CN111012332A

CN111012332A - Multi-label classification method for 12-lead electrocardiosignals based on neural network

Info

Publication number: CN111012332A
Application number: CN202010065982.8A
Authority: CN
Inventors: 李灯熬; 赵菊敏; 武行
Original assignee: Taiyuan University of Technology
Current assignee: Taiyuan University of Technology
Priority date: 2020-01-20
Filing date: 2020-01-20
Publication date: 2020-04-17

Abstract

The invention relates to a 12-lead electrocardiosignal multi-label classification method based on a neural network, belonging to the technical field of computer image processing; the method aims to fully excavate and analyze potential characteristics of 12-lead signals by utilizing a neural network, identify normal and heart failure through the signals, further subdivide the heart failure aiming at the heart failure, and find out specific heart failure types so as to facilitate doctors to give medicines according to symptoms and diagnose and treat the heart failure in time; the method comprises the following specific steps: determining a data input format, improving a convolutional neural network, fusing the characteristic information of 3 tributaries through a full connection layer by data input, and finally performing label classification; the invention fully excavates the clinical information of a patient by taking a 12-lead signal as a data source, finds out a rule from RNN time sequence of a double-layer LSTM structure by utilizing improved three-branch CNN extraction to medical characteristics of different scales, and improves the multi-classification accuracy.

Description

Multi-label classification method for 12-lead electrocardiosignals based on neural network

Technical Field

The invention belongs to the technical field of computer image processing, and relates to a 12-lead electrocardiosignal multi-label classification method based on a neural network.

Background

Heart Failure (HF), abbreviated as Heart Failure, is a group of clinical syndromes caused by ventricular filling and/or ejection disorder due to structural and functional changes of myocardium caused by initial myocardial damage (such as myocardial infarction, hemodynamic overload, inflammation, etc.) of any reason. Heart failure is an important component of chronic cardiovascular diseases in the world, is the final stage of development of various heart diseases, has the characteristics of high morbidity, high cost and poor healing, and has high early mortality and high risk of hospitalization, thereby becoming a significant public health problem in the world. Because the pathogenesis of the heart failure is complex, the clinical manifestations are different, generally accepted knowledge is difficult to obtain, the type of the heart failure of the heart cannot be correctly identified, and the diagnosis and treatment of the heart failure are influenced to a certain extent.

The 12-lead electrocardiosignals are 12-dimensional signals obtained by collecting pulsation of different parts of a body through an electrocardiograph, 12 different leads play an important role in finally evaluating whether the electrocardiosignals are normal or not, however, if 12-dimensional data are simultaneously input into a model according to time, the promotion of data volume and the potential association of each lead are more complicated, and the classification result may be worse. Therefore, single lead or double leads are often adopted to complete the classification of the electrocardiosignals at present. However, such operation may miss the abundant information in the 12-lead electrocardiographic signal data, and may cause deviation of the analysis result.

Disclosure of Invention

The invention overcomes the defects of the prior art, provides a 12-lead electrocardiosignal multi-label classification method based on a neural network, fully excavates and analyzes the potential characteristics of the 12-lead electrocardiosignal by utilizing the neural network, identifies normal and heart failure through the signal, further subdivides the heart failure aiming at the heart failure, and finds out the specific heart failure type so as to facilitate the doctor to give medicines according to symptoms and diagnose and treat the diseases in time.

The invention is realized by the following technical scheme.

The multi-label classification method of 12-lead electrocardiosignals based on the neural network comprises the following specific steps:

1) determining a data input format: the number of leads is 12, the length is 3s of electrocardiosignals sampled according to 500HZ, and a data format with 12 multiplied by 1500 is formed.

2) Improving the convolutional neural network: the convolutional neural network was modified to have 3 independent streams, each consisting of, in turn: two convolutional layers with the size of 64, a first largest pooling layer, two convolutional layers with the size of 128, a second largest pooling layer, two convolutional layers with the size of 256, a third largest pooling layer, a fully-connected layer with the size of 768 and a fully-connected layer with the size of 512; the convolution kernel size for each stream is 3, 5, 7, respectively.

3) Data input: the electrocardiosignals are input from each stream in sequence, the scale characteristics of a kernel are captured from the input electrocardiogram record, and then the characteristic information of 3 branches is fused through a full connecting layer.

4) And (4) label classification: removing the Softmax layer of the improved convolutional neural network, inputting the extracted features into the improved RNN through full connection, and classifying the features into 9 types: normality, atrial fibrillation, first degree atrioventricular block, left bundle branch block, right bundle branch block, premature atrial contraction, premature ventricular contraction, ST segment depression, ST segment elevation; the RNN is composed of a double-layer LSTM structure, the length of the two-way LSTM in the previous layer is 256, the length of the two-way LSTM in the next layer is 64, and the number of the hidden units in each layer is 128.

Preferably, the step size of the maximum cell layer of the first maximum pooling layer and the second maximum pooling layer is 3.

Preferably, the step size in the third maximum pooling layer is 2.

Preferably, the improved convolutional neural network uses a cross entropy method to calculate a loss function, and uses L2 regularization to improve generalization performance.

Compared with the prior art, the invention has the beneficial effects that.

The invention adopts a deep learning method, fully excavates and analyzes the potential characteristics of 12-lead signals by utilizing a neural network, identifies normal and heart failure through the signals, further subdivides the heart failure aiming at the heart failure, and finds out specific heart failure types so as to facilitate doctors to timely diagnose and treat symptoms by giving medicines. The CNN is further optimized and organically combined by utilizing the advantages of the CNN in feature learning and the RNN in time series to solve the technical problem.

1. The invention abandons the traditional single lead identification heart failure and utilizes 12 lead signals as data sources to fully mine the clinical information of patients.

2. And extracting medical features of different scales by using the improved three-branch CNN. And the RNN of the double-layer LSTM structure is utilized to find out the rule from the time sequence, the signal is divided into normal and subdivided 8 types of abnormal situations, and the multi-classification accuracy is improved.

Drawings

FIG. 1 is a schematic representation of a 12 lead cardiac signal.

Fig. 2 is a schematic diagram of the improved convolutional network of the three tributaries of the present invention.

FIG. 3 is a diagram illustrating the classification of the two-layer long/short term memory network 9.

Detailed Description

In order to make the technical problems, technical solutions and advantageous effects to be solved by the present invention more clearly apparent, the present invention is further described in detail with reference to the embodiments and the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical solutions of the present invention are described in detail below with reference to the embodiments and the drawings, but the scope of protection is not limited thereto.

A12-lead electrocardiosignal multi-label classification method based on a neural network comprises the following specific steps:

as shown in fig. 1 and 2, the specific input is a data format of 12 × 1500, i.e. the number of leads is 12, and the length is 3s of electrocardiosignals sampled at 500 HZ. The improved convolutional neural network has 3 independent streams, each of which is composed of: two convolutional layer layers with size of 64, a maximum pooling layer, two convolutional layer layers with size of 128, a maximum pooling layer, two convolutional layer layers with size of 256, a maximum pooling layer, a fully-connected layer with size of 768, and a fully-connected layer with size of 512. The difference is that the convolution kernel size of each stream is 3, 5 and 7 respectively, and the step length is adjusted according to the actual situation. Therefore, the electrocardiosignals are respectively input from each branch, the characteristics of different scales can be captured from the input electrocardiogram record due to the difference of the nuclear sizes, then the characteristic information of 3 branches is fused through the full-connection layer, and finally the signals are divided into: normal (Normal), Atrial Fibrillation (AF), Block (Block), Premature contraction (Premature contraction), ST-segment abnormalities (ST-segment abnormalities) class 5. Since the feature information is concentrated in deeper layers, the step size of the first two largest pool layers is taken to be 3, which makes the computational complexity decrease faster, while the feature extraction performance is not affected much because the features are very shallow. The step size in the last maximum pooling layer is taken to be 2. The improved CNN network may use cross-entropy to compute the loss function, using L2 regularization to improve generalization performance.

Removing the Softmax layer of CNN and inputting the extracted features via full concatenation to a two-layer RNN variant with a first layer bi-directional LSTM length of 256 (hidden unit 128) and a second layer bi-directional LSTM length of 64 (hidden unit 128) will achieve a finer classification of 9: normal (Normal), Atrial Fibrillation (AF), low-grade atrioventricular block (I-AVB), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Premature Atrial Contraction (PAC), Premature Ventricular Contraction (PVC),

The accuracy rate can be further improved by ST-segment depression (STD) and ST-segment elevation (STE).

While the invention has been described in further detail with reference to specific preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. The multi-label classification method of the 12-lead electrocardiosignals based on the neural network is characterized by comprising the following specific steps of:

1) determining a data input format: the number of lead connections is 12, the length of the lead connections is 3s of electrocardiosignals sampled according to 500HZ sampling, and a data format with 12 multiplied by 1500 is formed;

2) improving the convolutional neural network: the convolutional neural network was modified to have 3 independent streams, each consisting of, in turn: two convolutional layers with the size of 64, a first largest pooling layer, two convolutional layers with the size of 128, a second largest pooling layer, two convolutional layers with the size of 256, a third largest pooling layer, a fully-connected layer with the size of 768 and a fully-connected layer with the size of 512; the convolution kernel size of each stream is 3, 5, 7 respectively;

3) data input: inputting the electrocardiosignals into each stream in sequence, capturing scale characteristics of a kernel from the input electrocardiogram record, and then fusing the characteristic information of the 3 branches through the full-connection layer;

4) and (4) label classification: removing the Softmax layer of the improved convolutional neural network, inputting the extracted features into the RNN through full connection, and forming 9 classes of the features: normal, atrial fibrillation, low-grade atrioventricular block, left bundle branch block, right bundle branch block, premature atrial contraction, premature ventricular contraction, ST segment depression, ST segment elevation; the RNN is composed of a double-layer LSTM structure, the length of the two-way LSTM in the previous layer is 256, the length of the two-way LSTM in the next layer is 64, and the number of the hidden units in each layer is 128.

2. The neural network-based 12-lead cardiac signal multi-label classification method according to claim 1, characterized in that the step size of the maximum pooling layer of the first maximum pooling layer and the second maximum pooling layer is 3.

3. The neural network-based 12-lead cardiac signal multi-label classification method according to claim 1, characterized in that the step size in the third maximum pooling layer is 2.

4. The neural network-based 12-lead electrocardiosignal multi-label classification method according to claim 1, wherein the improved convolutional neural network uses a cross entropy method to calculate a loss function, and uses L2 regularization to improve generalization performance.