Automatic electrocardiosignal classification method and system based on feature adaptive weighted fusion

Abstract

The present application relates to the technical field of automatic classification, and discloses a kind of electrocardiosignal automatic classification method and system based on feature adaptive weighting fusion, comprising the following steps: S1, obtain the electrocardiosignal of twelve leads to be classified and the electrocardiosignal of twelve leads for testing, carry out feature extraction, obtain its respective data time series, and further obtain feature data;S2, extract the kinetic feature of feature data;The kinetic difference between the kinetic feature to be classified and the kinetic feature obtained by calculation;S3, the feature extraction of feature data to be classified is carried out again, and depth feature is obtained;S4, the weight of feature data to be classified in classification identification is calculated, and the depth feature is adaptively weighted feature fusion according to the weight of feature data to be classified in classification identification, and the classification identification result of the electrocardiosignal of twelve leads to be classified is output through full connection network.The present application solves the problem that different depth features cannot be adaptively weighted and fused in prior art, and has the characteristics of accurate classification identification.

Description

Technical Field

[0001] This invention relates to the field of automatic classification technology, and more specifically, to an automatic classification method and system for electrocardiogram signals based on feature adaptive weighted fusion. Background Technology

[0002] Electrocardiogram (ECG) signals are common time-series signals in the field of intelligent medical signals and are one of the most frequently used clinical diagnostic tools. Traditional ECG signal analysis mainly relies on the experience of clinicians and manual interpretation, lacking efficient and stable methods for automatic analysis, classification, and diagnosis of ECG signals. With the application and popularization of machine learning methods, more and more researchers are beginning to study automatic ECG signal classification methods based on machine learning. These methods often rely on large amounts of data, complex feature extraction, and parameter learning processes, which poses certain challenges to their application in the biomedical field.

[0003] An existing technology provides a C-LSTM-based arrhythmia classification algorithm for monitoring physiological parameters in the Internet of Things (IoT) for health, comprising the following steps: an ECG signal acquisition device collects ECG data and performs data preprocessing; the ECG signal acquisition device transmits the collected data to a server, where a deep convolutional neural network (CNN) performs deep encoding on the collected ECG signals to extract morphological features; by learning the temporal correlation between morphological feature representations through LSTM, the intrinsic features are further explored; using fully connected layers and the Softmax function, the predicted probability is calculated based on the output of the LSTM unit to achieve automatic arrhythmia classification based on ECG signal features.

[0004] However, existing technologies still have the problem of being unable to adaptively weight and fuse different depth features. How to invent a method that can adaptively weight and fuse different depth features is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the problem that existing technologies cannot adaptively weighted fuse different depth features, this invention provides an automatic classification method and system for electrocardiogram signals based on feature adaptive weighted fusion, which features accurate classification and recognition.

[0006] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0007] An automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion includes the following steps:

[0008] S1. Preprocess the twelve-lead ECG signal to be classified and tested; extract features from the data of each lead of the preprocessed ECG signal to be classified and tested to obtain their respective data time series, and further obtain the feature data to be classified and the test feature data.

[0009] S2. Construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; calculate the dynamic difference between the obtained dynamic features to be classified and the test dynamic features;

[0010] S3. Introduce a convolutional neural network; use a convolutional neural network to extract features from the data to be classified again to obtain deep features;

[0011] S4. Introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences, and perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition. Output the classification and recognition results of the twelve-lead ECG signals to be classified through the fully connected network.

[0012] Preferably, in step S1, feature extraction is performed on the data of each lead of the preprocessed ECG signal to be classified and used for testing to obtain their respective data time series, specifically as follows:

[0013] S11. Analyze the waveforms of the preprocessed ECG signals to be classified and tested, and extract the start point, peak, and end point of the QRS wave and T wave in each cardiac cycle, thereby obtaining the start point, peak, and end point of the QRS wave and T wave for all cardiac cycles; denot the extracted start point, peak, and end point of the i-th QRS wave as Rs. i Rp i Re i The starting point, peak, and ending point of the extracted i-th T-wave are denoted as Ts. i Tp i Te i This yields the ECG data sequence to be classified and tested;

[0014] S12. Based on the ECG data sequence to be classified and tested, obtain the time series of whole heart beat amplitude characteristic data, the time series of whole heart beat width characteristic data, and the time series of whole heart beat area characteristic data.

[0015] Furthermore, in step S11, the waveforms of the preprocessed electrocardiogram signals to be classified and tested are analyzed, specifically as follows:

[0016] S101. Calculate the amplitude difference between corresponding data points of adjacent heartbeats in the electrocardiogram data sequence as amplitude feature data, thereby obtaining the time series of whole heartbeat amplitude feature data:

[0017] L Rsi =abs(y(Rs) i )-y(Rs i+1 ))

[0018] L Rpi =abs(y(Rp) i )-y(Rp i+1 ))

[0019] L Rei =abs(y(Re i )-y(Re i+1 ))

[0020] L Tsi =abs(y(Ts) i )-y(Ts i+1 ))

[0021] L Tpi =abs(y(Tp) i )-y(Tp i+1 ))

[0022] L Tei =abs(y(Te i )-y(Te i+1 ))

[0023] Among them, L Rsi L represents the amplitude difference between the start points of two adjacent QRS waves in spatial coordinates. Rpi L represents the amplitude difference between the peaks of two adjacent QRS waves in spatial coordinates; Rei L represents the amplitude difference between the endpoints of two adjacent QRS waves in spatial coordinates. Tsi L represents the amplitude difference between the starting points of two adjacent T waves in spatial coordinates. Tpi L represents the amplitude difference between the apexes of two adjacent T-wave peaks in spatial coordinates; Tei This represents the amplitude difference between two adjacent T-wave termination points in spatial coordinates; x(.) represents the x-coordinate of the data point in the ECG data sequence in a two-dimensional coordinate system; y(·) represents the y-coordinate of the data point in the ECG data sequence in a two-dimensional coordinate system; abs(.) represents the absolute value operation;

[0024] S102. Calculate the time interval between corresponding data points for each heart beat in the ECG data sequence as width feature data, thereby obtaining the time series of full heart beat width feature data:

[0025] WRpi =abs(x(Rp) i )-x(Rp i+1 ))

[0026] W Tpi =abs(x(Tp) i )-x(Tp i+1 ))

[0027] W RTi =abs(x(Rs) i )-x(Te i ))

[0028] W Ti =abs(x(Tp) i )-x(Te i ))

[0029] W Ri =abs(x(Rs) i )-x(Re i ))

[0030] W Tdi =abs(x(Ts) i )-x(Tp i ))

[0031] Among them, W Rpi W represents the time interval between two adjacent QRS wave peaks on the time coordinate. Tpi W represents the time interval between two adjacent T-wave peaks on the time coordinate. RTi W represents the time interval between the start of the QRS complex and the end of the T-wave on the time axis. Ti This represents the time interval between the peak and end of the T-wave on the time axis; W Ri W represents the time interval between the start and end points of the QRS wave on the time axis. Tdi y(·) represents the time interval between the start point and peak of the T wave on the time axis; x(·) represents the horizontal coordinate of the data point in the ECG data sequence in the two-dimensional coordinate system; y(·) represents the vertical coordinate of the data point in the ECG data sequence in the two-dimensional coordinate system; abs(·) represents the absolute value operation;

[0032] S103. Calculate the area enclosed by the ECG data for each heartbeat and the ECG baseline in the ECG data sequence as the area feature data, where the area above the baseline is the positive area and the area below the baseline is the negative area, thereby obtaining the time series of the whole heartbeat area feature data:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] Among them, A Rsi A represents the area enclosed by the curve between the start point and the peak of the QRS wave; Rei A represents the area enclosed by the curve between the peak and the end of the QRS wave; Ri A represents the area enclosed by the curve between the start and end points of the QRS wave; Tsi A represents the area enclosed by the curve between the start point and peak of the T-wave; Tei A represents the area enclosed by the curve between the peak and the end of the T-wave; Ti x(·) represents the area enclosed by the curve between the start and end points of the T wave; x(·) represents the abscissa value of the data point in the ECG data sequence in the two-dimensional coordinate system; y(·) represents the ordinate value of the data point in the ECG data sequence in the two-dimensional coordinate system.

[0040] S104. Traverse all ECG data sequences for all cardiac cycles to obtain feature data.

[0041] Furthermore, the feature data consists of several feature curves that characterize the time-varying power energy variability of electrocardiogram signals.

[0042] Furthermore, in step S2, the intrinsic dynamic features of the feature data are extracted through a classification neural network, specifically as follows:

[0043]

[0044] Wherein, vector X() represents the value of the dynamic feature point to be classified on the feature curve at time t. Let f(X(t)) represent the derivative of X() at time t, and f(X(t)) is the nonlinear dynamic term of X() as it evolves over time, representing the nonlinear unknown dynamic of the characteristic curve.

[0045] Furthermore, in step S2, the calculated dynamic difference between the dynamic features to be classified and the test dynamic features is specifically as follows:

[0046] S201. Calculate the weight matrix of the classification neural network for the dynamic features to be classified and the dynamic features to be tested:

[0047] ξ=argmin ξ ||-X||2

[0048] Here, Θ() is a candidate function library composed of wavelet neural networks under multiple different scale factors, used to approximate the nonlinear dynamic term f(X(t)):

[0049]

[0050] ξ is the coefficient of the candidate function Θ(X), and the goal is to find the minimum value. It represents a combination of multiple wavelet neural networks of different scales that can approximate the nonlinear dynamic term. The symbols ||·||2 represent the L-2 norm operation respectively. Each ξ converges to the optimal constant value over a period of time, thus obtaining the weight matrix of the classification neural network for the dynamic features to be classified.

[0051] S202. Calculate the matrix difference between the weight matrix of the classification neural network for the test dynamic features and the weight matrix of the classification neural network for the dynamic features to be classified.

[0052] Furthermore, in S202, when calculating the matrix difference between the classification neural network weight matrix of the test dynamic features and the classification neural network weight matrix of the dynamic features to be classified, the root mean square value of all elements in the obtained difference matrix is ​​used as the recognition error term; for amplitude features, width features and area features, the dynamic difference between the test dynamic features and the dynamic features to be classified is calculated respectively, and the corresponding recognition error terms ε1, ε2 and ε3 are obtained.

[0053] Furthermore, in step S3, a convolutional neural network is used to extract features from the feature dataset again, specifically as follows:

[0054] The structure of a convolutional neural network is set up, and the convolutional neural network is used to perform secondary feature extraction on the feature dataset. After the convolutional neural network operation, the depth feature matrices D1, D2, and D3 corresponding to the amplitude feature, width feature, and area feature are obtained. The depth feature matrices corresponding to the amplitude feature, width feature, and area feature are denoted as D1. test D2 test D3 test .

[0055] Furthermore, in step S4, the weights of the features in the data to be classified are calculated based on the dynamic differences, and different depth features are adaptively weighted and fused. The classification results of the twelve-lead ECG signals to be classified are then output through a fully connected network. The specific steps are as follows:

[0056] S401. Calculate the weight coefficients of the features in the classification and recognition of the feature data to be classified based on the dynamic differences:

[0057]

[0058]

[0059]

[0060] Where, ε max ε min ε1, ε2, and ε3 represent the maximum and minimum values, respectively; ω1, ω2, and ω3 represent the weight coefficients of amplitude feature, width feature, and area feature in the final classification and recognition task, respectively.

[0061] S402. Multiply the depth features by the corresponding weight coefficients:

[0062]

[0063] S403, D1 after fusion multiplied by weighting coefficient test D2 test D3 test This data is then used as input to a fully connected network for classification, resulting in the final classification and recognition outcome.

[0064] An automatic classification system for electrocardiogram signals based on feature adaptive weighted fusion includes a data extraction module, a classification neural module, a convolutional neural network, and a fully connected network fusion module;

[0065] The data extraction module is used to preprocess the twelve-lead electrocardiogram (ECG) signals to be classified and tested; to extract features from the data of each lead of the preprocessed ECG signals to be classified and tested, to obtain their respective data time series, and to further obtain the feature data to be classified and the test feature data.

[0066] The classification neural module is used to construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; and calculate the dynamic difference between the obtained dynamic features to be classified and the test dynamic features.

[0067] The convolutional neural network mentioned above is used to introduce the convolutional neural network; the convolutional neural network is used to extract features from the data to be classified again to obtain deep features;

[0068] The fully connected network fusion module is used to introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences; perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition; and output the classification and recognition results of the twelve-lead electrocardiogram signals to be classified through the fully connected network.

[0069] The beneficial effects of this invention are as follows:

[0070] This invention discloses an automatic classification method for electrocardiogram (ECG) signals based on adaptive weighted feature fusion. For the robot path planning problem, this invention proposes an improved chaotic adaptive particle swarm optimization method. After extracting the temporal feature data of the original ECG signals, a neural network is constructed to extract intrinsic dynamic features. Based on the dynamic differences between the obtained dynamic features to be classified and the test dynamic features, the weights of the features in the classification data are determined. A convolutional neural network is used to extract features from the dynamic feature dataset again, and features of different depths are adaptively weighted and fused. Finally, a fully connected network outputs the classification result. The method proposed in this invention can effectively avoid getting trapped in local optima, has strong search capabilities and high convergence speed, and can effectively solve the robot path planning problem. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating an automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to the present invention.

[0072] Figure 2 This is a schematic diagram of the specific process of an automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to the present invention in Example 2.

[0073] Figure 3 This is a schematic diagram of a 12-lead electrocardiogram of the human body in Example 2. Detailed Implementation

[0074] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0075] Example 1

[0076] like Figure 1 As shown, an automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion includes the following steps:

[0077] S1. Preprocess the twelve-lead ECG signal to be classified and tested; extract features from the data of each lead of the preprocessed ECG signal to be classified and tested to obtain their respective data time series, and further obtain the feature data to be classified and the test feature data.

[0078] S2. Construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; calculate the dynamic difference between the obtained dynamic features to be classified and the test dynamic features;

[0079] S3. Introduce a convolutional neural network; use a convolutional neural network to extract features from the data to be classified again to obtain deep features;

[0080] S4. Introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences, and perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition. Output the classification and recognition results of the twelve-lead ECG signals to be classified through the fully connected network.

[0081] Example 2

[0082] like Figure 2 As shown, an automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion includes the following steps:

[0083] S1. Preprocess the twelve-lead ECG signal to be classified and tested; extract features from the data of each lead of the preprocessed ECG signal to be classified and tested to obtain their respective data time series, and further obtain the feature data to be classified and the test feature data.

[0084] S2. Construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; calculate the dynamic difference between the obtained dynamic features to be classified and the test dynamic features;

[0085] S3. Introduce a convolutional neural network; use a convolutional neural network to extract features from the data to be classified again to obtain deep features;

[0086] S4. Introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences, and perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition. Output the classification and recognition results of the twelve-lead ECG signals to be classified through the fully connected network.

[0087] The electrocardiogram (ECG) signals used in this embodiment are from the ECG database provided by the German National Metrology Institute, Physikalisch-Technische Bundesanstalt (PTB), such as... Figure 3As shown. The recorded ECG database contains ECG records from 290 subjects, each containing conventional 12 leads (i, ii, iii, apr, avl, avf, v1, v2, v3, v4, v5, v6). This invention treats the sampled continuous ECG signal time series as a data queue of fixed length N, and then extracts the R-wave and T-wave from each cardiac cycle after averaging and median filtering to obtain the R-wave and T-wave for all cardiac cycles. The preprocessed twelve-lead ECG signal waveform is analyzed to extract the start point, peak, and end point of the QRS wave and T-wave for each cardiac cycle, thus obtaining the start point, peak, and end point of the QRS wave and T-wave for all cardiac cycles. The extracted start point, peak, and end point of the i-th QRS wave are denoted as Rs. i Rp i Re i The starting point, peak, and ending point of the extracted i-th T-wave are denoted as Ts. i Tp i Te i .

[0088] In one specific embodiment, step S1 involves extracting features from the preprocessed ECG signals for classification and testing for each lead to obtain their respective data time series, specifically:

[0089] S11. Analyze the waveforms of the preprocessed ECG signals to be classified and tested, and extract the start point, peak, and end point of the QRS wave and T wave in each cardiac cycle, thereby obtaining the start point, peak, and end point of the QRS wave and T wave for all cardiac cycles; denot the extracted start point, peak, and end point of the i-th QRS wave as Rs. i Rp i Re i The starting point, peak, and ending point of the extracted i-th T-wave are denoted as Ts. i Tp i Te i This yields the ECG data sequence to be classified and tested;

[0090] S12. Based on the ECG data sequence to be classified and tested, obtain the time series of whole heart beat amplitude characteristic data, the time series of whole heart beat width characteristic data, and the time series of whole heart beat area characteristic data.

[0091] In one specific embodiment, step S11 involves analyzing the waveforms of the preprocessed electrocardiogram signals to be classified and tested, specifically as follows:

[0092] S101. Calculate the amplitude difference between corresponding data points of adjacent heartbeats in the electrocardiogram data sequence as amplitude feature data, thereby obtaining the time series of whole heartbeat amplitude feature data:

[0093] L Rsi =abs(y(Rs) i )-y(Rs i+1 ))

[0094] L Rpi =abs(y(Rp) i )-y(Rp i+1 ))

[0095] L Rei =abs(y(Re i )-y(Re i+1 ))

[0096] L Tsi =abs(y(Ts) i )-y(Ts i+1 ))

[0097] L Tpi =abs(y(Tp) i )-y(Tp i+1 ))

[0098] L Tei =abs(y(Te i )-y(Te i+1 ))

[0099] Among them, L Rsi L represents the amplitude difference between the start points of two adjacent QRS waves in spatial coordinates. Rpi L represents the amplitude difference between the peaks of two adjacent QRS waves in spatial coordinates; Rei L represents the amplitude difference between the endpoints of two adjacent QRS waves in spatial coordinates. Tsi L represents the amplitude difference between the starting points of two adjacent T waves in spatial coordinates. Tpi L represents the amplitude difference between the apexes of two adjacent T-wave peaks in spatial coordinates; Tei This represents the amplitude difference between two adjacent T-wave termination points in spatial coordinates; x(·) represents the abscissa value of the data point in the ECG data sequence in a two-dimensional coordinate system; y(·) represents the ordinate value of the data point in the ECG data sequence in a two-dimensional coordinate system; abs(·) represents the absolute value operation;

[0100] S102. Calculate the time interval between corresponding data points for each heart beat in the ECG data sequence as width feature data, thereby obtaining the time series of full heart beat width feature data:

[0101] WRpi =abs(x(Rp) i )-x(Rp i+1 ))

[0102] W Tpi =abs(x(Tp) i )-x(Tp i+1 ))

[0103] W RTi =abs(x(Rs) i )-x(Te i ))

[0104] W Ti =abs(x(Tp) i )-x(Te i ))

[0105] W Ri =abs(x(Rs) i )-x(Re i ))

[0106] W Tdi =abs(x(Ts) i )-x(Tp i ))

[0107] Among them, W Rpi W represents the time interval between two adjacent QRS wave peaks on the time coordinate. Tpi W represents the time interval between two adjacent T-wave peaks on the time coordinate. RTi W represents the time interval between the start of the QRS complex and the end of the T-wave on the time axis. Ti This represents the time interval between the peak and end of the T-wave on the time axis; W Ri W represents the time interval between the start and end points of the QRS wave on the time axis. Tdi y(·) represents the time interval between the start point and peak of the T wave on the time axis; x(·) represents the horizontal coordinate of the data point in the ECG data sequence in the two-dimensional coordinate system; y(·) represents the vertical coordinate of the data point in the ECG data sequence in the two-dimensional coordinate system; abs(·) represents the absolute value operation;

[0108] S103. Calculate the area enclosed by the ECG data for each heartbeat and the ECG baseline in the ECG data sequence as the area feature data, where the area above the baseline is the positive area and the area below the baseline is the negative area, thereby obtaining the time series of the whole heartbeat area feature data:

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115] Among them, A Rsi A represents the area enclosed by the curve between the start point and the peak of the QRS wave; Rei A represents the area enclosed by the curve between the peak and the end of the QRS wave; Ri A represents the area enclosed by the curve between the start and end points of the QRS wave; Tsi A represents the area enclosed by the curve between the start point and peak of the T-wave; Tei A represents the area enclosed by the curve between the peak and the end of the T-wave; Ti x(·) represents the area enclosed by the curve between the start and end points of the T wave; x(·) represents the abscissa value of the data point in the ECG data sequence in the two-dimensional coordinate system; y(·) represents the ordinate value of the data point in the ECG data sequence in the two-dimensional coordinate system.

[0116] S104. Traverse all ECG data sequences for all cardiac cycles to obtain feature data.

[0117] In this embodiment, the feature data are six feature curves that characterize the power energy variation of electrocardiogram signals over time.

[0118] In one specific embodiment, step S2 involves extracting the intrinsic dynamic features of the feature data using a classification neural network, specifically as follows:

[0119]

[0120] Wherein, vector X(t) represents the value of the dynamic feature point to be classified on the feature curve at time t. Let T(X(t)) represent the derivative of X(t) at time t, and let T(X(t)) be the nonlinear dynamic term of X(t) as it evolves over time, representing the nonlinear unknown dynamic of the characteristic curve.

[0121] In one specific embodiment, the dynamic difference between the dynamic features to be classified and the test dynamic features calculated in step S2 is as follows:

[0122] S201. Calculate the weight matrix of the classification neural network for the dynamic features to be classified and the dynamic features to be tested:

[0123] ξ=arg min ξ ||Θξ-X||2

[0124] Here, Θ(X) is a candidate function library composed of wavelet neural networks with multiple different scale factors, used to approximate the nonlinear dynamic term f(X(t)):

[0125]

[0126] ξ is the coefficient of the candidate function Θ(X), and the goal is to find the minimum value. It represents a combination of multiple wavelet neural networks of different scales that can approximate the nonlinear dynamic term. The symbols ||·||2 represent the L-2 norm operation respectively. Each ξ converges to the optimal constant value over a period of time, thus obtaining the weight matrix of the classification neural network for the dynamic features to be classified.

[0127] S202. Calculate the matrix difference between the weight matrix of the classification neural network for the test dynamic features and the weight matrix of the classification neural network for the dynamic features to be classified.

[0128] In a specific embodiment, in S202, when calculating the matrix difference between the classification neural network weight matrix of the test dynamic feature and the classification neural network weight matrix of the dynamic feature to be classified, the root mean square value of all elements in the obtained difference matrix is ​​used as the recognition error term; for amplitude feature, width feature and area feature, the dynamic difference between the test dynamic feature and the dynamic feature to be classified is calculated respectively, and the corresponding recognition error terms ε1, ε2 and ε3 are obtained.

[0129] In one specific embodiment, step S3 involves using a convolutional neural network to extract features from the feature dataset again, specifically as follows:

[0130] The structure of a convolutional neural network is set up, and the convolutional neural network is used to perform secondary feature extraction on the feature dataset. After the convolutional neural network operation, the depth feature matrices D1, D2, and D3 corresponding to the amplitude feature, width feature, and area feature are obtained. The depth feature matrices corresponding to the amplitude feature, width feature, and area feature are denoted as D1. test D2 test D3 test .

[0131] In one specific embodiment, in step S4, the weights of the features in the data to be classified are calculated based on the dynamic differences, and different depth features are adaptively weighted and fused. The classification and recognition results of the twelve-lead ECG signals to be classified are then output through a fully connected network. The specific steps are as follows:

[0132] S401. Calculate the weight coefficients of the features in the classification and recognition of the feature data to be classified based on the dynamic differences:

[0133]

[0134]

[0135]

[0136] Where, ε max ε min ε1, ε2, and ε3 represent the maximum and minimum values, respectively; ω1, ω2, and ω3 represent the weight coefficients of amplitude feature, width feature, and area feature in the final classification and recognition task, respectively.

[0137] S402. Multiply the depth features by the corresponding weight coefficients:

[0138]

[0139] S403, D1 after fusion multiplied by weighting coefficient test D2 test D3 test This data is then used as input to a fully connected network for classification, resulting in the final classification and recognition outcome.

[0140] Example 3

[0141] An automatic classification system for electrocardiogram signals based on feature adaptive weighted fusion includes a data extraction module, a classification neural module, a convolutional neural network, and a fully connected network fusion module;

[0142] The data extraction module is used to preprocess the twelve-lead electrocardiogram (ECG) signals to be classified and tested; to extract features from the data of each lead of the preprocessed ECG signals to be classified and tested, to obtain their respective data time series, and to further obtain the feature data to be classified and the test feature data.

[0143] The classification neural module is used to construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; and calculate the dynamic difference between the obtained dynamic features to be classified and the test dynamic features.

[0144] The convolutional neural network mentioned above is used to introduce the convolutional neural network; the convolutional neural network is used to extract features from the data to be classified again to obtain deep features;

[0145] The fully connected network fusion module is used to introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences; perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition; and output the classification and recognition results of the twelve-lead electrocardiogram signals to be classified through the fully connected network.

[0146] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. An automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion, characterized in that: Includes the following steps: S1. Preprocess the twelve-lead ECG signal to be classified and tested; extract features from the data of each lead of the preprocessed ECG signal to be classified and tested to obtain their respective data time series, and further obtain the feature data to be classified and the feature data to be tested. In step S1, during feature extraction, the waveforms of the preprocessed ECG signals to be classified and used for testing are analyzed. The specific steps are as follows: The amplitude difference between corresponding data points of adjacent heartbeats in the electrocardiogram data sequence is calculated as amplitude feature data, thereby obtaining the time series of whole heartbeat amplitude feature data; The time interval between corresponding data points for each heart beat in the electrocardiogram data sequence is calculated as the width feature data, thereby obtaining the time series of the whole heart beat width feature data; The area enclosed between each heartbeat ECG data point and the ECG baseline is calculated as the area feature data, where the area above the baseline is the positive area and the area below the baseline is the negative area, thus obtaining the time series of the whole heartbeat area feature data; S2. Construct a classification neural network; extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network; The dynamic difference between the calculated dynamic features to be classified and the tested dynamic features; Specifically, the intrinsic dynamic features of the feature data are extracted through a classification neural network, namely: Where, vector This represents the value of the dynamic feature point to be classified on the characteristic curve at time t. Represents time t The derivative of yes The nonlinear dynamic term that evolves over time represents the nonlinear unknown dynamic of the characteristic curve; The calculated dynamic differences between the dynamic features to be classified and the test dynamic features are as follows: S201. Calculate the weight matrix of the classification neural network for the dynamic features to be classified and the dynamic features to be tested: in, It is a candidate function library composed of wavelet neural networks under multiple different scale factors, used to approximate nonlinear dynamic terms. : It is a candidate function The coefficients, with the goal of finding their minimum, represent a combination of wavelet neural networks at different scales that can approximate the nonlinear dynamic term; (symbol) These represent L-2 norm operations; each The weight matrix of the classification neural network for the dynamic features to be classified is obtained by converging to the optimal constant value over a period of time. S202. Calculate the matrix difference between the weight matrix of the classification neural network for the test dynamic features and the weight matrix of the classification neural network for the dynamic features to be classified; S3. Introduce a convolutional neural network; use a convolutional neural network to extract features from the data to be classified again to obtain deep features; S4. Introduce a fully connected network; calculate the weights of the features in the classification data based on the dynamic differences, and perform adaptive weighted feature fusion of the deep features according to their weights. Output the classification results of the twelve-lead ECG signals through the fully connected network. The specific steps are as follows: S401. Calculate the weight coefficients of the features in the classification and recognition of the feature data to be classified based on the dynamic differences: in, , Represent The maximum and minimum values ​​in; These represent the weight coefficients of amplitude feature, width feature, and area feature in the final classification and recognition task, respectively. S402. Multiply the depth features by the corresponding weight coefficients: S403, after fusion multiplied by weighting coefficient , This data is then used as input to a fully connected network for classification, resulting in the final classification and recognition outcome.

2. The automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to claim 1, characterized in that: In step S1, feature extraction is performed on the data of each lead of the preprocessed ECG signal to be classified and used for testing to obtain their respective data time series, specifically as follows: S11. Analyze the waveforms of the preprocessed ECG signals to be classified and tested, extracting the start point, peak, and end point of the QRS wave and T wave in each cardiac cycle, thereby obtaining the start point, peak, and end point of the QRS wave and T wave for all cardiac cycles; denot the extracted start point, peak, and end point of the i-th QRS wave as... The starting point, peak, and ending point of the extracted i-th T-wave are denoted as . This yields the ECG data sequence to be classified and tested; S12. Based on the ECG data sequences to be classified and tested, obtain the time series of whole-heart beat amplitude characteristic data, the time series of whole-heart beat width characteristic data, and the time series of whole-heart beat area characteristic data to be classified and tested.

3. The automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to claim 2, characterized in that: In step S11, the waveforms of the preprocessed electrocardiogram signals to be classified and tested are analyzed, specifically as follows: S101. Calculate the amplitude difference between corresponding data points of adjacent heartbeats in the electrocardiogram data sequence as amplitude feature data, thereby obtaining the time series of whole heartbeat amplitude feature data: in, This represents the amplitude difference between the start points of two adjacent QRS waves in spatial coordinates; This represents the amplitude difference between the peaks of two adjacent QRS waves in spatial coordinates. This represents the amplitude difference between the endpoints of two adjacent QRS waves in spatial coordinates. This represents the amplitude difference between the start points of two adjacent T waves in spatial coordinates; This represents the amplitude difference between the apex values ​​of two adjacent T-wave peaks in spatial coordinates; This represents the amplitude difference between the termination points of two adjacent T waves in spatial coordinates; This represents the x-coordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. This represents the ordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. This indicates the absolute value operation; S102. Calculate the time interval between corresponding data points for each heart beat in the ECG data sequence as the width feature data, thereby obtaining the time series of the whole heart beat width feature data: in, This represents the time interval between two adjacent QRS wave peaks on the time coordinate. This represents the time interval between two adjacent T-wave peaks on the time coordinate. This represents the time interval between the start point of the QRS wave and the end point of the T wave on the time axis. This represents the time interval between the peak of the T-wave and the end of the T-wave on the time coordinate. This represents the time interval between the start and end points of the QRS wave on the time coordinate. This represents the time interval between the start point and peak point of the T-wave on the time coordinate. This represents the x-coordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. This represents the ordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. This indicates the absolute value operation; S103. Calculate the area enclosed by the ECG data for each heartbeat and the ECG baseline in the ECG data sequence as the area feature data, where the area above the baseline is the positive area and the area below the baseline is the negative area, thereby obtaining the time series of the whole heartbeat area feature data: in, This represents the area enclosed by the curve between the start point and the peak of the QRS wave; This represents the area enclosed by the curve between the peak and the end of the QRS wave; This represents the area enclosed by the curve between the start and end points of the QRS wave; This represents the area enclosed by the curve between the start point of the T-wave and the peak point of the T-wave; This represents the area enclosed by the curve between the peak of the T-wave and the end of the T-wave. This represents the area enclosed by the curve between the start and end points of the T-wave; This represents the x-coordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. This represents the ordinate value of a data point in a two-dimensional coordinate system within an electrocardiogram (ECG) data sequence. S104. Traverse all ECG data sequences for all cardiac cycles to obtain feature data.

4. The automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to claim 1, characterized in that: The feature data consists of several feature curves that characterize the time-varying power energy variation parameters of electrocardiogram signals.

5. The automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to claim 1, characterized in that: In S202, when calculating the matrix difference between the weight matrix of the classification neural network for the test dynamic features and the weight matrix of the classification neural network for the dynamic features to be classified, the root mean square value of all elements in the resulting difference matrix is ​​used as the recognition error term. For amplitude features, width features, and area features, the dynamic differences between the test dynamic features and the dynamic features to be classified are calculated respectively, and the corresponding recognition error terms are obtained. .

6. The automatic classification method for electrocardiogram signals based on feature adaptive weighted fusion according to claim 1, characterized in that: In step S3, a convolutional neural network is used to extract features from the feature dataset again, specifically as follows: The structure of a convolutional neural network is set up, and the convolutional neural network is used to perform secondary feature extraction on the feature dataset; the depth feature matrix corresponding to amplitude feature, width feature, and area feature is obtained through convolutional neural network operation. , The depth feature matrix corresponding to the amplitude feature, width feature, and area feature is denoted as... , .

7. An automatic classification system for electrocardiogram signals based on feature adaptive weighted fusion, characterized in that: The method for implementing the method as described in any one of claims 1 to 6 includes a data extraction module, a classification neural module, a convolutional neural network, and a fully connected network fusion module; The data extraction module is used to preprocess the twelve-lead ECG signals to be classified and tested; to extract features from the data of each lead of the preprocessed ECG signals to obtain their respective data time series, and further obtain the feature data to be classified and the test feature data; the classification neural module is used to construct a classification neural network; and to extract the dynamic features of the feature data to be classified and the test feature data through the classification neural network. The dynamic difference between the calculated dynamic features to be classified and the tested dynamic features; The convolutional neural network is used to introduce the convolutional neural network; the convolutional neural network is used to extract features from the data to be classified again to obtain deep features; The fully connected network fusion module is used to introduce a fully connected network; calculate the weight of the features of the data to be classified in the classification and recognition based on the dynamic differences; perform adaptive weighted feature fusion of the deep features based on the weight of the features of the data to be classified in the classification and recognition; and output the classification and recognition results of the twelve-lead electrocardiogram signals to be classified through the fully connected network.

Images