据权利要求1-6任一项所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，ST段异常类型包括ST段水平抬高、ST段水平压低和ST弓背抬高，与正常心电图共同组成4种类型。
  8. 根据权利要求1-5任一项所述的高阶多项式激活函数的ST段分类神经网络，其特征在于，训练时采用的训练算法为：随机梯度下降算法、Adam算法、RMSProp算法、Adagrad算法、Adadelta算法、Adamax算法。
  9. 一种高阶多项式激活函数的ST段分类神经网络的应用方法，其特征在于，包括以下步骤：
    S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号；
    S2：将新的长链心电信号输入到如权利要求1-8任一项所述的高阶多项式激活函数的ST段分类神经网络中；
    S3：根据高阶多项式激活函数的ST段分类神经网络输出的标签向量确定多 导联心电图的类型。
  10. 一种高阶多项式激活函数的ST段分类神经网络的应用方法，其特征在于，使用如权利要求1所述的高阶多项式激活函数的ST段分类神经网络，包括以下步骤：
    S1：获取多导联心电信号，所述多导联心电图的采样频率相同或者被预处理成相同，并且所述多导联心电图经过相同的均值滤波器进行滤波，再通过小波滤波方法进行滤波，并拼接多导联心电信号中不同导联心电信号形成新的长链心电信号，按照不同导联的顺序进行排列组合获取所有排列顺序下的长链心电信号；
    S2：将不同顺序下的长链心电信号，分别输入到如权利要求1中对应顺序的训练集所训练的卷积神经网络中；
    S3：若所有卷积神经网络中输出的某个标签向量占卷积神经网络输出总数的80％以上时，则以该标签向量确定多导联心电图的类型。
PCT/CN2020/109292 2019-12-09 2020-08-14 高阶多项式激活函数的st段分类神经网络及其应用 WO2021114704A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911250657.2A CN110974211A (zh) 2019-12-09 2019-12-09 高阶多项式激活函数的st段分类神经网络及其应用
CN201911250657.2 2019-12-09

Publications (1)

Publication Number Publication Date
WO2021114704A1 true WO2021114704A1 (zh) 2021-06-17

Family

ID=70091307

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/109292 WO2021114704A1 (zh) 2019-12-09 2020-08-14 高阶多项式激活函数的st段分类神经网络及其应用

Country Status (2)

Country Link
CN (1) CN110974211A (zh)
WO (1) WO2021114704A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110974211A (zh) * 2019-12-09 2020-04-10 上海数创医疗科技有限公司 高阶多项式激活函数的st段分类神经网络及其应用
CN111666865B (zh) * 2020-06-02 2023-05-23 上海数创医疗科技有限公司 多导联心电信号卷积神经网络分类方法及其使用方法
CN112883803B (zh) * 2021-01-20 2023-09-01 武汉中旗生物医疗电子有限公司 一种基于深度学习的心电信号分类方法、装置及存储介质
CN114662678B (zh) * 2022-05-17 2023-04-07 华中科技大学 一种基于变激活函数卷积神经网络的图像识别方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107007279A (zh) * 2017-03-17 2017-08-04 浙江大学 一种基于堆栈式自编码器的无创心内异常激动点定位方法
CN110141219A (zh) * 2019-06-20 2019-08-20 鲁东大学 基于导联融合深度神经网络的心肌梗死自动检测方法
CN110151168A (zh) * 2019-06-12 2019-08-23 上海移视网络科技有限公司 多病种识别及预警算法
US10426364B2 (en) * 2015-10-27 2019-10-01 Cardiologs Technologies Sas Automatic method to delineate or categorize an electrocardiogram
CN110522440A (zh) * 2019-08-12 2019-12-03 广州视源电子科技股份有限公司 基于分组卷积神经网络的心电信号识别装置
CN110974211A (zh) * 2019-12-09 2020-04-10 上海数创医疗科技有限公司 高阶多项式激活函数的st段分类神经网络及其应用

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102961129B (zh) * 2012-10-26 2015-11-25 上海交通大学无锡研究院 一种远程医疗的异常心电张量分析方法
CN105997055A (zh) * 2016-07-11 2016-10-12 吉林大学 一种心电信号st段的自动分类方法、系统及装置
EP3511871A4 (en) * 2016-09-06 2020-06-24 Nippon Telegraph And Telephone Corporation DEVICE, METHOD AND PROGRAM FOR EXTRACTING CHARACTERISTIC QUANTITIES OF DATA IN CHRONOLOGICAL SERIES
CN107714023B (zh) * 2017-11-27 2020-09-01 上海优加利健康管理有限公司 基于人工智能自学习的静态心电图分析方法和装置
CN110414565B (zh) * 2019-05-06 2021-06-08 北京邮电大学 一种用于功率放大器的基于Group Lasso的神经网络裁剪方法
CN110543901A (zh) * 2019-08-22 2019-12-06 阿里巴巴集团控股有限公司 图像识别方法、装置及设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10426364B2 (en) * 2015-10-27 2019-10-01 Cardiologs Technologies Sas Automatic method to delineate or categorize an electrocardiogram
CN107007279A (zh) * 2017-03-17 2017-08-04 浙江大学 一种基于堆栈式自编码器的无创心内异常激动点定位方法
CN110151168A (zh) * 2019-06-12 2019-08-23 上海移视网络科技有限公司 多病种识别及预警算法
CN110141219A (zh) * 2019-06-20 2019-08-20 鲁东大学 基于导联融合深度神经网络的心肌梗死自动检测方法
CN110522440A (zh) * 2019-08-12 2019-12-03 广州视源电子科技股份有限公司 基于分组卷积神经网络的心电信号识别装置
CN110974211A (zh) * 2019-12-09 2020-04-10 上海数创医疗科技有限公司 高阶多项式激活函数的st段分类神经网络及其应用

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
REN XIAOXIA: "ST Segment Waveform Classification Algorithm Based on Dropout Deep Convolutional Neural Network", JOURNAL OF TRANSDUCTION TECHNOLOGY, DONGNAN DAXUE CHUBANSHE, NANJING, CN, vol. 31, no. 8, 31 August 2018 (2018-08-31), CN, pages 1217 - 1222, XP055794230, ISSN: 1004-1699, DOI: 10.3969/j.issn.1004-1699.2018.08.013 *

Also Published As

Publication number Publication date
CN110974211A (zh) 2020-04-10

Similar Documents

Publication Publication Date Title
WO2021114704A1 (zh) 高阶多项式激活函数的st段分类神经网络及其应用
WO2021057328A1 (zh) 基于特征选取的st段分类卷积神经网络及其使用方法
US9949714B2 (en) Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection
Kanani et al. ECG heartbeat arrhythmia classification using time-series augmented signals and deep learning approach
KR102463764B1 (ko) 부정맥 검출 방법, 장치, 전자장치 및 컴퓨터 기억 매체
JP6986724B2 (ja) 人工知能に基づく心電図干渉識別方法
CN110046604B (zh) 一种基于残差网络的单导联ecg心律失常检测分类方法
CN110717416B (zh) 基于特征选取的st段分类识别用神经网络训练方法
CN108968941B (zh) 一种心律失常检测方法、装置及终端
CN111666865B (zh) 多导联心电信号卷积神经网络分类方法及其使用方法
CN110037683B (zh) 用于识别心律类型的改进卷积神经网络及其训练方法
Kang et al. A study on several critical problems on arrhythmia detection using varying-dimensional electrocardiography
Feng et al. Dcmn: Double core memory network for patient outcome prediction with multimodal data
Yang et al. Biot: Biosignal transformer for cross-data learning in the wild
Jia et al. Automatic detection and classification of 12-lead ECGs using a deep neural network
CN111956202B (zh) 基于样本分布改进的卷积神经网络的训练方法及模型
Gaudilliere et al. Generative pre-trained transformer for cardiac abnormality detection
Kennedy et al. Device agnostic AI-based analysis of ambulatory ECG recordings
CN112603324B (zh) 基于改进损失函数的神经网络的训练方法
Bodini et al. Classification of 12-lead ECG with an ensemble machine learning approach
CN112022140B (zh) 一种心电图的诊断结论自动诊断方法及系统
Pratiher et al. A dilated residual vision transformer for atrial fibrillation detection from stacked time-frequency ECG representations
Vázquez et al. Two will do: Convolutional neural network with asymmetric loss, self-learning label correction, and hand-crafted features for imbalanced multi-label ECG data classification
CN110897630A (zh) 一种基于迁移学习的室性心动过速心律识别方法和装置
Zhang et al. Continual Learning for Cuffless Blood Pressure Measurement using PPG and ECG Signals

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20897733

Country of ref document: EP

Kind code of ref document: A1