Method for classifying various arrhythmia signals

Abstract

The invention discloses a method for classifying various arrhythmia signals, belongs to the technical field of arrhythmia signal classification methods, and the accurate classification of arrhythmia has important significance for disease diagnosis and clinical treatment. However, the traditional detection method is often interfered by various factors such as signal non-stationarity, feature non-specificity, artifacts and noise caused by arrhythmia when processing various arrhythmia, and the classification precision and efficiency are difficult to meet the requirements. Therefore, the invention provides an improved power normalized cepstrum coefficient (SPNCC) method based on singular spectrum analysis (SSA), which is used for extracting key features from a photoplethysmogram (PPG) signal and reducing the influence of noise, and constructing a deep learning classification model in combination with a convolutional neural network and a bidirectional long short-term memory network (CNN-BiLSTM). Therefore, accurate identification of various arrhythmias can be realized.

Description

Technical Field

[0001] The present invention relates to a method for classifying arrhythmia signals, in particular to a method for classifying multiple arrhythmia signals, and belongs to the technical field of arrhythmia signal classification methods. Background Art

[0002] Traditional arrhythmia diagnosis mainly relies on manual analysis by clinicians. This method highly depends on doctors' experience and professional knowledge, and has problems of strong subjectivity and low efficiency.

[0003] To improve the diagnosis efficiency, researchers have introduced automated signal processing technologies, such as fast Fourier transform, principal component analysis, and wavelet transform, to extract spectral features and main features.

[0004] On this basis, support vector machine and random forest machine learning algorithms have also been applied to classification tasks to enhance the automated recognition ability.

[0005] In recent years, with the development of deep learning technologies, convolutional neural networks, recurrent neural networks, and long short-term memory networks have begun to be applied to the classification of arrhythmia signals due to their superior performance in feature extraction, time series modeling, and capturing long-term dependencies respectively. Deep learning models can automatically learn features, reduce the dependence on manual feature extraction, and significantly improve the accuracy and efficiency of classification.

[0006] However, the non-specificity, non-stationarity of arrhythmia signals, and noise interference still pose challenges to existing methods. Therefore, developing more robust feature extraction methods and efficient classification methods has become the focus of current research. Summary of the Invention

[0007] The main object of the present invention is to provide a method for classifying multiple arrhythmia signals.

[0008] The object of the present invention can be achieved by adopting the following technical solutions:

[0009] A method for classifying multiple arrhythmia signals includes the following steps:

[0010] Step 1: Preprocess the arrhythmia signals by using the SPNCC preprocessing method;

[0011] Step 2: Transmit the preprocessed data to the CNN-BILSTM structure model for data processing;

[0012] Step 3: Transmit the data processed in Step 2 to the multi-dimensional performance index evaluation standard model for classifying arrhythmia signals.

[0013] Preferably, in Step 1, the SPNCC preprocessing method includes the following steps:

[0014] Select an appropriate window length L to lag the original time series to obtain the trajectory matrix M 1 ;

[0015] Calculate the covariance matrix M of the covariance matrix C of the trajectory matrix 1

[0016]

[0017] M 1 = UΣV T ;

[0018] Perform eigenvalue decomposition on the covariance matrix C to obtain the eigen triple UΣV T , and sort the eigenvalues and corresponding eigenvectors by magnitude to obtain {λ 1 > λ 2 > … > λ L ≥ 0}{V 1 , V 2 , …, V n};

[0019] where i.e., the singular value and the eigenvector V i is the singular vector.

[0020] Reconstruct each singular value obtained from the decomposition separately to convert it back to the trajectory matrix M 2 and then use the anti-diagonal averaging method to convert the trajectory matrix into a sequence form to obtain n components;

[0021] Perform correlation analysis based on the obtained components to evaluate the consistency between each component and the original PPG record. The calculation formula for the coherence coefficient is as follows:

[0022]

[0023] where Sig represents the original PPG record, and select the required components according to a preset threshold (the initial threshold is 0.9, which can be further adjusted according to experiments);

[0024] imf i is the time series obtained by reconstructing the original sequence through the eigenvector and performing diagonal averaging, and select the required components according to a preset threshold;

[0025] cov represents the calculation of covariance, and Var represents the calculation of variance.

[0026] Finally, x appropriate components are obtained through the calculation of the coherence coefficient and screening using the corresponding threshold, and the x components are arranged into a matrix of size x×L;

[0027] Preferably, SPNCC feature extraction is performed on the feature matrix composed of components, which specifically includes the following steps:

[0028] Calculate the power spectrum of P i (ω) using the following formula:

[0029]

[0030] Input the power estimate value P i (ω) into the gamma-pass filter for filtering. The time-domain impulse response of the gamma-pass filter is:

[0031] G(t) = αt n-1 e -2πωt cos(2πf 0 t + φ), (t > 0);

[0032] where α is a scaling factor used to adjust the response intensity of the filter, t is the time-domain input variable of the filter, f 0 is the center frequency of the filter, which determines the main frequency response of the filter to the input signal, φ is the phase shift of the filter, also known as phase rotation, which controls the phase of the filter response, ω is the filter bandwidth, and n is the filter order;

[0033] After filtering the signal using the gamma-pass filter, the filtered power spectrum will be normalized. The normalization process uses the following formula:

[0034]

[0035] where POW n is the value obtained from the gamma-pass filter, μ[ω] is the average power, and the frequency ω g reflects the spectrum after the filtering process;

[0036] Finally, perform non-linear processing on the normalized power spectrum:

[0037]

[0038] where θ is the exponential factor, 0 < θ < 1;

[0039] The SPNCC feature coefficient is the obtained feature matrix;

[0040] After obtaining the SPNCC feature matrix through the above processing, combine it with the time-domain features and input it into the CNN-BILSTM-Attention structure model for training;

[0041] Preferably, the CNN-BILSTM-Attention structure model specifically includes an input layer: composed of a combination of multiple data augmentation modules, including randomly adding Gaussian noise, signal scaling and cropping, time translation, and signal inversion;

[0042] The CNN layer: includes multiple convolutional layers and pooling layers. The convolutional layer simultaneously learns multiple different convolutional kernels, and each convolutional kernel extracts features of the data from different aspects, selects important feature data, and obtains the output data OutPut1. Then, the pooling layer performs dimensionality reduction on the data, retains the most significant features, and removes invalid features to obtain the output data Output2;

[0043] The BILSTM layer: contains two independent LSTM layers, one for obtaining information in the forward time step and the other for obtaining information in the backward time step. These two LSTM layers are independent of each other, each having its own hidden state and cell state. The forward LSTM processes data starting from the beginning of the sequence, while the backward LSTM processes data starting from the end of the sequence;

[0044] The ATTENTION layer: contains multiple independent attention heads at different angles, maps the temporal features to different sub-representation spaces, and guides the model to learn the correlation information in different sub-representation spaces to obtain richer representations;

[0045] The output layer: consists of a Dropout layer and a full connect layer. Dropout reduces the model's dependence on specific neurons by randomly discarding some neurons, making the model more robust. Full connect outputs the final classification result and regression value.

[0046] Preferably, the specific training process and signal transmission process of the CNN-BILSTM-ATTENTION model are as follows:

[0047] Input the data used for training the CNN-BILSTM-ATTENTION model;

[0048] Reconstruct the input original data;

[0049] Initialize the CNN-BILSTM-ATTENTION model with the optimal hyperparameters obtained by Bayesian optimization;

[0050] The reconstructed signal sequentially passes through multiple convolutional layers and pooling layers for CNN calculation to obtain the CNN output OutPut-One;

[0051] OutPut-One passes through multiple hidden layers of the BILSTM in sequence to obtain the output OutPut-Two of the BILSTM;

[0052] OutPut-Two passes through a new convolutional layer and a pooling layer again for CNN calculation to obtain the CNN output OutPut-Three;

[0053] OutPut-Three undergoes attention operation through the Multi-head Attention layer to obtain the output OutPut-Four;

[0054] OutPut-Four finally undergoes the calculation of the fully connected layer to obtain the final result and calculate the predicted loss value;

[0055] Verify whether the finally obtained loss value triggers the early stopping mechanism. If so, save the trained model;

[0056] If not, return to step four and repeat the above process until the early stopping condition is met.

[0057] Preferably, in step three, extract the misjudgment situations of each category from the confusion matrix;

[0058] Calculate the misjudgment rate of each category according to the misjudged samples and the total number of category samples:

[0059]

[0060] Dynamically adjust the weights according to the misjudgment rate:

[0061]

[0062] Dynamic adjustment of weights based on task objectives:

[0063]

[0064] Among them, Base is the initial weight, reflecting the basic preference of the task objective for the index;

[0065] ErrorFactorAcc is the mean of the overall misjudgment rate, measuring the overall classification accuracy of the model;

[0066] ErrorFactorRecall is the sum of the proportions of samples of each category misjudged as other categories, emphasizing the misjudgment impact on recall;

[0067] ErrorFactorPrecision is the sum of the proportions of samples of each category misjudged as the target category, emphasizing the misjudgment impact on precision;

[0068] α is a dynamic adjustment coefficient used to control the influence degree of the misjudgment rate on the weight.

[0069] Preferably, calculate the harmonic mean of all evaluation methods, including the accuracy, recall rate, F1 score, and precision of the heart rate signal classification;

[0070]

[0071] Subsequently, calculate the time index (T n ), representing a metric proportional to the calculation time.

[0072]

[0073] Among them, the coefficient σ ∈ (0, 1] represents the time ratio, which is determined according to the time sensitivity of the task; the higher σ is, the stronger the time sensitivity of the task indicates;

[0074] σ is set to 0.2, and normalize(Time) refers to normalizing the calculation time of all methods;

[0075] MDPI = T n × H n ;

[0076] Multiply the time index by the harmonic mean to comprehensively evaluate the calculation time and accuracy of all methods.

[0077] The beneficial technical effects of the present invention:

[0078] The present invention provides a variety of arrhythmia signal classification methods, which are used to extract key features from photoplethysmogram (PPG) signals and reduce the influence of noise, and combine a convolutional neural network and a bidirectional long short-term memory network (CNN-BiLSTM) to construct a deep learning classification model to achieve accurate identification of various arrhythmias. Experimental results show that the classification accuracy of this method reaches 99.6% on the MIT-BIH and MIMIC-III datasets, and is superior to existing methods in multiple metrics (such as recall rate and F1 score). In addition, this paper also introduces a comprehensive evaluation index T-Mean, and this method achieves the best balance between classification performance and computational efficiency. Therefore, the SPNCC-CNN-BiLSTM method proposed in this paper not only effectively solves the non-stationarity and noise problems of PPG signals, but also significantly improves the classification accuracy and robustness of various arrhythmias, providing reliable support for intelligent healthcare and early disease screening. Brief Description of the Drawings

[0079] Figure 1 It is a model training and testing flowchart of a preferred embodiment of the multiple arrhythmia signal classification method according to the present invention;

[0080] Figure 2 It is a framework diagram of the CNN-BILSTM-ATTENTION model according to a preferred embodiment of the method for classifying multiple arrhythmia signals of the present invention;

[0081] Figure 3 It is a training process diagram of the CNN-BILSTM-ATTENTION model according to a preferred embodiment of the method for classifying multiple arrhythmia signals of the present invention;

[0082] Figure 4 It is a flowchart according to a preferred embodiment of the method for classifying multiple arrhythmia signals of the present invention. Detailed implementation manners

[0083] To make the technical solutions of the present invention clearer and more definite to those skilled in the art, the present invention will be further described in detail below with reference to the embodiments and the accompanying drawings, but the implementation manners of the present invention are not limited thereto.

[0084] It includes the following steps:

[0085] Step 1: Preprocess the arrhythmia signals by using the SPNCC preprocessing method;

[0086] Step 2: Transmit the data after preprocessing to the CNN-BILSTM structural model for data processing;

[0087] Step 3: Transmit the data processed in Step 2 to the multi-dimensional performance index evaluation standard model for classifying arrhythmia signals.

[0088] Preferably, in Step 1, the SPNCC preprocessing method includes the following steps:

[0089] Select an appropriate window length L to lag the original time series to obtain the trajectory matrix M 1 ;

[0090] Calculate the covariance matrix M of the covariance matrix C of the trajectory matrix 1

[0091]

[0092] M 1 = U∑V T ;

[0093] Perform eigenvalue decomposition on the covariance matrix C to obtain the eigen-triplet U∑V T , and sort the eigenvalues and the corresponding eigenvectors according to their magnitudes to obtain {λ 1 > λ 2 > … > λ L ≥ 0}{V 1 , V2 , …, V n}; where i.e., the singular value and the eigenvector V i is the singular vector.

[0094] The singular values are divided into x non - connected groups, representing different trends, periodic components, and x is the number of target components;

[0095] Reconstruct the x non - connected groups to obtain a matrix of size x×K;

[0096] After performing SSA decomposition on the PPG record, each unique component and the residual are obtained;

[0097] Perform correlation analysis to evaluate the consistency between each component and the original PPG record. The calculation formula for the coherence coefficient is as follows:

[0098]

[0099] where n represents the total number of components obtained;

[0100] Sig represents the original PPG record, and the required components are selected according to a pre - set threshold;

[0101] imf i is the time series obtained by eigenvector reconstruction and after diagonal averaging, and the required components are selected according to a pre - set threshold.

[0102] cov represents the calculation of covariance, and Var represents the calculation of variance.

[0103] After performing SSA processing on the PPG record, these components are organized and represented by the eigenmatrix composed of these components.

[0104] Preferably, perform SPNCC feature extraction on the eigenmatrix composed of components, specifically including the following steps:

[0105] Use the following formula to calculate the power spectrum of P i (ω):

[0106]

[0107] Input the power estimate P i (ω) into the gamma - pass filter for filtering. The time - domain impulse response of the gamma - pass filter is:

[0108]

[0109] where ω is the filter bandwidth and n is the filter order;

[0110] After filtering, the filter is normalized to the power spectrum using the following expression:

[0111]

[0112] where POW n is the value obtained by the gamma-pass filter, and μ[ω] is the average power;

[0113] Nonlinear processing: Perform power function nonlinear processing using the expression:

[0114]

[0115] where θ is the exponential factor, 0 < θ < 1;

[0116] The SPNCC feature coefficient is the obtained feature matrix;

[0117] After obtaining the SPNCC feature matrix through the above processing, combine it with the time-domain features and input it into the CNN-BILSTM-Attention structure model for training;

[0118] Preferably, the CNN-BILSTM-Attention structure model specifically includes an input layer: composed of a combination of multiple data augmentation modules, including randomly adding Gaussian noise, signal scaling and cropping, time translation, and signal inversion;

[0119] The CNN layer: includes multiple convolutional layers and pooling layers. The convolutional layer simultaneously learns multiple different convolutional kernels, each convolutional kernel extracts features of different aspects from the data, selects important feature data, and obtains the output data OutPut1. Then, through the pooling layer, the data is dimensionally reduced, the most significant features are retained, and the invalid features are removed to obtain the output data Output2;

[0120] The BILSTM layer: contains two independent LSTM layers, one for obtaining information in the forward time step and the other for obtaining information in the backward time step. These two LSTM layers are independent of each other, each having its own hidden state and cell state. The forward LSTM processes the data starting from the beginning of the sequence, while the backward LSTM processes the data starting from the end of the sequence;

[0121] The ATTENTION layer: contains multiple independent attention heads at different angles, maps the temporal features to different sub-representation spaces, and guides the model to learn the correlation information in different sub-representation spaces to obtain richer representations;

[0122] Output layer: It consists of a Dropout layer and a full connect layer. Dropout randomly discards some neurons, reducing the model's dependence on specific neurons, thus making the model more robust. The full connect outputs the final classification result and regression value.

[0123] Preferably, the specific training process and signal transmission process of the CNN-BILSTM-ATTENTION model are as follows:

[0124] Input the data used for training the CNN-BILSTM-ATTENTION model;

[0125] Reconstruct the input original data;

[0126] Initialize the CNN-BILSTM-ATTENTION model with the optimal hyperparameters obtained by Bayesian optimization;

[0127] The reconstructed signal passes through multiple convolutional layers and pooling layers in sequence for CNN calculation to obtain the CNN output OutPut-One;

[0128] OutPut-One passes through multiple hidden layers of BILSTM in sequence to obtain the BILSTM output OutPut-Two;

[0129] OutPut-Two passes through a new convolutional layer and pooling layer again for CNN calculation to obtain the CNN output OutPut-Three;

[0130] OutPut-Three undergoes attention operation through the Multi-head Attention layer to obtain the output OutPut-Four;

[0131] OutPut-Four finally undergoes the calculation of the full connect layer to obtain the final result and calculate the predicted loss value;

[0132] Verify whether the finally obtained loss value triggers the early stopping mechanism. If so, save the trained model;

[0133] If not, return to step four and repeat the above process until the early stopping condition is met.

[0134] Preferably, in step three, extract the misjudgment situations of each category from the confusion matrix;

[0135] According to the misjudged samples and the total number of category samples, calculate the misjudgment rate of each category:

[0136]

[0137] Dynamically adjust the weight according to the misjudgment rate:

[0138]

[0139] Dynamic adjustment of weight based on task objectives:

[0140]

[0141] Among them, Base is the initial weight, reflecting the basic preference of the task objective for the indicator;

[0142] ErrorFactorAcc is the mean value of the overall misjudgment rate, measuring the overall classification accuracy of the model;

[0143] ErrorFactorRecall is the sum of the proportions of samples misjudged as other categories for each category, emphasizing the misjudgment impact on recall;

[0144] ErrorFactorPrecision is the sum of the proportions of samples misjudged as the target category for each category, emphasizing the misjudgment impact on precision;

[0145] α is the dynamic adjustment coefficient, used to control the influence degree of the misjudgment rate on the weight.

[0146] Preferably, calculate the harmonic mean of all evaluation methods, including the accuracy, recall, F1 score, and precision of heart rate signal classification;

[0147]

[0148] Subsequently, calculate the time index (T n ), representing a metric proportional to the calculation time.

[0149]

[0150] Among them, the coefficient σ ∈ (0, 1] represents the time ratio, determined according to the time sensitivity of the task; the higher σ is, the stronger the time sensitivity of the task is;

[0151] σ is set to 0.2, and normalize(Time) refers to normalizing the calculation time of all methods;

[0152] MDPI = T n ×H n ;

[0153] Multiply the time index by the harmonic mean to comprehensively evaluate the calculation time and accuracy of all methods.

[0154] As described above, it is only a further embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention, according to the technical solution and its concept of the present invention, makes equivalent substitutions or changes, all belong to the protection scope of the present invention.

Claims

1. A method for classifying multiple arrhythmia signals, characterized in that: The steps include: Step 1: Use SPNCC preprocessing method to denoise and extract features of arrhythmia signals; Step 2: Transfer the preprocessed data to the CNN-BILSTM structure model for data processing; Step 3: The data processed in step 2 is transmitted to the multidimensional performance index evaluation standard model for arrhythmia signal classification.

2. The method for classifying multiple arrhythmia signals according to claim 1, characterized in that: In step 1, the SPNCC preprocessing method is used to denoise the original time series, including the following steps: Select an appropriate window length L to lag the original time series and obtain the trajectory matrix M1; Calculate the covariance matrix M1 of the covariance matrix C of the trajectory matrix; M1=U∑V T ; Perform eigenvalue decomposition C on the covariance matrix to obtain the characteristic triple U∑V T , and sort the eigenvalues ​​and corresponding eigenvectors by size to obtain {λ1>λ2>…>λ L ≥0}{V1,V2,…,V n }; in Singular Values The eigenvector V i It is the singular vector; Each singular value obtained by decomposition is reconstructed separately and converted back to the trajectory matrix M2. The trajectory matrix is ​​converted into a sequence form using the anti-diagonal averaging method to obtain n components; Correlation analysis was performed based on the obtained components to evaluate the consistency between each component and the original PPG record. The coherence coefficient was calculated as follows: Where Sig represents the original PPG record, and the required components are selected according to a pre-set threshold (the initial threshold is 0.9, which can be further adjusted according to the experiment); imf i It is a time series obtained by reconstructing the original sequence through the eigenvector and diagonally averaging it, and selecting the required components according to the pre-set threshold; cov means the calculation of covariance, and Var means the calculation of variance; By calculating the coherence coefficient and using the corresponding threshold for screening, x suitable components are finally obtained, and the x components are arranged into a matrix of size x×L.

3. The SPNCC preprocessing method according to claim 2, wherein after performing a denoising operation on the original time series, a corresponding matrix is ​​obtained, and a feature extraction operation is performed on the matrix, characterized in that: The SPNCC feature extraction is performed on the feature matrix composed of components, which specifically includes the following steps: Calculate P using the following formula i The power spectrum of (ω): The power estimate P i (ω) is input into the gamma filter for filtering. The time domain impulse response of the gamma filter is: G(t)=αt n-1 e -2πωt cos(2πf0t+φ),(t>0); Among them, α is a scaling factor used to adjust the response strength of the filter, t is the time domain input variable of the filter, f0 is the center frequency of the filter, which determines the main frequency response of the filter to the input signal, φ is the phase offset of the filter, also known as phase rotation, which controls the phase of the filter response, ω is the filter bandwidth, and n is the filter order; After filtering the signal using the gammatone filter, the filtered power spectrum will be normalized. The normalization process uses the following formula: Among them, POW n is the value obtained by the gammatone filter, μ[ω] is the average power, and the frequency ω g Reflects the spectrum after the filtering process; Finally, the normalized power spectrum is processed nonlinearly: in, is the exponential factor, The SPNCC characteristic coefficient is the obtained characteristic matrix; After obtaining the SPNCC feature matrix through the above processing, it is input into the CNN-BILSTM-Attention structure model for training.

4. The method for classifying multiple arrhythmia signals according to claim 3, characterized in that: The CNN-BILSTM-Attention structure model specifically includes the input layer: composed of a combination of multiple data enhancement modules, including random addition of Gaussian noise, signal scaling and cropping, time shift, and signal inversion; CNN layer: includes multiple convolution layers and pooling layers. The convolution layer learns multiple different convolution kernels at the same time. Each convolution kernel extracts different features of the data, selects important feature data, and obtains output data OutPut1. The data is then processed by the pooling layer to reduce the dimension, retain the most significant features, remove invalid features, and obtain output data Output2. BILSTM layer: contains two independent LSTM layers, one for forward time step information acquisition, and the other for backward time step information acquisition. These two LSTM layers are independent of each other, each with its own hidden state and cell state. The forward LSTM processes data from the beginning of the sequence, while the backward LSTM processes data from the end of the sequence. ATTENTION layer: contains multiple independent attention heads from different angles, mapping temporal features to different sub-representation spaces to guide the model to learn the associated information of different sub-representation spaces to obtain richer representations; Output layer: It consists of a Dropout layer and a full connect layer. Dropout reduces the model's dependence on specific neurons by randomly discarding some neurons, making the model more robust. Full connect outputs the final classification results and regression values.

5. The method for classifying multiple arrhythmia signals according to claim 4, characterized in that: The specific training process and signal transmission process of the CNN-BILSTM-ATTENTION model are as follows: Input the data used for CNN-BILSTM-ATTENTION model training; Reconstruct the input raw data; The CNN-BILSTM-ATTENTION model will be initialized with the optimal hyperparameters found using Bayesian optimization; The reconstructed signal is sequentially passed through multiple convolutional layers and pooling layers for CNN calculation to obtain the CNN output OutPut-One; OutPut-One passes through multiple hidden layers of BILSTM in sequence to obtain the output OutPut-Two of BILSTM; OutPut-Two is again subjected to CNN calculations through new convolutional and pooling layers to obtain CNN output OutPut-Three; OutPut-Three is processed by the Multi-head Attention layer to obtain the output OutPut-Four; OutPut-Four is finally calculated by the fully connected layer to obtain the final result and calculate the predicted loss value; Verify whether the final loss value triggers the early stopping mechanism. If so, save the training model. If not, go back to step 4 and repeat the above process until the early stopping condition is met.

6. The method for classifying multiple arrhythmia signals according to claim 5, characterized in that: In step 3, the misclassification of each category is extracted from the confusion matrix; According to the total number of misclassified samples and class samples, the misclassification rate of each class is calculated: Dynamically adjust weights based on misjudgment rate: Dynamic adjustment of weights based on task objectives: Among them, Base is the initial weight, which reflects the basic preference of the task target for the indicator; ErrorFactorAcc is the mean of the overall misclassification rate, which measures the overall classification accuracy of the model; ErrorFactorRecall is the sum of the proportions of samples in each category that are misclassified as other categories, emphasizing the impact of misclassification on recall rate; ErrorFactorPrecision is the sum of the proportions of samples in each category that are misclassified as the target category, emphasizing the impact of misclassification on precision; α is a dynamic adjustment coefficient, which is used to control the influence of the misjudgment rate on the weight.

7. The method for classifying multiple arrhythmia signals according to claim 6, characterized in that: Calculate the harmonic mean of all evaluation methods, including accuracy, recall, F1 score, and precision for heart rate signal classification; Then, the time index (T n ), representing a metric proportional to computation time. Among them, the coefficient σ∈(0,1] represents the time ratio, which is determined according to the time sensitivity of the task; the higher σ is, the stronger the time sensitivity of the task is; σ is set to 0.2, and normalize(Time) means to normalize the computation time of all methods; MDPI=T n ×H n ; The time index was multiplied by the harmonic mean to provide a comprehensive assessment of the computation time and accuracy of all methods.

Images