Self-adaptive increment fusion system for intelligent recognition of arrhythmia
By employing an adaptive weight fusion mechanism and a dynamic threshold adjustment strategy, the update lag problem of BLS when adding new data or categories is solved, achieving efficient and stable intelligent arrhythmia recognition, which is suitable for wearable devices and real-time monitoring systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-03-27
AI Technical Summary
Existing BLS and its incremental learning strategies generally suffer from limitations such as being able to handle only known categories, lagging model updates, and sensitivity to data imbalance, making it difficult to meet the needs of dynamic growth and diversification of ECG data in clinical settings.
We introduce an adaptive weight fusion mechanism and a dynamic threshold adjustment strategy. By deriving the trace and norm of the Gram matrix, we reveal the intrinsic relationship between the fusion weights and data features, enabling efficient updates of new data or categories. This supports continuous learning and self-updating of the model, avoiding performance degradation and catastrophic forgetting in traditional methods.
It enables efficient model updates when new data or categories are added, improves model scalability and training efficiency, maintains high recognition accuracy, and is suitable for wearable devices and real-time heart rate monitoring systems.
Smart Images
Figure CN121744002A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of intelligent recognition of cardiovascular diseases, in particular to a self-adaptive incremental fusion system for intelligent recognition of arrhythmia. BACKGROUND
[0002] Cardiovascular disease (CVD) is one of the chronic non-communicable diseases with the highest mortality rate worldwide. As one of the main pathological types, arrhythmia not only exists widely in all kinds of people, but also is an important inducement of sudden cardiac death. The core feature of arrhythmia is the abnormality of cardiac electrical activity rhythm, which has complex and diverse manifestations, including tachycardia, bradycardia, and rhythm disorders. Since arrhythmia is often occult and sudden, its early identification and accurate classification are of great significance to reduce the risk of death and improve the clinical prognosis.
[0003] With the rapid development of artificial intelligence and biomedical signal processing technology, intelligent recognition of cardiovascular diseases has become an important research direction of medical informationization. Based on machine learning and deep learning, electrocardiogram (ECG) analysis methods can realize the automatic extraction and classification of complex arrhythmia patterns, greatly improving the accuracy and efficiency of arrhythmia detection. However, traditional deep learning models have complex structure and large amount of calculation, which are difficult to be efficiently deployed in wearable devices or real-time monitoring systems. At the same time, their training relies on large-scale labeled data, and they lack the ability to adapt to new samples and unknown categories.
[0004] Under this background, the Broad Learning System (BLS) as a kind of efficient lightweight learning framework, with the advantages of simple structure and fast training, has shown broad prospects in the field of intelligent recognition of medical signals. However, the existing BLS and its incremental learning strategies generally have limitations such as only processing known categories, model updating lag, and sensitivity to data imbalance, which are difficult to meet the dynamic growth and diversification needs of ECG data in clinical environment. To solve these problems, researchers have begun to explore new incremental learning frameworks that introduce self-adaptive weight fusion mechanisms and dynamic threshold adjustment strategies to realize continuous learning, rapid updating and robust recognition of arrhythmia, and provide more scalable and practical technical support for intelligent diagnosis of cardiovascular diseases. SUMMARY
[0005] The purpose of the present application is to solve the problem that the existing BLS and its incremental learning strategies generally have limitations such as only processing known categories, model updating lag, and sensitivity to data imbalance, which are difficult to meet the dynamic growth and diversification needs of ECG data in clinical environment, and to propose a self-adaptive incremental fusion system for intelligent recognition of arrhythmia.
[0006] An adaptive incremental fusion system for intelligent arrhythmia recognition includes:
[0007] ECG data acquisition module, ECG sequence acquisition module, standardized ECG sequence acquisition module, training set and test set acquisition module, original training set and new training set acquisition module, trained AWBIFS network model 1 acquisition module, trained AWBIFS network model 2 acquisition module, new trained AWBIFS network model 1 acquisition module, prediction module;
[0008] The electrocardiogram (ECG) data acquisition module is used to acquire ECG data;
[0009] The ECG sequence acquisition module is used to segment the ECG sampling data to obtain... Each ECG sequence and its corresponding label;
[0010] The standardized ECG sequence acquisition module is used to standardize each ECG sequence to obtain a standardized ECG sequence.
[0011] The training set and test set acquisition module is used to... The ECG sequences and their corresponding labels are divided into training and testing sets.
[0012] The original training set and new training set acquisition module is used to randomly divide the electrocardiogram sequence of each training set into an original training set and a new training set.
[0013] The module for acquiring the trained AWBIFS network model 1 is used to obtain the data matrix from the training set. Input an AWBIFS network model, and the AWBIFS network model outputs a category label matrix corresponding to the data matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve for the network weights corresponding to the data matrix Obtain the trained AWBIFS network model 1;
[0014] The trained AWBIFS network model 2 acquisition module is used to obtain new data. and the category label matrix corresponding to the new data ; New data Input an AWBIFS network model, and the AWBIFS network model outputs a class label matrix corresponding to the new data. The category label matrix corresponding to the new data output by the AWBIFS network model Solve for the network weights corresponding to the new data matrix 2. Obtain the trained AWBIFS network model.
[0015] The newly trained AWBIFS network model 1 acquisition module is used to obtain the fusion network weights based on the acquired trained AWBIFS network model 1 and the acquired trained AWBIFS network model 2. ; Integrate network weights Replace the network weights in the trained AWBIFS network model 1 This yields a new, well-trained AWBIFS network model 1;
[0016] The prediction module is used to obtain each test sample. Each test sample Input the newly trained AWBIFS network model 1, and the newly trained AWBIFS network model 1 will output the class label corresponding to each test sample.
[0017] The beneficial effects of this invention are as follows:
[0018] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an adaptive incremental fusion system for intelligent arrhythmia recognition. This system enables dynamic incremental updates for new data or new data classes. Specifically, when new training data or new classes are introduced, although the dimension of the output weights expands with the number of classes, we only need to focus on the weight parameters after training. By fusing and reconstructing the connection weights of two independent models, fusion operations can be performed between different models. In this way, it is no longer necessary to rebuild or fine-tune the network structure, thus achieving an efficient incremental learning process. More importantly, this method is unrestricted when facing class expansion: even if the parameter dimension of the output layer changes, as long as the fusion operation targets the weight parameters trained under the same structure, the model's capabilities can be directly expanded and transferred.
[0019] This invention innovatively introduces an adaptive weighted fusion mechanism based on traditional wide-range learning systems, which can dynamically allocate fusion weights according to the size of the dataset and the dispersion of features. This mechanism automatically assigns higher weights to data with larger datasets or more dispersed features, thereby achieving efficient model updates when new data or new categories are added, without the need to retrain all historical data, significantly improving the model's scalability and training efficiency.
[0020] This invention quantitatively reveals the intrinsic relationship between fusion weights and data features through the derivation of the trace and norm of the Gram matrix. It proves that when the labels of the new category and the original category do not overlap, the sub-model parameters are directly spliced into a non-optimal solution, thereby ensuring the rationality and stability of model fusion and avoiding the performance degradation problem commonly found in traditional incremental strategies.
[0021] This strategy introduces a class-adaptive threshold mechanism in the model output stage, which dynamically adjusts the decision criteria based on the sample proportion of each class. It lowers the discrimination requirement for the minority class to improve the recognition rate, while maintaining a strict standard for the majority class to prevent overfitting. This mechanism effectively alleviates the bias of the traditional argmax decision criterion towards the minority class, achieving robust classification under both balanced and extremely imbalanced data conditions.
[0022] This invention can simultaneously handle data expansion for both known and unknown categories during incremental learning, supporting continuous learning and self-updating of the model, and avoiding the "catastrophic forgetting" problem in traditional methods. Experimental results show that this invention maintains an overall accuracy of over 99% and an F1 score on various public databases, verifying its adaptability in open and dynamic environments.
[0023] This invention inherits the advantages of BLS, such as its simple structure and analytical solution, resulting in fast computation speed and low resource consumption. Compared with deep learning models, AWBIFS maintains high recognition accuracy while significantly reducing training and testing time, making it suitable for deployment in wearable devices, real-time heart rate monitoring systems, and resource-constrained edge devices. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating the overall implementation of the present invention;
[0025] Figure 2 This invention provides a diagram illustrating the central beat segmentation process of the arrhythmia classification model.
[0026] Figure 3 The overall flowchart of the arrhythmia classification model proposed in this invention is shown below. Detailed Implementation
[0027] Specific implementation method one: Combining Figure 1 , Figure 2 This embodiment describes an adaptive incremental fusion system for intelligent arrhythmia identification, comprising:
[0028] ECG data acquisition module, ECG sequence acquisition module, standardized ECG sequence acquisition module, training set and test set acquisition module, original training set and new training set acquisition module, trained AWBIFS network model 1 acquisition module, trained AWBIFS network model 2 acquisition module, new trained AWBIFS network model 1 acquisition module, prediction module;
[0029] The electrocardiogram (ECG) data acquisition module is used to acquire ECG data;
[0030] The ECG sequence acquisition module is used to segment the ECG sampling data to obtain... Each ECG sequence and its corresponding label;
[0031] The standardized ECG sequence acquisition module is used to standardize each ECG sequence to obtain a standardized ECG sequence.
[0032] The training set and test set acquisition module is used to... The ECG sequences and their corresponding labels are divided into training and testing sets.
[0033] The original training set and new training set acquisition module is used to randomly divide the electrocardiogram sequence of each training set into an original training set and a new training set.
[0034] The module for acquiring the trained AWBIFS network model 1 is used to obtain the data matrix from the training set. Input an AWBIFS network model, and the AWBIFS network model outputs a category label matrix corresponding to the data matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve for the network weights corresponding to the data matrix Obtain the trained AWBIFS network model 1; such as Figure 3 The diagram shows the overall structure of this classification network model;
[0035] The trained AWBIFS network model 2 acquisition module is used to obtain new data. and the category label matrix corresponding to the new data ; New data Input an AWBIFS network model, and the AWBIFS network model outputs a class label matrix corresponding to the new data. The category label matrix corresponding to the new data output by the AWBIFS network model Solve for the network weights corresponding to the new data matrix Obtain the trained AWBIFS network model 2; the training process is the same as S6.
[0036] The newly trained AWBIFS network model 1 acquisition module is used to obtain the fusion network weights based on the acquired trained AWBIFS network model 1 and the acquired trained AWBIFS network model 2. ; Integrate network weights Replace the network weights in the trained AWBIFS network model 1 This yields a new, well-trained AWBIFS network model 1;
[0037] The prediction module is used to obtain each test sample. Each test sample Input the newly trained AWBIFS network model 1, and the newly trained AWBIFS network model 1 will output the class label corresponding to each test sample.
[0038] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the ECG data acquisition module is used to acquire ECG data; the specific process is as follows:
[0039] Use an electrocardiogram (ECG machine or 24-hour Holter monitor) to record ECG data from different subjects. The recording time for each ECG data should be greater than 10 seconds.
[0040] Label the recorded electrocardiogram (ECG) data of different subjects with category labels;
[0041] The arrhythmia classification method based on the AMMI standard was adopted, with the categories being normal heartbeat, supraventricular ectopic beat, ventricular ectopic beat, fusion beat, and unknown.
[0042] The other steps and parameters are the same as in Specific Implementation Method 1.
[0043] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: the ECG sequence acquisition module is used to segment the ECG sampling data to obtain... Each ECG sequence and its corresponding tag; the specific process is as follows:
[0044] Obtain the R-peak position and the corresponding category label for each ECG sampling data;
[0045] The first 149 sampling points to the last 150 sampling points of each ECG sampling data were selected as a single ECG sequence, and each ECG sequence was used as the input to the AWBIFS network model.
[0046] Other steps and parameters are the same as in specific implementation method one or two.
[0047] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: the standardized ECG sequence acquisition module is used to standardize each ECG sequence to obtain a standardized ECG sequence; the specific process is as follows:
[0048] Using the Z-score standardization method, the first... A single electrocardiogram (ECG) sequence is converted into a vector with a mean of zero and a standard deviation of 1. ;
[0049] The formula for calculating the Z-score using the normalization method is as follows:
[0050] (1)
[0051] in, Indicates the first One electrocardiogram sequence, Represents the standardized first One electrocardiogram sequence, and The first ECG sequence The mean and standard deviation.
[0052] The other steps and parameters are the same as those in one of the specific implementation methods one to three.
[0053] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: the training set and test set acquisition module is used to... Each electrocardiogram (ECG) sequence and its corresponding label are divided into a training set and a test set; the specific process is as follows:
[0054] S41. According to Determining individual ECG sequences and their corresponding labels Number of categories corresponding to each ECG sequence , Indicates the total number of categories;
[0055] S42. Order , Indicates category , ;
[0056] S43. Will In the ECG sequence belonging to the number The electrocardiogram sequences of each category were randomly divided into 10 groups;
[0057] The electrocardiogram sequences in each group belong to the same classification category;
[0058] The electrocardiogram sequences in the 10 groups are completely different from each other;
[0059] All ECG sequences of the same category in the same group are spliced together in random order to form a complete ECG sequence data;
[0060] Will belong to the The first nine groups out of 10 in each category are used as the training set, and the tenth group is used as the test set;
[0061] S44. Order Repeat S43 until... This yields a training set and a test set, where the number of groups in the training set is... The number of test sets is .
[0062] The other steps and parameters are the same as those in one of the specific implementation methods one to four.
[0063] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: the original training set and new training set acquisition module is used to randomly divide the electrocardiogram sequences of each training set into an original training set and a new training set; the specific process is as follows:
[0064] The electrocardiogram sequences in each training set are randomly divided into the original training set and the newly added training set.
[0065] The original training set and the newly added training set do not contain the same ECG sequences. This is to test the recognition effect of sample augmentation or sample category augmentation.
[0066] The other steps and parameters are the same as those in any of the specific implementation methods one to five.
[0067] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: the module for acquiring the trained AWBIFS network model 1 is used to obtain the data matrix from the training set. Input an AWBIFS network model, and the AWBIFS network model outputs a category label matrix corresponding to the data matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve for the network weights corresponding to the data matrix Obtain the trained AWBIFS network model 1; such as ( Figure 3 The diagram shows the overall structure of this classification network model; the specific process is as follows:
[0068] S61. Training Set Data Matrix Data matrix The corresponding label matrix is represented as follows ;
[0069] in, express Medium sample size express The dimensions of the data Indicates the number of categories. Represent real numbers;
[0070] S62. [It is provided] Group feature mapping nodes and Group enhancement nodes;
[0071] S63. No. Group feature mapping node It is given by the following formula:
[0072] (2)
[0073] in, Indicates the first Group feature mapping nodes, ;
[0074] This indicates the initialization of random weights. ;
[0075] This indicates the initial random bias. ;
[0076] for The number of nodes in; Represents a linear mapping function;
[0077] S64. Combine all feature mapping nodes to obtain , ;
[0078] S65. Group Enhancement Node It is given by the following formula:
[0079] (3)
[0080] in, Indicates the first Group enhancement nodes, , Don't The number of nodes in;
[0081] This indicates the initialization of random weights. ;
[0082] This indicates the initial random bias. ;
[0083] To represent an activation function, typically used function;
[0084] S66. Combine all enhanced nodes to obtain , ;
[0085] S67. Connect all feature mapping nodes and all enhancement nodes into a feature expansion matrix, expressed as formula (4):
[0086] (4)
[0087] in, Represents the characteristic extended matrix;
[0088] This indicates that all feature mapping nodes and all enhancement nodes are connected;
[0089] S68. Category label matrix corresponding to the output data matrix of the AWBIFS network model The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve ; Obtain the trained AWBIFS network model 1; represented as:
[0090] (5)
[0091] in, This represents the output weight to be determined;
[0092] Category label matrix corresponding to the data matrix output by the AWBIFS network model Solve The specific process is as follows:
[0093] Construct the regularized least squares objective function, expressed as:
[0094] (6)
[0095] in, Represent the objective function; Denotes the 2-norm; the first term It is the least squares error term, the second term. It is a regularization term. It is a regularization coefficient used to prevent overfitting;
[0096] Let equation (6) be applied to The derivative is zero, so the output weights are obtained. as follows:
[0097] (7)
[0098] in, It is an identity matrix, with superscript. This indicates the transpose;
[0099] Obtain the trained AWBIFS network model 1.
[0100] The other steps and parameters are the same as those in one of the specific implementation methods one to six.
[0101] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that: the trained AWBIFS network model 2 acquisition module is used to obtain new data. and the category label matrix corresponding to the new data ; New data Input an AWBIFS network model, and the AWBIFS network model outputs a class label matrix corresponding to the new data. The category label matrix corresponding to the new data output by the AWBIFS network model Solve for the network weights corresponding to the new data matrix Obtain the trained AWBIFS network model 2; the training process is the same as S6; the specific process is as follows:
[0102] S71. Obtain New Data New data The corresponding label matrix is represented as follows ;
[0103] in, express Medium sample size express The dimensions of the data Indicates the number of categories. Represent real numbers;
[0104] S72. [It is equipped with] Group feature mapping nodes and Group enhancement nodes;
[0105] S73. No. Group feature mapping node It is given by the following formula:
[0106] (8)
[0107] in, Indicates the first Group feature mapping nodes, ; This indicates the initialization of random weights. ; This indicates the initial random bias. ; for The number of nodes in; Represents a non-linear activation function;
[0108] S74. Combine all feature mapping nodes to obtain , ;
[0109] S75. Group Enhancement Node It is given by the following formula:
[0110] (9)
[0111] in, Indicates the first Group enhancement nodes, , for The number of nodes in; This indicates the initialization of random weights. ; This indicates the initial random bias. ; To represent an activation function, typically used function;
[0112] S76. Combine all the enhanced nodes to obtain , ;
[0113] S77. Connect all feature mapping nodes and all enhancement nodes into a feature expansion matrix, expressed as formula (10):
[0114] (10)
[0115] in, Represents the characteristic extended matrix;
[0116] This indicates that all feature mapping nodes and all enhancement nodes are connected;
[0117] S78. The AWBIFS network model outputs a new data matrix corresponding to the category label matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve ; Obtain the trained AWBIFS network model2; represented as:
[0118] (11)
[0119] in, This represents the output weight to be determined;
[0120] Category label matrix corresponding to the data matrix output by the AWBIFS network model Solve The specific process is as follows:
[0121] Construct the regularized least squares objective function, expressed as:
[0122] (12)
[0123] in, Represent the objective function; Denotes the 2-norm; the first term It is the least squares error term, the second term. It is a regularization term. It is a regularization coefficient used to prevent overfitting;
[0124] Let equation (12) be applied to The derivative is zero, so the output weights are obtained. as follows:
[0125] (13)
[0126] in, It is an identity matrix, with superscript. This indicates the transpose;
[0127] Obtain the trained AWBIFS network model 2.
[0128] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.
[0129] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: the newly trained AWBIFS network model 1 acquisition module is used to obtain the fusion network weights based on the acquired trained AWBIFS network model 1 and the acquired trained AWBIFS network model 2. ; Integrate network weights Replace the network weights in the trained AWBIFS network model 1 The new, trained AWBIFS network model 1 is obtained; the specific process is as follows:
[0130] S81. The obtained trained AWBIFS network model 1 and the obtained trained AWBIFS network model 2 are represented as follows:
[0131] (14)
[0132] To find a new fusion weight matrix This allows the weights of the two models to be fused to obtain a new output. To solve this fusion problem, we first modify equation (14) to:
[0133] (15)
[0134] in, Indicates the weights of the fused network;
[0135] If the new data matrix in step seven corresponds to the category label matrix Category label matrix corresponding to the data matrix in step 6 When the number of categories is the same, do not use equation (15) and Process it;
[0136] If the new data matrix in step seven corresponds to the category label matrix Category label matrix corresponding to the data matrix in step 6 When the number of categories is inconsistent, in equation (15) and Positions with inconsistent dimensions are padded with zeros; this operation does not affect the BLS solution results.
[0137] As can be seen from the BLS model structure in equation (15), the label matrix... and One-hot encoding is used, meaning that when the number of categories in the new data is inconsistent with the number of categories in the original data, we only need to pad the positions where the dimensions of the two label matrices are inconsistent with zeros. This operation does not affect the solution result of BLS.
[0138] S82. Equation (15) is written as:
[0139] (16)
[0140] in, Represents the fusion feature extension matrix, ;
[0141] Represents the fused category label matrix. ;
[0142] S83. Construct an objective function for equation (16), expressed as:
[0143] (17)
[0144] Among them, the first item This is the least squares error term, used to measure the deviation between the predicted value and the actual value;
[0145] Second item This is a regularization term used to constrain the weights and prevent overfitting.
[0146] Here, is the regularization coefficient. ;
[0147] S84. According to formula (7), we can obtain:
[0148] (18)
[0149] S85. Will and Substituting into formula (14), we get:
[0150] (19)
[0151] Simplifying equation (19), we get:
[0152] (20)
[0153] The fusion weight matrix is derived from equation (20). It is a weight matrix and The weighted combination, with weight coefficients as follows: and ;
[0154] parameter , , , This is achieved when a trained AWBIFS network model 1 is obtained, meaning that no additional computational burden is added.
[0155] parameter , This is achieved when a pre-trained AWBIFS network model 2 is obtained, meaning no additional computational burden is added.
[0156] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.
[0157] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that: the prediction module is used to obtain each test sample. Each test sample Input the newly trained AWBIFS network model 1, and the newly trained AWBIFS network model 1 will output the class label corresponding to each test sample; the specific process is as follows:
[0158] (twenty one)
[0159] in, Indicates test sample The corresponding feature extension matrix;
[0160] It is a test sample The feature extension matrix (4) is obtained through feature mapping formula (2) and feature enhancement formula (3);
[0161] This is the output vector; Indicates the total number of categories;
[0162] The output vector of the model Normalization is performed using the Softmax function to obtain the... The first sample Category confidence of each category ; indicates as:
[0163] (twenty two)
[0164] in, Indicates the first Test sample Indicates the first The first sample Predicted labels for each category dimension; ;
[0165] For a classifier, the target class is set as positive examples and other classes are set as negative examples. The number of positive examples is counted. Number of counterexamples The probability of observation can be calculated. Since we typically assume that the training set is an unbiased sample of the real population, the observed probability represents the true probability. Therefore, if the classifier's predicted value is greater than the observed probability, it should be classified as a positive example, i.e.:
[0166] (twenty three)
[0167] in, Indicates the number of positive examples. Indicates the number of counterexamples;
[0168] Based on the The first sample Category confidence of each category Adjust formula (23) to obtain the decision threshold for each category, where the first... The decision threshold for the class is ; ;
[0169] in, Indicates the first Class decision threshold Indicates the sample is the first The number of classes Indicates that the sample is not the first The number of classes;
[0170] For the test samples If only one category's confidence level satisfies formula (24), then the predicted category is... ;
[0171] For the test samples If the confidence scores of two or more categories satisfy formula (24), then the predicted category should be the category with the smallest data volume. ;
[0172] For the test samples If no category has a confidence level that satisfies formula (24), then the predicted category is preferentially identified as the category with the highest confidence level. ;
[0173] (twenty four)
[0174] in, Indicates the first test samples Belongs to the first Confidence level of the class.
[0175] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.
[0176] In evaluating the effectiveness of the method in this invention, three metrics were used: accuracy, F1 score, and G_mean. Accuracy reflects the overall accuracy of the model's final classification and the classification accuracy of each sample. F1 score and G_mean reflect the effect of imbalanced class identification. The larger the values of F1 score and G_mean, the better the imbalanced classification effect.
[0177] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.
Claims
1. An adaptive incremental fusion system for intelligent identification of cardiac arrhythmias, characterized in that: The system includes: ECG data acquisition module, ECG sequence acquisition module, standardized ECG sequence acquisition module, training set and test set acquisition module, original training set and new training set acquisition module, trained AWBIFS network model 1 acquisition module, trained AWBIFS network model 2 acquisition module, new trained AWBIFS network model 1 acquisition module, prediction module; The electrocardiogram (ECG) data acquisition module is used to acquire ECG data; The ECG sequence acquisition module is used to segment the ECG sampling data to obtain... Each ECG sequence and its corresponding label; The standardized ECG sequence acquisition module is used to standardize each ECG sequence to obtain a standardized ECG sequence. The training set and test set acquisition module is used to... The ECG sequences and their corresponding labels are divided into training and testing sets. The original training set and new training set acquisition module is used to randomly divide the electrocardiogram sequence of each training set into an original training set and a new training set. The module for acquiring the trained AWBIFS network model 1 is used to obtain the data matrix from the training set. Input an AWBIFS network model, and the AWBIFS network model outputs a category label matrix corresponding to the data matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve for the network weights corresponding to the data matrix Obtain the trained AWBIFS network model 1; The trained AWBIFS network model 2 acquisition module is used to obtain new data. and the category label matrix corresponding to the new data ; New data Input an AWBIFS network model, and the AWBIFS network model outputs a class label matrix corresponding to the new data. The category label matrix corresponding to the new data output by the AWBIFS network model Solve for the network weights corresponding to the new data matrix 2. Obtain the trained AWBIFS network model. The newly trained AWBIFS network model 1 acquisition module is used to obtain the fusion network weights based on the acquired trained AWBIFS network model 1 and the acquired trained AWBIFS network model 2. ; Integrate network weights Replace the network weights in the trained AWBIFS network model 1 This yields a new, well-trained AWBIFS network model 1; The prediction module is used to obtain each test sample. Each test sample Input the newly trained AWBIFS network model 1, and the newly trained AWBIFS network model 1 will output the class label corresponding to each test sample.
2. The adaptive incremental fusion system for intelligent arrhythmia identification according to claim 1, characterized in that: The electrocardiogram (ECG) data acquisition module is used to acquire ECG data; the specific process is as follows: Use an electrocardiogram (ECG) signal acquisition device to record ECG sampling data from different subjects. The recording time for each ECG sampling data should be greater than 10 seconds. Label the recorded electrocardiogram (ECG) data of different subjects with category labels; The categories are normal heartbeat, supraventricular ectopic beat, ventricular ectopic beat, fusion beat, and unknown.
3. The adaptive incremental fusion system for intelligent arrhythmia identification according to claim 2, characterized in that: The ECG sequence acquisition module is used to segment the ECG sampling data to obtain... Each ECG sequence and its corresponding label; The specific process is as follows: Obtain the R-peak position and the corresponding category label for each ECG sampling data; The first 149 sampling points to the last 150 sampling points of each ECG sampling data were selected as a single ECG sequence, and each ECG sequence was used as the input to the AWBIFS network model.
4. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 3, characterized in that: The standardized ECG sequence acquisition module is used to standardize each ECG sequence to obtain a standardized ECG sequence; the specific process is as follows: Using the Z-score standardization method, the first... A single electrocardiogram (ECG) sequence is converted into a vector with a mean of zero and a standard deviation of 1. ; The formula for calculating the Z-score using the normalization method is as follows: （1） in, Indicates the first One electrocardiogram sequence, Represents the standardized first One electrocardiogram sequence, and The first ECG sequence The mean and standard deviation.
5. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 4, characterized in that: The training set and test set acquisition module is used to... Each electrocardiogram (ECG) sequence and its corresponding label are divided into a training set and a test set; the specific process is as follows: S41. According to Determining individual ECG sequences and their corresponding labels Number of categories corresponding to each ECG sequence , Indicates the total number of categories; S42. Order , Indicates category , ; S43. Will In the ECG sequence belonging to the number The electrocardiogram sequences of each category were randomly divided into 10 groups; The electrocardiogram sequences in each group belong to the same classification category; The electrocardiogram sequences in the 10 groups are completely different from each other; All ECG sequences of the same category in the same group are spliced together in random order to form a complete ECG sequence data; Will belong to the The first nine groups out of 10 in each category are used as the training set, and the tenth group is used as the test set; S44. Order Repeat S43 until... This yields a training set and a test set, where the number of groups in the training set is... The number of test sets is .
6. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 5, characterized in that: The original training set and new training set acquisition module is used to randomly divide the electrocardiogram sequences of each training set into an original training set and a new training set; the specific process is as follows: The electrocardiogram sequences in each training set are randomly divided into the original training set and the newly added training set. The original training set and the newly added training set do not contain the same ECG sequences.
7. The adaptive incremental fusion system for intelligent arrhythmia identification according to claim 6, characterized in that: The module for acquiring the trained AWBIFS network model 1 is used to obtain the data matrix from the training set. Input an AWBIFS network model, and the AWBIFS network model outputs a category label matrix corresponding to the data matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve for the network weights corresponding to the data matrix The trained AWBIFS network model 1 is obtained; the specific process is as follows: S61. Training Set Data Matrix Data matrix The corresponding label matrix is represented as follows ; in, express Medium sample size express The dimensions of the data Indicates the number of categories. Represent real numbers; S62. [It is provided] Group feature mapping nodes and Group enhancement nodes; S63. No. Group feature mapping node It is given by the following formula: （2） in, Indicates the first Group feature mapping nodes, ; This indicates the initialization of random weights. ; This indicates the initial random bias. ; for The number of nodes in; Represents a linear mapping function; S64. Combine all feature mapping nodes to obtain , ; S65. Group Enhancement Nodes It is given by the following formula: （3） in, Indicates the first Group enhancement nodes, , Don't The number of nodes in; This indicates the initialization of random weights. ; This indicates the initial random bias. ; To represent an activation function, typically used function; S66. Combine all the enhanced nodes to obtain , ; S67. Connect all feature mapping nodes and all enhancement nodes into a feature expansion matrix, expressed as formula (4): （4） in, Represents the characteristic extended matrix; This indicates that all feature mapping nodes and all enhancement nodes are connected; S68. Category label matrix corresponding to the output data matrix of the AWBIFS network model The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve ; Obtain the trained AWBIFS network model 1; represented as: （5） in, This represents the output weight to be determined; Category label matrix corresponding to the data matrix output by the AWBIFS network model Solve The specific process is as follows: Construct the regularized least squares objective function, expressed as: （6） in, Represent the objective function; Denotes the 2-norm; the first term It is the least squares error term, the second term. It is a regularization term. It is the regularization coefficient; Let equation (6) be applied to The derivative is zero, so the output weights are obtained. as follows: （7） in, It is an identity matrix, with superscript. This indicates the transpose; Obtain the trained AWBIFS network model 1.
8. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 7, characterized in that: The trained AWBIFS network model 2 acquisition module is used to obtain new data. and the category label matrix corresponding to the new data ; New data Input an AWBIFS network model, and the AWBIFS network model outputs a class label matrix corresponding to the new data. The category label matrix corresponding to the new data output by the AWBIFS network model Solve for the network weights corresponding to the new data matrix The trained AWBIFS network model 2 is obtained; the specific process is as follows: S71. Obtain New Data New data The corresponding label matrix is represented as follows ; in, express Medium sample size express The dimensions of the data Indicates the number of categories. Represent real numbers; S72. [It is equipped with] Group feature mapping nodes and Group enhancement nodes; S73. No. Group feature mapping node It is given by the following formula: （8） in, Indicates the first Group feature mapping nodes, ; This indicates the initialization of random weights. ; This indicates the initial random bias. ; for The number of nodes in; Represents a non-linear activation function; S74. Combine all feature mapping nodes to obtain , ; S75. Group Enhancement Nodes It is given by the following formula: （9） in, Indicates the first Group enhancement nodes, , for The number of nodes in; This indicates the initialization of random weights. ; This indicates the initial random bias. ; To represent an activation function, typically used function; S76. Combine all the enhanced nodes to obtain , ; S77. Connect all feature mapping nodes and all enhancement nodes into a feature expansion matrix, expressed as formula (10): （10） in, Represents the characteristic extended matrix; This indicates that all feature mapping nodes and all enhancement nodes are connected; S78. The AWBIFS network model outputs a new data matrix corresponding to the category label matrix. The category label matrix corresponding to the data matrix output by the AWBIFS network model Solve ; Obtain the trained AWBIFS network model2; represented as: （11） in, This represents the output weight to be determined; Category label matrix corresponding to the data matrix output by the AWBIFS network model Solve The specific process is as follows: Construct the regularized least squares objective function, expressed as: （12） in, Represent the objective function; Denotes the 2-norm; the first term It is the least squares error term, the second term. It is a regularization term. It is the regularization coefficient; Let equation (12) be applied to The derivative is zero, so the output weights are obtained. as follows: （13） in, It is an identity matrix, with superscript. This indicates the transpose; Obtain the trained AWBIFS network model 2.
9. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 8, characterized in that: The newly trained AWBIFS network model 1 acquisition module is used to obtain the fusion network weights based on the acquired trained AWBIFS network model 1 and the acquired trained AWBIFS network model 2. ; Integrate network weights Replace the network weights in the trained AWBIFS network model 1 The new, trained AWBIFS network model 1 is obtained; the specific process is as follows: S81. The obtained trained AWBIFS network model 1 and the obtained trained AWBIFS network model 2 are represented as follows: （14） Change equation (14) to: （15） in, Indicates the weights of the fused network; If the new data matrix in step seven corresponds to the category label matrix Category label matrix corresponding to the data matrix in step 6 When the number of categories is the same, do not use equation (15) and Process it; If the new data matrix in step seven corresponds to the category label matrix Category label matrix corresponding to the data matrix in step 6 When the number of categories is inconsistent, in equation (15) and Positions with inconsistent dimensions are padded with zeros. S82. Equation (15) is written as: (16) in, Represents the fusion feature extension matrix, ; Represents the fused category label matrix. ; S83. Construct an objective function for equation (16), expressed as: (17) Among them, the first item The least squares error term; the second term For regularization terms; Here, is the regularization coefficient. ; S84. According to formula (7), we can obtain: (18) S85. Will and Substituting into formula (14), we get: (19) Simplifying equation (19), we get: (20) The fusion weight matrix is derived from equation (20). It is a weight matrix and The weighted combination, with weight coefficients as follows: and .
10. The adaptive incremental fusion system for intelligent identification of arrhythmias according to claim 9, characterized in that: The prediction module is used to obtain each test sample. Each test sample Input the newly trained AWBIFS network model 1, and the newly trained AWBIFS network model 1 will output the class label corresponding to each test sample; The specific process is as follows: (21) in, Indicates test sample The corresponding feature extension matrix; It is a test sample The feature extension matrix (4) is obtained through feature mapping formula (2) and feature enhancement formula (3); This is the output vector; Indicates the total number of categories; The output vector of the model Normalization is performed using the Softmax function to obtain the... The first sample Category confidence of each category ; indicates as: (22) in, Indicates the first Test sample Indicates the first The first sample Predicted labels for each category dimension; ; If the classifier's predicted value is greater than the observed probability, it should be classified as a positive example, that is: (23) in, Indicates the number of positive examples. Indicates the number of counterexamples; Based on the The first sample Category confidence of each category Adjust formula (23) to obtain the decision threshold for each category, where the first... The decision threshold for the class is ; ; in, Indicates the first Class decision threshold Indicates the sample is the first The number of classes Indicates that the sample is not the first The number of classes; For the test samples If only one category's confidence level satisfies formula (24), then the predicted category is... ; For the test samples If the confidence scores of two or more categories satisfy formula (24), then the predicted category should be the category with the smallest data volume. ; For the test samples If no category has a confidence level that satisfies formula (24), then the predicted category is preferentially identified as the category with the highest confidence level. ; (24) in, Indicates the first test samples Belongs to the first Confidence level of the class.