A heart disease auxiliary diagnosis classification method based on synaptic plasticity random configuration network

Abstract

The application belongs to the field of artificial intelligence, and discloses a heart disease auxiliary diagnosis classification method based on a synapse plasticity random configuration network. The method is to input a processed sample set to be tested with unknown labels into a prediction model constructed and trained, obtain an output matrix of a class of hidden heart diseases, and then analyze the output matrix to obtain a predicted class of heart diseases. The construction and training process of the prediction model is as follows: after initializing parameters of the random configuration network, hidden layer neurons are added one by one, then the connection weights between the input layer and the hidden layer of the random configuration network are unsupervised trained through a synapse plasticity learning rule, the connection weights between the hidden layer and the output layer of the random configuration network are calculated by using a least square method, the above process is repeated until a termination condition is reached. The application improves the weight learning ability and performance representation of the network model, and further improves the classification accuracy of the network model in the heart disease auxiliary diagnosis classification task.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and relates to a method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity random configuration networks. Background Technology

[0002] As one of the major diseases threatening human health, early screening and auxiliary diagnosis of heart disease are of great significance for reducing mortality and improving the quality of life of patients.

[0003] Currently, heart disease is typically diagnosed clinically using methods such as electrocardiograms (ECG), echocardiography, and blood tests. These diagnostic methods rely to some extent on the doctor's clinical experience and professional judgment, and some methods are costly and time-consuming. With the development of artificial intelligence and data mining technologies, data-driven classification methods for heart disease diagnosis are increasingly being applied to the early screening and auxiliary diagnosis of heart disease. By analyzing clinical indicators such as age, gender, type of chest pain, resting blood pressure, cholesterol, fasting blood glucose, and maximum heart rate, classification models are established to determine whether a patient has heart disease, providing doctors with auxiliary decision-making support and improving diagnostic efficiency and accuracy.

[0004] Traditional classification methods for cardiac auxiliary diagnosis are represented by machine learning models such as logistic regression, support vector machines, decision trees, and random forests, which have good classification performance in small sample scenarios. However, cardiac data often exhibits complex characteristics such as nonlinearity and multi-scale coupling, and traditional methods are insufficient in modeling such complex data. To improve the model's ability to model nonlinearity of complex data, stochastic vector function connection networks (SVMs) have been increasingly applied in the field of cardiac auxiliary diagnosis. SVMs have advantages such as simple structure, high training efficiency, and strong nonlinear fitting ability. However, insufficient number of hidden layer nodes or improper parameter settings can lead to a significant decrease in the model's nonlinear fitting ability. Stochastic configuration networks (SVMs) propose a method to gradually generate hidden layer neurons through a unique supervised mechanism to build the network and reduce the burden of parameter tuning. However, in the traditional stochastic configuration network, after the input layer connection weights of the hidden layer neurons are determined by the supervised mechanism and added to the network, their parameters remain fixed and no longer participate in subsequent optimization. As new hidden layer neurons are continuously added to the network, the connection weights of the existing hidden layer neurons cannot be adaptively adjusted according to the sample feature distribution, and they cannot achieve co-optimization with the new hidden layer neurons, thus limiting the model's performance.

[0005] Meanwhile, synaptic plasticity, as a core foundation of the biological nervous system, offers a new approach to solving the problem of fixed random weight allocation through brain-inspired computational modeling of synaptic plasticity. Specifically, synaptic plasticity refers to the specific changes in synaptic strength that depend on the firing activity between neurons; this mechanism is widely considered fundamental to learning and memory. Computational modeling of synaptic plasticity, as an unsupervised learning mechanism, can effectively improve the weight learning ability of networks.

[0006] In the prior art, for example, patent application CN121798595A discloses a method for adjusting the connection weights of a spiking neural network using a synaptic plasticity learning mechanism. However, the performance of the spiking neural network in real-world task scenarios is still limited, and the synaptic plasticity learning rule used is based on encoding the discharge peak moment. In addition, the connection weights between the input layer and the hidden layer of a traditional randomly configured network remain fixed after random generation and cannot be adaptively adjusted according to the features of the input samples, which limits the feature representation ability and classification performance of the randomly configured network.

[0007] Patent application CN116992935A discloses a neural network model training method based on an improved random allocation algorithm. This method selects optimal hidden layer nodes and fixes their parameters using inequality constraints. However, in this method, after adjusting and adding candidate hidden layer neurons that satisfy the constraints (i.e., fixing their parameters), the parameters of existing hidden layer neurons remain unchanged when new hidden layer neurons are added subsequently. This fixing method causes early hidden layer neurons to be unable to adapt to the feature distribution of subsequent data, leading to error accumulation and limiting model performance and generalization ability.

[0008] Therefore, this invention constructs a brain-inspired unsupervised synaptic plasticity learning method to realize the classification of cardiac disease auxiliary diagnosis using a randomized network, aiming to improve the weight learning ability and performance representation of the randomized network, thereby improving the accuracy and generalization ability of the randomized network in the cardiac disease auxiliary diagnosis classification task. Summary of the Invention

[0009] The purpose of this invention is to address the problems existing in the prior art and provide a method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity random configuration networks.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A classification method for cardiac disease auxiliary diagnosis based on synaptic plasticity randomized configuration networks, which processes missing values, Standardized test sample set with unknown labels The input is fed into the constructed and trained prediction model to obtain the output matrix of implicit heart disease categories. For the output matrix The analysis yields the predicted type of heart disease;

[0012] The construction and training process of the prediction model is as follows: After initializing the parameters of the randomized network (a randomized network is a three-layer feedforward neural network, including an input layer, a hidden layer, and an output layer), hidden layer neurons are added one by one. During this process, candidate hidden layer neurons are randomly generated each time. The optimal candidate hidden layer neurons are selected and added to the randomized network through the supervision mechanism of the randomized network. Then, the connection weights between the input layer and the hidden layer of the randomized network are trained unsupervisedly through the synaptic plasticity learning rule (i.e., unsupervised training is performed on the newly added hidden layer neurons and all the added hidden layer neurons together). The connection weights between the hidden layer and the output layer of the randomized network are calculated using the least squares method. The above process is repeated until the termination condition is met.

[0013] As a preferred technical solution:

[0014] The above describes a cardiac diagnosis classification method based on a stochastically configured network with synaptic plasticity. The parameters of the stochastically configured network include: the maximum number of neurons in the hidden layer. Allowable output error The maximum number of candidate hidden layer neurons generated in each round The set of weight distribution parameters between the input layer and the hidden layer ,in , These are the upper and lower boundaries of the weight distribution. The variable is the amount of change.

[0015] The steps for constructing and training the prediction model for a cardiac auxiliary diagnostic classification method based on synaptic plasticity stochastic configuration networks, as described above, are as follows:

[0016] (a) Constructing the training sample set Training label set Test sample set and test tag set ;

[0017] Test sample set with unknown labels Feature space and training sample set ;

[0018] (b) Initialize the randomly configured network parameters;

[0019] (c) Let ;

[0020] (d) Randomly generate the first The candidate hidden layer neurons of each hidden layer are selected by a supervision mechanism of a randomized network, and the optimal candidate hidden layer neuron is added to the randomized network.

[0021] (e) Learning rules based on synaptic plasticity for individuals with The connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons are trained unsupervised, and the least squares method is used to calculate the connection weights. The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons;

[0022] (f) Determine if the termination condition has been met. If yes, stop and use the randomly configured network from the last iteration as the prediction model; otherwise, let... Then, return to step (d);

[0023] The termination condition is the root mean square error of the randomly configured network. Less than the allowable output error or .

[0024] The cardiac auxiliary diagnostic classification method based on synaptic plasticity randomized configuration network, as described above, has the following specific steps (a):

[0025] (a1) Obtain cardiac data, which includes Each sample and its corresponding label will The sample set consists of 10 samples. , For the first The first sample One characteristic variable, , , This represents the total number of feature variables in each sample;

[0026] Will The labels corresponding to each sample form a label set. , For the first The label corresponding to each sample , The total number of categories of heart disease;

[0027] (a2) For the sample set Missing values ​​are handled by filling in the missing values ​​of a feature variable with missing values ​​using the mean of that feature variable across all non-missing samples.

[0028] (a3) For the sample set conduct Standardization processing, the formula is as follows:

[0029] ;

[0030] In the formula, for After standardization The first sample One characteristic variable, for The first sample The mean of each characteristic variable, for The first sample The standard deviation of each characteristic variable;

[0031] Cardiac data includes indicators with different dimensions, such as cholesterol, fasting blood glucose, and maximum heart rate. The large differences in the numerical ranges of these indicators make model convergence difficult. After standardization, different indicators are transformed to a uniform scale with a mean of 0 and a standard deviation of 1, which can accelerate the convergence speed of the model.

[0032] (a4) For the tag set Perform one-hot encoding; the tag set after one-hot encoding. , For the first The label corresponding to the nth sample Each encoded value, , ;

[0033] (a5) Divide the sample set into a 4:1 ratio. Divided into training sample set and test sample set At the same time, the tag set Divided into training label sets and test tag set .

[0034] The cardiac auxiliary diagnostic classification method based on synaptic plasticity randomized configuration network, as described above, has the following specific steps (d):

[0035] (d1) The set of weight distribution parameters between the input layer and the hidden layer Randomly generate connection weights between the input layer and the hidden layer. Bias between input layer and hidden layer To generate Candidate hidden layer neurons , For the first One candidate hidden layer neuron, For input layer and Connection weights between them For input layer and The offset between them ;

[0036] (d2) The above Substitute the candidate hidden layer neurons sequentially into the following formula:

[0037] ;

[0038] ;

[0039] = ;

[0040] ;

[0041] ;

[0042] in, for The judgment value, for and the The decision value corresponding to each output layer neuron. , , For having The first hidden layer of the randomly configured network of neurons The output residual vector of each output layer neuron For the input training sample set After having A network with randomly configured hidden layer neurons Output vector (training sample set) Total For each sample, there are samples, and for each sample, there are A network with randomly configured hidden layer neurons Each of them will have a corresponding output value. The output values ​​together constitute the output vector. This is an inner product operation, where r is a constant (ranging from 0.001 to 0.1). The sequence is a non-negative real number sequence, which increases with the number of neurons in the hidden layer. The increase gradually decreases and approaches 0. This is the activation function (here, the Sigmoid function). For training label set The One encoded vector (training label set) Total Each tag has a corresponding tag number. Each encoded value, The first The coded values ​​together constitute the first (one encoded vector) , For the input training sample set After having The first hidden layer of the randomly configured network of neurons The output vectors of each output layer neuron (training sample set) Total For each sample, there are samples, and for each sample, there are The first hidden layer of the randomly configured network of neurons Each neuron in the output layer will have a corresponding output value. The output values ​​together constitute the output vector. For the input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons (from the training sample set). Total For each sample, there are samples, and for each sample, there are Each randomly configured hidden layer neuron in the network will output a corresponding output vector. The output vectors together constitute the output matrix. For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons. For having The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons;

[0043] (d3) when satisfying Select from candidate hidden layer neurons The candidate hidden layer neuron with the largest value is added to the layer with... In a network with randomly configured neurons in the nth hidden layer, as the th One hidden layer of neurons.

[0044] As described above, in the cardiac auxiliary diagnostic classification method based on synaptic plasticity randomized configuration networks, step (e) involves learning rules based on synaptic plasticity to classify the synaptic plasticity network. The formula used for unsupervised training of the connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons is as follows:

[0045]

[0046] ;

[0047] ;

[0048]

[0049] In the formula, For having The first hidden layer of the randomly configured network of neurons The input layer neuron and the first The connection weights between hidden layer neurons For having The first hidden layer of the randomly configured network of neurons The input layer neuron and the first The connection weight adjustment term between hidden layer neurons , , , This is the learning rate parameter (with a value range of 0 to 0.01). For having The first hidden layer of the randomly configured network of neurons The output value of each input layer neuron. For having The first hidden layer of the randomly configured network of neurons The output value of each hidden layer neuron For having The first hidden layer of the randomly configured network of neurons The dynamic threshold of each hidden layer neuron (this parameter is determined by calculation). To build a randomly configured network up to the th When the first hidden layer neuron is... The output value of each hidden layer neuron , The length of the historical response window. For the first The number of historical response states of each hidden layer neuron.

[0050] As described above, in the cardiac auxiliary diagnostic classification method based on synaptic plasticity stochastic configuration networks, step (e) uses the least squares method to calculate the... The connection weights between the hidden and output layers of a randomly configured network with hidden layer neurons are calculated using the following formula:

[0051] ;

[0052] ;

[0053] In the formula, For calculation using the least squares method, with The connection weights between the hidden and output layers of a network are determined by the randomized configuration of neurons in each hidden layer. For having The output matrix of the hidden layer of a network with randomly configured hidden layer neurons (from the training sample set). Total For each sample, there are samples, and for each sample, there are In a network with randomly configured hidden layer neurons, each hidden layer will output a corresponding output vector. The output vectors together constitute the output matrix. This is the activation function (here, the Sigmoid function). For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons.

[0054] As described above, in the cardiac auxiliary diagnostic classification method based on synaptic plasticity randomized configuration networks, in step (f), the root mean square error of the randomized configuration network is... Less than the allowable output error , The calculation formula is as follows:

[0055] ;

[0056] in, Representative input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons is the first... The output vector of the th ... Each output value (training sample set) Total For each sample, there are samples, and for each sample, there are Each randomly configured hidden layer neuron in the network will output a corresponding output vector. The output vectors together constitute the output matrix. Representative training label set The Middle The first label Each encoded value, , .

[0057] The cardiac auxiliary diagnostic classification method based on synaptic plasticity stochastic configuration networks as described above further includes step (g): preparing the test sample set... The data is input into the prediction model to obtain the classification results. The classification accuracy is then calculated using the following formula:

[0058] ;

[0059] In the formula, This represents the number of correctly classified test samples. This represents the number of test samples.

[0060] This invention is implemented using computer algorithms and models, and belongs to information processing methods, not disease diagnosis methods. All steps are executed automatically by computer programs, without involving subjective judgment or clinical decision-making by doctors. The output classification results do not constitute medical diagnostic conclusions, but are only used as supplementary reference information. Computer models cannot make comprehensive judgments by combining patients' clinical symptoms and past medical history like doctors do; their role is to improve the classification accuracy of network models in the auxiliary diagnostic classification task of heart disease, not to replace professional medical diagnosis.

[0061] Beneficial effects:

[0062] This invention, after adding candidate hidden layer neurons to the network, performs unsupervised training on both the newly added and existing hidden layer neurons to achieve global adjustment of all existing hidden layer neurons in the network. This allows the connection weights to be adaptively updated based on the response characteristics of the hidden layer neurons, thereby improving the weight learning ability and overall performance of the network model. As a result, the classification accuracy of the network model in the task of assisting in the diagnosis and classification of heart disease is improved, solving the problems of parameter fixation and poor performance in the prior art. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the framework of the cardiac auxiliary diagnosis classification method based on synaptic plasticity randomized configuration network of the present invention;

[0064] Figure 2 The bar chart shows the classification accuracy comparison of Example 1 and Comparison Methods 1 to 3 on the test set. Detailed Implementation

[0065] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0066] A classification method for cardiac disease auxiliary diagnosis based on synaptic plasticity randomized configuration networks, the framework of which is as follows: Figure 1 As shown, the steps are as follows:

[0067] (1) Constructing the training sample set Training label set Test sample set and test tag set ;

[0068] (1.1) Obtain cardiac data, said cardiac data including Each sample and its corresponding label will The sample set consists of 10 samples. , For the first The first sample One characteristic variable, , , This represents the total number of feature variables in each sample;

[0069] Will The labels corresponding to each sample form a label set. , For the first The label corresponding to each sample , The total number of categories of heart disease;

[0070] (1.2) For the sample set Missing value handling involves imputing missing values ​​for feature variables with missing values ​​using the mean of that feature variable across all non-missing samples. For example, when... When missing, use mean of Perform filling processing;

[0071] (1.3) For the sample set conduct Standardization processing, the formula is as follows:

[0072] ;

[0073] In the formula, for After standardization The first sample One characteristic variable, for The first sample The mean of each characteristic variable, for The first sample The standard deviation of each characteristic variable;

[0074] (1.4) For the tag set Perform one-hot encoding; the tag set after one-hot encoding. , For the first The label corresponding to the nth sample Each encoded value, , For example, when When it is 1, =1, All other encoded values ​​are 0, when When it is 4, =1, All other encoded values ​​are 0, when When it is C, =1, All other encoded values ​​are 0;

[0075] (1.5) Divide the sample set into a 4:1 ratio. Divided into training sample set and test sample set At the same time, the tag set Divided into training label sets and test tag set ;

[0076] Among them, the training sample set The number of training samples is , ;

[0077] Test sample set The number of test samples is , ;

[0078] Training label set ;

[0079] Test tag set ;

[0080] (2) Initialize the network parameters by random configuration; wherein, the parameters of the network by random configuration include: the maximum number of hidden layer neurons. Allowable output error The maximum number of candidate hidden layer neurons generated in each round The set of weight distribution parameters between the input layer and the hidden layer ,in , These are the upper and lower boundaries of the weight distribution. The change;

[0081] (3) Let ;

[0082] (4) Randomly generate the first The candidate hidden layer neurons of each hidden layer are selected by a supervision mechanism of a randomized network, and the optimal candidate hidden layer neuron is added to the randomized network.

[0083] (4.1) From the set of weight distribution parameters between the input layer and the hidden layer Randomly generate connection weights between the input layer and the hidden layer. Bias between input layer and hidden layer To generate Candidate hidden layer neurons , For the first One candidate hidden layer neuron, For input layer and Connection weights between them For input layer and The offset between them ;

[0084] (4.2) The above Substitute the candidate hidden layer neurons sequentially into the following formula:

[0085] ;

[0086] ;

[0087] = ;

[0088] ;

[0089] ;

[0090] in, for The judgment value, for and the The decision value corresponding to each output layer neuron. , , For having The first hidden layer of the randomly configured network of neurons. The output residual vector of each output layer neuron For the input training sample set After having A network with randomly configured hidden layer neurons Output vector (training sample set) Total For each sample, there are samples, with A network with randomly configured hidden layer neurons Each of them will have a corresponding output value. The output values ​​together constitute the output vector. This is an inner product operation, where r is a constant (ranging from 0.001 to 0.1). The sequence is a non-negative real number sequence, which increases with the number of neurons in the hidden layer. The increase gradually decreases and approaches 0. This is the activation function (here, the Sigmoid function). For training label set The One encoded vector (training label set) Total Each tag has a corresponding tag number. Each encoded value, The first The coded values ​​together constitute the first (one encoded vector) , For the input training sample set After having The first hidden layer of the randomly configured network of neurons. The output vectors of each output layer neuron (training sample set) Total For each sample, there are samples, and for each sample, there are The first hidden layer of the randomly configured network of neurons. Each neuron in the output layer will have a corresponding output value. The output values ​​together constitute the output vector. For the input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons (from the training sample set). Total For each sample, there are samples, with Each randomly configured hidden layer neuron in the network will output a corresponding output vector. The output vectors together constitute the output matrix. For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons. For having The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons;

[0091] (4.3) When satisfying Select from candidate hidden layer neurons The candidate hidden layer neuron with the largest value is added to the layer with... In a network with randomly configured neurons in the nth hidden layer, as the th One hidden layer neuron;

[0092] (5) Learning rules based on synaptic plasticity for individuals with... The connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons are trained unsupervised, and the least squares method is used to calculate the connection weights. The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons;

[0093] Among them, the learning rules of synaptic plasticity are used to learn the following: The formula used for unsupervised training of the connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons is as follows:

[0094]

[0095] ;

[0096] ;

[0097]

[0098] In the formula, For having The first hidden layer of the randomly configured network of neurons. The input layer neuron and the first The connection weights between hidden layer neurons For having The first hidden layer of the randomly configured network of neurons. The input layer neuron and the first The connection weight adjustment term between hidden layer neurons , , , This is the learning rate parameter (with a value range of 0 to 0.01). For having The first hidden layer of the randomly configured network of neurons. The output value of each input layer neuron. For having The first hidden layer of the randomly configured network of neurons. The output value of each hidden layer neuron For having The first hidden layer of the randomly configured network of neurons. The dynamic threshold of each hidden layer neuron (this parameter is determined by calculation). To build a randomly configured network up to the th When the first hidden layer neuron is... The output value of each hidden layer neuron , The length of the historical response window. For the first The number of historical response states of each hidden layer neuron;

[0099] Calculate using the least squares method with The connection weights between the hidden and output layers of a randomly configured network with hidden layer neurons are calculated using the following formula:

[0100] ;

[0101] ;

[0102] In the formula, For calculation using the least squares method, with The connection weights between the hidden and output layers of a network are determined by the randomized configuration of neurons in each hidden layer. For having The output matrix of the hidden layer of a network with randomly configured hidden layer neurons (from the training sample set). Total For each sample, there are samples, with In a network with randomly configured hidden layer neurons, each hidden layer will output a corresponding output vector. The output vectors together constitute the output matrix. This is the activation function (here, the Sigmoid function). For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons;

[0103] (6) Determine if the termination condition has been met. If yes, stop and use the randomized network from the last iteration as the prediction model; otherwise, let... Then, return to step (4);

[0104] The termination condition is the root mean square error of the randomly configured network. Less than the allowable output error or ;

[0105] Among them, the root mean square error of the randomly configured network Less than the allowable output error , The calculation formula is as follows:

[0106] ;

[0107] in, Representative input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons is the first... The output vector of the th ... Each output value (training sample set) Total For each sample, there are samples, with Each randomly configured hidden layer neuron in the network will output a corresponding output vector. The output vectors together constitute the output matrix. Representative training label set The Middle The first label Each encoded value, , ;

[0108] (7) Test sample set The input is fed into the prediction model to obtain the output matrix. The calculation formula is as follows:

[0109] ;

[0110] ;

[0111] In the formula, For the input test sample set Output matrix of the output layer of the post-prediction model (test sample set) Total For each test sample, the output layer of the prediction model will output a corresponding output vector. The output vectors together constitute the output matrix of the output layer. For the input test sample set Output matrix of the hidden layer of the post-prediction model (test sample set) Total For each test sample, the hidden layer of the prediction model outputs a corresponding output vector. The output vectors together constitute the output matrix of the output layer. These are the connection weights between the hidden and output layers of the prediction model;

[0112] (8) Obtain the output matrix With test tag set The same row vectors are compared one by one. When they are the same row vector, if the output matrix... The column number containing the maximum output value of the row vector and the test label set If the number of columns containing elements with a code value of 1 in a row vector is the same, then the test sample corresponding to that row vector is considered correctly classified; otherwise, it is considered incorrectly classified. After all test samples have been compared, the number of correctly classified test samples is counted. Calculate the classification accuracy according to the formula. The calculation formula is as follows:

[0113] ×100%;

[0114] In the formula, This represents the number of test samples that were correctly classified. This represents the number of test samples;

[0115] (9) Test sample set with unknown labels (Feature space is the same as the training sample set) Perform the above missing value processing. After standardization, the data is input into the constructed and trained prediction model to obtain the output matrix of implicit heart disease categories. For the output matrix The analysis yields the predicted heart disease category, such as the output matrix. If the column number containing the maximum output value of the first row vector is 2, then the test sample set... The first sample in the dataset has a heart condition classified as class 2. The output matrix is... If the column number of the maximum output value of the 5th row vector is 4, then the test sample set... The fifth sample in the study had a heart condition classified as category 4.

[0116] To verify the accuracy of the above-mentioned cardiac auxiliary diagnostic classification method based on synaptic plasticity stochastic allocation network, it was tested against existing technologies (logistic regression (LR), support vector machine (SVM), stochastic vector function connection network (RVFL), and stochastic allocation network (SCN)) on the same cardiac data. The specific process is as follows:

[0117] Cardiac data: The dataset contains 303 samples, each with 13 clinical features, including age, sex, type of chest pain, resting blood pressure, cholesterol, fasting blood glucose, electrocardiogram results, maximum heart rate, whether exercise-induced angina, ST depression, ST slope, number of vessels, and thalassemia type.

[0118] Example 1: The above-mentioned classification method for cardiac auxiliary diagnosis based on synaptic plasticity randomized configuration network is used for classification. The process is as follows:

[0119] In step (1.1), , ; It is 302. It is 13. It is 2.

[0120] In step (1.5), the training sample set There are 242 samples in the test sample set. There are 61 samples in the training label set. There are 242 tags, test tag set There are 61 tags;

[0121] In step (2), the maximum number of neurons in the hidden layer The allowable output error is 10. The maximum number of candidate hidden layer neurons generated per round is 0.01. The set of weight distribution parameters between the input layer and the hidden layer is 100. Where 0.5 and 4.0 are the upper and lower boundaries of the weight distribution, and the change is... It is 0.5;

[0122] In step (4), the constant r is 0.001;

[0123] In step (5), the learning rate parameter The historical response window length is 0.002. It is 5;

[0124] The termination condition in step (6) is: the network is randomly configured. ;

[0125] In step (7), the test sample set The data was input into the prediction model, and the classification result was obtained with a classification accuracy of 86.07%.

[0126] Comparison Method 1: The Support Vector Machine (SVM) method was used to classify the above heart disease data. This method was implemented using the SVM model in the Python machine learning library sklearn, and the classification accuracy was used as the performance evaluation index. The classification accuracy obtained was 83.61%.

[0127] Comparison Method 2: The Random Vector Functional Link Network (RVFL) method in Reference 1 (Learning and generalization characteristics of the random vector Functional-link net[J].Neurocomputing, 1994, 6(2):163-180.) was used to classify the above heart disease data. This method uses the RVFL model implemented in Python programming language and uses classification accuracy as the performance evaluation index. The obtained classification accuracy is 83.03%.

[0128] Comparison Method 3: The traditional stochastic configuration network (SCN) method in Reference 2 (Stochastic Configuration Networks: Fundamentals and Algorithms[J].IEEE Transactions on Cybernetics, 2017, 47(10):3466-3479.) was used to classify the above heart disease data. This method uses an SCN model implemented in Python programming language and uses classification accuracy as the performance evaluation index. The classification accuracy obtained is 83.61%.

[0129] The cardiac auxiliary diagnostic classification method based on synaptic plasticity randomized configuration network in Example 1 was compared with Comparison Method 1, Comparison Method 2, and Comparison Method 3. The results are as follows: Figure 2 As shown, from Figure 2 As can be seen, the classification accuracy of the cardiac auxiliary diagnosis classification method based on synaptic plasticity randomized configuration network in Example 1 is 86.07%, which is higher than that of support vector machine, randomized vector function connection network and traditional randomized configuration network.

[0130] Among them, compared with comparison method three (i.e., traditional randomized network), the classification accuracy of the synaptic plasticity randomized network in Example 1 is improved by 2.46 percentage points; compared with comparison method two (i.e., randomized vector function connection network), the classification accuracy of the synaptic plasticity randomized network in Example 1 is improved by 3.04 percentage points; compared with comparison method one (i.e., support vector machine method), the classification accuracy of the synaptic plasticity randomized network in Example 1 is improved by 2.46 percentage points. The present invention has the highest classification accuracy compared with all existing classification methods.

[0131] The above results show that, during the construction of a randomly configured network, dynamically adjusting the connection weights between the input layer and the hidden layer through synaptic plasticity learning rules can enhance the weight learning ability and feature representation ability of the randomly configured network, thereby improving the classification accuracy of the randomly configured network in the task of assisting diagnosis and classification of heart disease.

Claims

1. A method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks, characterized in that, After missing value processing, Standardized test sample set with unknown labels The input is fed into the constructed and trained prediction model to obtain the output matrix of implicit heart disease categories. For the output matrix The analysis yields the predicted type of heart disease; The construction and training process of the prediction model is as follows: After initializing the parameters of the randomized network, hidden layer neurons are added one by one. During this process, candidate hidden layer neurons are randomly generated each time. The optimal candidate hidden layer neurons are selected and added to the randomized network through the supervision mechanism of the randomized network. Then, the connection weights between the input layer and the hidden layer of the randomized network are trained unsupervised through the synaptic plasticity learning rule. The connection weights between the hidden layer and the output layer of the randomized network are calculated using the least squares method. The above process is repeated until the termination condition is reached. The parameters for randomly configuring a network include: the maximum number of neurons in the hidden layer. Allowable output error The maximum number of candidate hidden layer neurons generated in each round The set of weight distribution parameters between the input layer and the hidden layer ,in , These are the upper and lower boundaries of the weight distribution. The change; The steps for building and training the prediction model are as follows: (a) Constructing the training sample set Training label set Test sample set and test tag set ; Test sample set with unknown labels Feature space and training sample set ; (b) Initialize the randomly configured network parameters; (c) Let ; (d) Randomly generate the first The candidate hidden layer neurons of each hidden layer are selected by a supervision mechanism of a randomized network, and the optimal candidate hidden layer neuron is added to the randomized network. (e) Learning rules based on synaptic plasticity for individuals with The connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons are trained unsupervised, and the least squares method is used to calculate the connection weights. The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons; Learning rules based on synaptic plasticity The formula used for unsupervised training of the connection weights between the input and hidden layers of a network with randomly configured hidden layer neurons is as follows: ； ； In the formula, For having The first hidden layer of the randomly configured network of neurons. The input layer neuron and the first The connection weights between hidden layer neurons For having The first hidden layer of the randomly configured network of neurons. The input layer neuron and the first The connection weight adjustment term between hidden layer neurons , , , The learning rate parameter, For having The first hidden layer of the randomly configured network of neurons. The output value of each input layer neuron. For having The first hidden layer of the randomly configured network of neurons. The output value of each hidden layer neuron For having The first hidden layer of the randomly configured network of neurons. The dynamic threshold of each hidden layer neuron To build a randomly configured network up to the th When the first hidden layer neuron is... The output value of each hidden layer neuron , The length of the historical response window. For the first The number of historical response states of each hidden layer neuron; (f) Determine if the termination condition has been met. If yes, stop and use the randomly configured network from the last iteration as the prediction model; otherwise, let... Then, return to step (d); The termination condition is the root mean square error of the randomly configured network. Less than the allowable output error or .

2. The method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks according to claim 1, characterized in that, Step (a) is as follows: (a1) Obtain cardiac data, which includes Each sample and its corresponding label will The sample set consists of 10 samples. , For the first The first sample One characteristic variable, , , This represents the total number of feature variables in each sample; Will The labels corresponding to each sample form a label set. , For the first The label corresponding to each sample , The total number of categories of heart disease; (a2) For the sample set Missing values ​​are handled by filling in the missing values ​​of a feature variable with missing values ​​using the mean of that feature variable across all non-missing samples. (a3) For the sample set conduct Standardization processing, the formula is as follows: ； In the formula, for After standardization The first sample One characteristic variable, for The first sample The mean of each characteristic variable, for The first sample The standard deviation of each characteristic variable; (a4) For the tag set Perform one-hot encoding; the tag set after one-hot encoding. , For the first The label corresponding to the nth sample Each encoded value, , ; (a5) Proportionately divide the sample set Divided into training sample set and test sample set At the same time, the tag set Divided into training label sets and test tag set .

3. The method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks according to claim 2, characterized in that, Step (d) is as follows: (d1) The set of weight distribution parameters between the input layer and the hidden layer Randomly generate connection weights between the input layer and the hidden layer. Bias between input layer and hidden layer To generate Candidate hidden layer neurons , For the first One candidate hidden layer neuron, For input layer and Connection weights between them For input layer and The offset between them ; (d2) The above Substitute the candidate hidden layer neurons sequentially into the following formula: ； ； = ； ； ； in, for The judgment value, for and the The decision value corresponding to each output layer neuron. , , For having The first hidden layer of the randomly configured network of neurons. The output residual vector of each output layer neuron For the input training sample set After having A network with randomly configured hidden layer neurons The output vector, This is an inner product operation, where r is a constant. The sequence is a non-negative real number sequence, which increases with the number of neurons in the hidden layer. The increase gradually decreases and approaches 0. For activation function, For training label set The One encoded vector, , For the input training sample set After having The first hidden layer of the randomly configured network of neurons. The output vector of each output layer neuron For the input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons. For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons. For having The connection weights between the hidden and output layers of a network with randomly configured hidden layer neurons; (d3) when satisfying Select from candidate hidden layer neurons The candidate hidden layer neuron with the largest value is added to the layer with... In a network with randomly configured neurons in the nth hidden layer, as the th One hidden layer of neurons.

4. The method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks according to claim 1, characterized in that, In step (e), the least squares method is used to calculate the... The connection weights between the hidden and output layers of a randomly configured network with hidden layer neurons are calculated using the following formula: ； ； In the formula, For calculation using the least squares method, with The connection weights between the hidden and output layers of a network are determined by the randomized configuration of neurons in each hidden layer. For having The output matrix of the hidden layer of a network with randomly configured hidden layer neurons. For activation function, For having The connection weights between the input and hidden layers of a network are randomly configured with neurons in each hidden layer. For having The bias between the input and hidden layers of a network with randomly configured hidden layer neurons.

5. The method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks according to claim 1, characterized in that, In step (f), the root mean square error of the network is randomly configured. Less than the allowable output error , The calculation formula is as follows: ； in, Representative input training sample set After having The output matrix of the output layer of a network with randomly configured hidden layer neurons is the first... The output vector of the th ... One output value, Representative training label set The Middle The first label Each encoded value, , .

6. The method for auxiliary diagnosis and classification of heart disease based on synaptic plasticity stochastic configuration networks according to claim 1, characterized in that, It also includes step (g): preparing the test sample set The data is input into the prediction model to obtain the classification results. The classification accuracy is then calculated using the following formula: ； In the formula, This represents the number of test samples that were correctly classified. This represents the number of test samples.

Images