A medical monitoring data expansion method and auxiliary classification method
By using adaptive feature manifold projection to generate adversarial networks, adaptive wolf pack optimization algorithms, mask-enhanced self-coding networks, and adaptive memory network architectures in medical monitoring data analysis, data acquisition and labeling costs are solved The problems of high, low feature extraction and dimensionality reduction efficiency, and poor classifier adaptability have achieved efficient medical monitoring data analysis and auxiliary diagnosis.
Patent Information
- Application Number
- CN202411163827.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-23
AI Technical Summary
The prior art relies on manual operations in the collection and labeling of medical monitoring data, which is time-consuming and costly, and the sample size is limited. Traditional neural networks are prone to gradient problems and local optimization in feature extraction, and autoencoders are characterized by feature reduction. Dimensional time may lead to information loss, classifiers are prone to overfitting when processing nonlinear and complex data, and lack the ability to quickly adapt to changes in new data.
Data expansion is adopted for data expansion based on adaptive feature manifold projection, combined with adaptive wolf pack optimization algorithm and mask-enhanced autoencoded neural network for feature extraction and dimensionality reduction, and the limit learning machine algorithm of the adaptive memory network architecture is used. Classification, dynamically adjust the loss function and weights to enhance the diversity of data and the generalization ability of the model.
It significantly improves the efficiency of medical monitoring data analysis and the accuracy of auxiliary diagnosis, alleviates the problem of insufficient sample size, improves the accuracy of feature extraction and the effectiveness of feature dimensionality reduction, reduces the risk of overfitting, and enhances the classifier adaptability and robustness.
Smart Images

Figure CN118690202B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a medical monitoring data expansion method and an auxiliary classification method, and belongs to the technical field of data processing. Background Art
[0002] Medical auxiliary diagnosis is receiving more and more attention as a key means to improve the accuracy and efficiency of disease diagnosis. With the advancement of medical technology, especially the development of wearable devices and remote monitoring technology, doctors can obtain a large amount of physiological and biochemical data from patients. These data are of great value for early detection of diseases, treatment monitoring, and personalized medicine.
[0003] However, existing technologies have some limitations when processing and analyzing these data. First, the collection and labeling of medical data usually requires professionals. This process is not only time-consuming and labor-intensive, but also limited by the number and experience of professionals, resulting in high data collection and labeling costs and limited sample size. Secondly, medical data usually has high dimensions and complex intrinsic structures. Traditional data processing methods are difficult to effectively capture and express the characteristics of these data, which can easily lead to information loss and inaccurate diagnosis. In addition, existing classifiers often lack sufficient adaptability and generalization capabilities when processing highly nonlinear and complex data, and are prone to overfitting, which affects the reliability of diagnostic results.
[0004] The Chinese invention patent with application number CN202410483624.7 proposes a multimodal medical imaging data analysis method based on machine learning, which includes the following steps: MRI data collection and preprocessing, MRI feature extraction, clinical diagnosis model of trigeminal neuralgia, and model result analysis. The method of the present invention uses machine learning to extract features from medical images and establish a clinical diagnosis model, demonstrating its potential application value in assisting the diagnosis of complex diseases.
[0005] The Chinese invention patent with application number CN202111014602.9 proposes an auxiliary analysis method, device, computer equipment and storage medium for medical image pictures, the method comprising: identifying the medical image pictures uploaded by the medical terminal, obtaining the location information of the medical image pictures that meets the pathological characteristics; outlining the lesion on the medical image pictures according to the location information, and obtaining a lesion outlining information picture including the lesion area; outputting the medical image pictures and the lesion outlining information pictures and displaying them on the display screen, so as to generate an edit box at the edge of the lesion area of the lesion outlining information picture on the display screen; upon receiving the touch modification operation of the medical staff on the display screen, generating a lesion box according to the touch modification operation and the edit box; generating a case picture according to the lesion box, and generating a case report by associating the electronic signature of the medical staff. This application can improve the efficiency of auxiliary analysis of medical image pictures.
[0006] The Chinese invention patent with application number CN202211702406.5 proposes a pathological analysis method, device, equipment and storage medium based on machine learning, wherein the method includes: obtaining the patient's medical physiological data; inputting the medical physiological data into a pre-trained disease prediction model for prediction to obtain a disease prediction result, the disease prediction model includes an encoder, a decoder and a discriminator, the encoder encodes the medical physiological data to obtain content features representing the patient's basic information and style features representing the patient's disease information, the decoder fuses and reconstructs the content features and style features, and the discriminator predicts the disease prediction result based on the reconstruction results. The present invention uses the disease prediction model to embed the medical physiological data into the content latent space representing the patient's basic information and the style latent space representing the patient's disease information, and then fuses and reconstructs the two to perform disease prediction, which greatly improves the decoupling, thereby improving the interpretability and having higher generalization.
[0007] In the process of implementing the embodiments of the present disclosure, it is found that there are at least the following problems in the related art:
[0008] Existing technologies rely on manual operations in the collection and labeling of medical monitoring data, which is time-consuming and costly, and the sample size is limited, affecting the quality and efficiency of model training. Traditional neural networks are prone to gradient problems and local optimality during feature extraction, resulting in inaccurate feature expression; traditional autoencoders may not be able to effectively capture the implicit structure of the data during feature dimensionality reduction, resulting in information loss after data compression. Existing classifiers may have overfitting problems when processing nonlinear and complex data, and lack the ability to quickly adapt to new data changes, affecting the accuracy and reliability of classification results. Summary of the invention
[0009] The purpose of the present invention is to provide a medical monitoring data expansion method and an auxiliary classification method, which significantly improves the efficiency of medical monitoring data analysis and the accuracy of auxiliary diagnosis.
[0010] In order to achieve the above object, the present invention is implemented through the following technical solutions:
[0011] The method comprises:
[0012] Collect real physiological and biochemical data;
[0013] Establishing a generative adversarial network based on adaptive feature manifold projection, wherein the generative adversarial network based on adaptive feature manifold projection includes a generator and a discriminator, and enhances the generated data through a dynamic loss function adjustment mechanism;
[0014] The collected physiological and biochemical data are used to train a generative adversarial network based on adaptive feature manifold projection, and the dataset is expanded by the trained generative adversarial network based on adaptive feature manifold projection.
[0015] Preferably, the training process of the generative adversarial network based on adaptive feature manifold projection is as follows:
[0016] Initialize the network parameters of the generator and discriminator, and define the weights of the feature manifold projection module of the generator;
[0017] Pre-train the discriminator and use real datasets to optimize its parameters;
[0018] In the adversarial training phase, the data generated by the generator is adjusted through the feature manifold projection module;
[0019] The generated data is fed into the discriminator, and the loss functions of the generator and the discriminator are calculated at the same time, and the loss functions of the generator and the discriminator are adjusted according to the feature manifold projection error;
[0020] The discriminator is optimized based on the loss function of synthetic data and real data, and the parameters of the discriminator are updated;
[0021] After each round of training, the weight and structure of the loss function are dynamically adjusted according to the similarity of the feature manifold between the generated data and the real data;
[0022] Enhance the generated data using an adaptive feature weight adjustment mechanism;
[0023] Repeat the above steps until the preset stop iteration condition is met.
[0024] Preferably, the specific formula for adjusting the data generated by the generator through the feature manifold projection module is as follows:
[0025] ,
[0026] ,
[0027] ,
[0028] ,
[0029] In the formula, is random noise, The current batch synthetic data, For the generator parameter, is the characteristic manifold projection function, For the The generated data, is the weight of the feature manifold projection module; is the bias of the feature manifold projection module, is the LeakyReLU nonlinear transformation function.
[0030] Preferably, the loss functions of the generator and the discriminator are adjusted according to the feature manifold projection error as follows:
[0031] ,
[0032] ,
[0033] In the formula, represents the loss function of the generator, are the discriminator parameters, represents the loss function of the discriminator, is the regularization parameter, is the L2 norm;
[0034] Dynamic Adjustment The calculation method is as follows:
[0035] ,
[0036] In the formula, and They are vectors and No. elements.
[0037] Preferably, the weight and structure of the loss function are dynamically adjusted according to the feature manifold similarity between the generated data and the real data, which is specifically achieved by dynamically adjusting the regularization parameter based on the feature manifold similarity, and the formula is as follows:
[0038] ,
[0039] ,
[0040] ,
[0041] In the formula, To dynamically adjust the factor, its update rule is as follows:
[0042] ,
[0043] In the formula, is the learning rate parameter, is the dynamic adjustment factor before updating, is the updated dynamic adjustment factor, Indicates The feature manifold similarity of the samples.
[0044] Preferably, the feature weight vector is updated by dynamic adjustment, and the discriminant generates data according to the discriminant result. , if a feature has a high error rate in the discriminator, then increase the weight of the feature; otherwise, reduce it; feature weight vector The adjustment method is expressed as:
[0045] ,
[0046] In the formula, is the weight vector The weight of the feature, is the weight vector before adjustment. The weight of the feature, is the first The weight of each feature; is the learning rate parameter; Denotes the discriminator loss function About The partial derivative of a feature.
[0047] An artificial intelligence-based medical monitoring data auxiliary classification method, comprising:
[0048] Collect physiological and biochemical data and annotate the collected data;
[0049] Expanding the data set by using the medical monitoring data expansion method;
[0050] A medical monitoring data analysis and auxiliary classification model is established, including a feature extraction module, a feature dimension reduction module and a classifier module. The expanded data set is processed by the feature extraction module, the feature dimension reduction module and the classifier module in sequence; the feature extraction module is a neural network based on an adaptive wolf pack optimization algorithm, the feature dimension reduction module is an autoencoder neural network based on mask enhancement, and the classifier module is an extreme learning machine algorithm based on an adaptive memory network architecture;
[0051] The expanded data set is used to train the medical monitoring data analysis and auxiliary diagnosis model, and the collected physiological and biochemical data to be classified are input into the trained medical monitoring data analysis and auxiliary classification model for classification.
[0052] Preferably, the neural network training process based on the adaptive wolf pack optimization algorithm is as follows:
[0053] Initializing parameters of a neural network, wherein the parameters of the neural network include weights and biases;
[0054] Based on the adaptive wolf pack optimization algorithm, the parameters of the neural network are adjusted by simulating the social behavior of the wolf pack. Each wolf represents a set of potential solutions of the neural network.
[0055] In each algorithm iteration, the search strategy is adjusted according to the optimization results of the previous generation, and the weights and biases of the neural network are updated;
[0056] Repeat the above steps until the preset stop iteration condition is met.
[0057] Preferably, the training process of the autoencoder neural network based on mask enhancement is as follows:
[0058] Initialize the parameters of the autoencoder network, including the weights and bias parameters of the autoencoder network,
[0059] According to the framework of the greedy algorithm, set the independent training cycle and learning rate for each layer;
[0060] At the beginning of each training cycle, for each input vector, a mask matrix is randomly generated;
[0061] The input data is sent to the encoder after masking, and the encoder compresses the data into a low-dimensional feature representation after processing through an activation function;
[0062] The low-dimensional features are fed into the decoder, which attempts to reconstruct the original input data. By calculating the loss function, the back propagation algorithm updates the parameters in the network.
[0063] The greedy algorithm is used to optimize the parameters of each layer, and the Adam optimizer is used to update the parameters;
[0064] After all layers are trained, the models of each layer are integrated to form a complete autoencoder model for overall fine-tuning.
[0065] Repeat the above steps until the preset stop iteration condition is met.
[0066] Preferably, the training process of the extreme learning machine algorithm based on the adaptive memory network architecture is as follows:
[0067] Initialize the memory unit, transform the input data into features through the adaptive memory network, and update the memory state based on the relationship between the current data and the historical data;
[0068] The data after feature conversion is subjected to sparse processing of the weight matrix, and the sparsely processed feature data is input into the extreme learning machine to train the extreme learning machine;
[0069] Automatically adjust regularization parameters of memory networks and extreme learning machines based on performance during training;
[0070] Repeat the above steps until the preset stop iteration condition is met.
[0071] The advantages of the present invention are that the present invention solves the deficiencies of the prior art in the field of medical monitoring data analysis and auxiliary diagnosis through innovative technical means. First, in view of the limitations of data collection and annotation in the prior art, the present invention adopts a generative adversarial network based on adaptive feature manifold projection, expands the data set by generating high-quality synthetic data, effectively alleviates the problem of insufficient real sample size, and improves the diversity of data and the generalization ability of the model. Secondly, in view of the challenges in the process of feature extraction and dimensionality reduction, the present invention adopts a neural network based on an adaptive wolf pack optimization algorithm and an autoencoder neural network based on mask enhancement, and improves the accuracy of feature extraction and the effectiveness of feature dimensionality reduction by optimizing network parameters and using random mask enhancement models to learn the implicit structure of data. Finally, in view of the adaptability and generalization ability of the classifier, the present invention adopts an extreme learning machine algorithm based on an adaptive memory network architecture, and enhances the rapid adaptability of the classifier to new data changes by dynamically adjusting the memory state and weights, reducing the risk of overfitting, and improving the accuracy and robustness of the classification results. Through the comprehensive application of these technical means, the present invention significantly improves the efficiency of medical monitoring data analysis and the accuracy of auxiliary diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.
[0073] Figure 1 Schematic diagram of the flow of the medical monitoring data expansion method of the present invention.
[0074] Figure 2 Schematic diagram of the flow chart of the medical monitoring data auxiliary classification method of the present invention.
[0075] Figure 3 The figure is a schematic diagram of the training process of the neural network based on the adaptive wolf pack optimization algorithm of the present invention.
[0076] Figure 4 The training process of the mask-enhanced autoencoder neural network of the present invention is as follows DETAILED DESCRIPTION
[0077] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0078] Example 1
[0079] In order to solve the problem of high cost and insufficient sample size of medical monitoring data collection, this embodiment provides a medical monitoring data expansion method based on artificial intelligence, which adopts a generative adversarial network based on adaptive feature manifold projection, generates high-quality synthetic data through a generator, and distinguishes the authenticity of data through a discriminator, so as to expand the data volume and improve the quality. Figure 1 As shown, it is achieved through the following technical solutions:
[0080] S1: Collect real physiological and biochemical data.
[0081] S2: Establish a generative adversarial network based on adaptive feature manifold projection, wherein the generative adversarial network based on adaptive feature manifold projection includes a generator and a discriminator, and enhances the generated data through a dynamic loss function adjustment mechanism.
[0082] S3: Use the collected physiological and biochemical data to train a generative adversarial network based on adaptive feature manifold projection, and expand the dataset through the trained generative adversarial network based on adaptive feature manifold projection.
[0083] As a refinement of the above step S1, the data set of this embodiment comes from monitoring equipment and health management application systems in medical institutions, including electrocardiogram systems, sphygmomanometers, blood glucose meters, etc. The collected data is stored in a structured JSON format. In one embodiment, the collected data attributes include: heart rate (Ra1), blood pressure (Ra2), blood glucose level (Ra3), body temperature (Ra4), respiratory rate (Ra5), blood oxygen saturation (Ra6), weight (Ra7), pulse (Ra8), drug reaction (Ra9), and pain index (Ra10).
[0084] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.
[0085] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include: normal, warning, early disease, mid-disease, and late disease, a total of 5 categories.
[0086] As a refinement of the above step S2, it can be understood that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a generative adversarial network based on adaptive feature manifold projection to generate samples, thereby achieving data expansion. The generative adversarial network based on adaptive feature manifold projection consists of two parts: a generator (G) and a discriminator (D). The purpose of the generator is to learn the distribution of real data and generate high-quality synthetic data, while the discriminator attempts to distinguish between generated data and real data. In traditional generative adversarial networks, the game between the generator and the discriminator usually relies on a static loss function. The present invention adopts a dynamic loss function adjustment mechanism, that is, a loss function based on a momentum coordination mechanism, and dynamically adjusts the loss function according to the difference in the feature manifold between the generated data and the real data, prompting the generator to more accurately capture complex data features and internal structures.
[0087] The training process of the generative adversarial network based on adaptive feature manifold projection is as follows:
[0088] S201: Initialize the generator and the discriminator The network parameters of the generator are as follows: and , the initial weights and biases of the discriminator are and , and define the weights of the feature manifold projection module of the generator .
[0089] In one embodiment, the generator and the discriminator The weight and bias parameters are initialized randomly.
[0090] S202: Pre-train the discriminator and use real data sets to optimize its parameters so that the discriminator can have better data recognition ability in the early stage. The objective function for optimizing the parameters of the discriminator is:
[0091] ,
[0092] In the formula, Representation Discriminator Parameters, For the discriminator The parameters determined by pre-training, is the number of real samples selected in the pre-training phase, For the The real data in the pre-training dataset, is the cross entropy loss function.
[0093] S203: Adversarial training phase. At the beginning of each training cycle, the generator Generate a batch of synthetic data based on the current parameters The data generated by the generator is adjusted through the feature manifold projection module. The specific formula is as follows:
[0094] ,
[0095] ,
[0096] In the formula, is random noise, The current batch synthetic data, For the generator parameter, is the characteristic manifold projection function, For the The generated data, is the weight of the feature manifold projection module, The number of synthetic data in the current batch.
[0097] In the generator, the calculation method of the feature manifold projection function is expressed as:
[0098] ,
[0099] ,
[0100] In the formula, is the bias of the feature manifold projection module, is the LeakyReLU nonlinear transformation function.
[0101] S204: The generated data is fed into the discriminator, and the loss functions of the generator and the discriminator are calculated at the same time, and the generator loss is adjusted according to the manifold projection error. The loss functions of the generator and the discriminator are as follows:
[0102] ,
[0103] ,
[0104] In the formula, represents the loss function of the generator, are the discriminator parameters, represents the loss function of the discriminator, is the regularization parameter, is the L2 norm.
[0105] Dynamic Adjustment The calculation method is as follows:
[0106] ,
[0107] In the formula, and They are vectors and No. elements.
[0108] S205: The discriminator is optimized according to the loss function of the synthetic data and the real data, and the parameters of the discriminator are updated. The formula is as follows:
[0109] ,
[0110] In the formula, is the learning rate, are the discriminator parameters, is the loss function About the discriminator parameters The gradient of The symbol for the update operation.
[0111] S206: After each round of training, dynamically adjust the weight and structure of the loss function according to the feature manifold similarity between the generated data and the real data to ensure that the generator can be optimized in the right direction. Specifically, the present invention is based on the feature manifold similarity. Dynamically adjust the regularization parameter, the adjustment method is expressed as:
[0112] ,
[0113] ,
[0114] ,
[0115] In the formula, To dynamically adjust the factor, its update rule is as follows:
[0116] ,
[0117] In the formula, is the learning rate parameter, is the dynamic adjustment factor before updating, is the updated dynamic adjustment factor, Indicates The feature manifold similarity of samples. Preferably, Set to 0.1.
[0118] S207: Enhance the generated data using an adaptive feature weight adjustment mechanism. Specifically, define a feature weight vector , each element of the vector corresponds to a feature in the generated data, and different weights are assigned to each feature. These weights are dynamically adjusted according to the feedback of the discriminator. The way to generate data enhancement is expressed as:
[0119] ,
[0120] In the formula, Indicates the enhanced Generate data, is the feature weight vector, Represents element-wise multiplication of vectors.
[0121] The feature weight vector is updated by dynamic adjustment, and the discriminant generates data according to the discriminant result. , if a feature has a high error rate in the discriminator, then increase the weight of the feature; otherwise, reduce it; feature weight vector The adjustment method is expressed as:
[0122] ,
[0123] In the formula, is the weight vector The weight of the feature, is the weight vector before adjustment. The weight of the feature, is the first The weight of each feature; is the learning rate parameter; Denotes the discriminator loss function About The partial derivative of the features. Preferably, is 0.01.
[0124] S208: Repeat the above steps until a preset condition for stopping iteration is met. In one embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.
[0125] It should be noted that the dynamic loss function adjustment mechanism effectively improves the quality and diversity of synthetic data, enhances the generalization ability of the model, and solves the problem caused by insufficient sample size.
[0126] Example 2
[0127] like Figure 2As shown, an artificial intelligence-based auxiliary classification method for medical monitoring data is used to solve the gradient vanishing, gradient exploding or local optimal problems that traditional neural networks may encounter in feature extraction. The present invention uses a neural network based on an adaptive wolf pack optimization algorithm to optimize network parameters by simulating the social behavior of wolf packs, thereby improving the efficiency and accuracy of feature extraction. In view of the complexity of high-dimensional data in storage and processing, the present invention uses a mask-enhanced autoencoder neural network to perform feature dimensionality reduction. Through random masks and a greedy algorithm framework, the model's ability to learn the implicit structure of the data is enhanced to achieve effective data compression. In order to cope with the classification difficulties caused by data nonlinearity and complexity, the present invention uses an extreme learning machine algorithm based on an adaptive memory network architecture to improve the classifier's responsiveness to new data and reduce the risk of overfitting through memory units and adaptive weight adjustments. Specifically including:
[0128] S1: Collect physiological and biochemical data and annotate the collected data.
[0129] S2: Establish a generative adversarial network based on adaptive feature manifold projection, wherein the generative adversarial network based on adaptive feature manifold projection includes a generator and a discriminator, and enhances the generated data through a dynamic loss function adjustment mechanism. Use the collected physiological and biochemical data to train the generative adversarial network based on adaptive feature manifold projection, and expand the data set through the trained generative adversarial network based on adaptive feature manifold projection.
[0130] S3: Establish a medical monitoring data analysis and auxiliary classification model, including a feature extraction module, a feature dimension reduction module and a classifier module. The expanded data set is processed by the feature extraction module, the feature dimension reduction module and the classifier module in sequence; the feature extraction module is a neural network based on an adaptive wolf pack optimization algorithm, the feature dimension reduction module is an autoencoder neural network based on mask enhancement, and the classifier module is an extreme learning machine algorithm based on an adaptive memory network architecture.
[0131] S4: Use the expanded data set to train the medical monitoring data analysis and auxiliary diagnosis model, and input the collected physiological and biochemical data to be classified into the trained medical monitoring data analysis and auxiliary classification model for classification.
[0132] As a refinement of the above embodiment, the expanded data is first input into the feature extraction model for training the feature extraction model. This embodiment uses a 3-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion, or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. This embodiment uses a neural network based on an adaptive wolf pack optimization algorithm. On the basis of the traditional wolf pack optimization algorithm, an adaptive mechanism is used to adjust the search strategy according to the previous optimization results to improve the algorithm's search efficiency and accuracy in high-dimensional data space.
[0133] Specific as Figure 3 As shown in the figure, the neural network training process based on the adaptive wolf pack optimization algorithm is as follows:
[0134] S3011: Initialize the parameters of the neural network, the parameters of the neural network include weights and biases. In one embodiment, the initialization method is expressed as:
[0135] ,
[0136] In the formula, represents the initial value of the neural network parameters, is the initialization standard deviation, represents a Gaussian distribution with mean 0 and covariance as the identity matrix, preferably, Set to 0.01.
[0137] S3012: Based on the adaptive wolf pack optimization algorithm, the parameters of the neural network are adjusted by simulating the social behavior of the wolf pack. Each wolf represents a set of potential solutions of the neural network, and the social behavior of the wolf pack helps to explore the solution space and find the optimal solution. The wolf pack position update method is expressed as:
[0138] ,
[0139] In the formula, and They are and The position of the wolf in the iteration, It is the best position for an individual in a wolf pack. is the global best position, and is a parameter that balances exploration and exploitation. Preferably, and Set them to 0.3 and 0.5 respectively.
[0140] S3013: In each algorithm iteration, the search strategy is adjusted according to the optimization results of the previous generation. The adaptive adjustment method is expressed as:
[0141] ,
[0142] In the formula, It is updated , is the adjustment factor used to adjust Size represents the Euclidean distance.
[0143] In one embodiment, the adjustment factor Adjustments are made based on feedback from historical optimization performance, and the calculation method is expressed as:
[0144] ,
[0145] In the formula, is the number of historical iterations considered, indicating how many generations of past performance affect the current adjustment; is the attenuation factor, is the decay rate, It is in The loss function value of the iteration; For the The neural network parameters of the iteration, that is, The optimal position of individuals in the wolf pack in iteration The corresponding neural network parameters. Preferably, Set to 3.
[0146] S3014: Update the weights and biases of the neural network. The update method is expressed as:
[0147] ,
[0148] In the formula, and They are The iteration and The neural network parameters of the iteration; is the learning rate; It is in the parameter The gradient of the loss function at ; preferably, Set to 0.01.
[0149] Furthermore, the gradient The calculation method is expressed as:
[0150] ,
[0151] In the formula, is the number of samples in each batch of training data; It is The input data of samples; It is The true labels of samples; is the cross entropy loss function; is the predicted output of the neural network model.
[0152] 3015: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.
[0153] As a refinement of the above embodiment, the data after feature extraction is input into the feature dimension reduction model for training the feature dimension reduction model. This embodiment adopts a mask-enhanced autoencoder neural network as a feature dimension reduction algorithm. The mask-enhanced autoencoder neural network includes an encoder and a decoder, and realizes data compression and reconstruction by learning the effective expression of the input data. On the basis of the traditional autoencoder, this embodiment uses a random mask in the input layer to selectively hide some features of the input data, forcing the network to reconstruct the complete output from the partially visible input data, thereby enhancing the model's ability to learn the implicit relationship in the data, so as to ensure that the data after dimensionality reduction can still maintain the original data structure characteristics.
[0154] Specific as Figure 4 As shown in the figure, the training process of the autoencoder neural network based on mask enhancement is as follows:
[0155] S3021: Initialize the parameters of the autoencoder network, including the weights and bias parameters of the autoencoder network. For the The set of parameters of the layer, including weights and biases, is ,and For the The weight matrix of the layer, For the The bias vector of the layer. In one embodiment, the initialization method is random initialization.
[0156] S3022: According to the framework of the greedy algorithm, set the independent training cycle and learning rate of each layer to ensure that each layer can reach the local optimum before joining the overall network. The number of independent training loop iterations for the layer Set to:
[0157] ,
[0158] In the formula, is the total number of iterations, is the total number of layers, is the Kronecker function, ensuring that the last layer can have more training cycles.
[0159] S3023: At the beginning of each training cycle, for each input vector , randomly generate a mask matrix , used to randomly mask part of the input data, and the way to generate the mask is expressed as:
[0160] ,
[0161] In the formula, is the mask matrix No. elements, is the mask ratio; is a Bernoulli distribution discrete function, which is used to generate a mask matrix, each element from An independent draw from the distribution indicates that the element has The probability of is set to 1 (i.e., selected or activated), The probability of is set to 0 (i.e. covered or inactive).
[0162] Furthermore, the mask ratio The calculation method is expressed as:
[0163] ,
[0164] In the formula, Represents input data The variance of is the preset variance threshold, Represents the minimum value function.
[0165] S3024: After the input data is masked, it is sent to the encoder. The encoder compresses the data into a low-dimensional feature representation after processing through an activation function. The forward propagation method of the encoder is expressed as:
[0166] ,
[0167] In the formula, is the encoded feature vector, is the Sigmoid activation function, and are the encoder weights and biases, Denotes the Hadamard product.
[0168] S3025: The low-dimensional features are fed into the decoder, which attempts to reconstruct the original input data. By calculating the loss function, the back propagation algorithm updates the parameters in the network. The output of the decoder is expressed as:
[0169] ,
[0170] In the formula, is the output of the decoder, that is, the reconstructed input data.
[0171] Furthermore, the calculation method of the loss function is expressed as:
[0172] ,
[0173] In the formula, is the total loss, is the reconstruction error, is the weight to control the isolated loss; is an isolated loss, indicating The number of non-zero elements of is used to measure the sparsity of the code; preferably, Set to 0.3.
[0174] Furthermore, isolated losses The calculation method is expressed as:
[0175] ,
[0176] In the formula, is a vector Dimensions; For the The encoded feature vector of dimension, is the indicator function, when The value is 1 when , otherwise it is 0.
[0177] S3026: Optimize the parameters of each layer using a greedy algorithm. In one embodiment, the parameter update method uses an Adam optimizer, and the update method is expressed as:
[0178] ,
[0179] In the formula, and Respectively represent Tier Iterations and The parameters of the iterations, is the Adam optimizer function, is the learning rate, is the calculated loss, For the Tier The gradient of the parameters of the iteration with respect to the loss function. Preferably, Set the dynamic adjustment strategy, and its initial value is set to 0.01.
[0180] In one embodiment, The learning rate for this iteration is , and its dynamic adjustment method is expressed as:
[0181] ,
[0182] In the formula, is the initial learning rate, is the initial reconstruction weight, For the The reconstruction weight of the iteration, is the adjustment influence coefficient. Preferably, Set to 3, Set to 5.
[0183] Further, The reconstruction weight of the iteration The adjustment of is based on the current effect of the reconstruction error and the sparsity weight, and is calculated as:
[0184] ,
[0185] In the formula, is the adjustment rate parameter, is the current reconstruction error, is the sparsity weight, For the Preferably, Set to 0.3, Set to 5.
[0186] Furthermore, the gradient adjustment method in the Adam optimizer is expressed as:
[0187] ,
[0188] In the formula, and They are The bias-corrected first- and second-order moment estimates for the iterations, is a small constant. Preferably, Set to 0.0001.
[0189] Further, The bias-corrected first- and second-order moment estimates for the iterations and The calculation method is:
[0190] ,
[0191] ,
[0192] In the formula, and It is The exponential moving average and squared exponential moving average of the gradients of the iterations, and It is Preferably, and Set to 0.9 and 0.999 respectively.
[0193] S3027: After the training of all layers is completed, the models of each layer are integrated to form a complete autoencoder model, and the overall fine-tuning is performed to optimize the global performance of the model. The fine-tuning method is expressed as:
[0194] ,
[0195] In the formula, is the fine-tuned parameter, is the number of training samples, It is training samples, Indicates that the parameter is and The loss function of .
[0196] S3028: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.
[0197] As a refinement of the above embodiment, the data after dimensionality reduction is input into the classifier model to train the classifier model. The present invention adopts an extreme learning machine algorithm based on an adaptive memory network architecture as the classification algorithm, and adopts memory units and an adaptive weight adjustment mechanism to enable the classifier to dynamically adjust its internal parameters according to changes in the input data, so as to better adapt to the nonlinearity and complexity of the data, while ensuring the classification performance and reducing the risk of overfitting of the model.
[0198] The training process of the extreme learning machine algorithm based on the adaptive memory network architecture is as follows:
[0199] S3031: Define memory unit as , the way to initialize the memory unit is expressed as:
[0200] ,
[0201] In the formula, is the initial state of the memory unit; Represents the feature dimension, that is, the feature dimension after dimensionality reduction; represents the number of memory cells, To generate Gaussian distributed random numbers.
[0202] S3032: In each training iteration, the input data first passes through the adaptive memory network. The network not only performs feature conversion, but also updates the memory state according to the relationship between the current data and the historical data, ensuring that the network can accumulate and utilize historical information and improve the response to new data. Specifically, the way in which the data is converted through the adaptive memory network is expressed as:
[0203] ,
[0204] In the formula, is the input feature data, is the feature data after preliminary processing. and are the first weight and bias of the memory network layer, is the activation function.
[0205] Furthermore, the way in which the adaptive memory network updates the memory state is expressed as:
[0206] ,
[0207] In the formula, For the current The memory state of the iteration, For the current The memory state of the iteration, is the updated memory state, is the memory decay factor, is the second weight of the memory network layer. Preferably, Set to 0.3.
[0208] S3033: Before the feature data is passed to the extreme learning machine, the weight matrix is sparsely processed. By removing less active connections, the network structure is optimized and the weights are sparsely processed, which is expressed as:
[0209] ,
[0210] In the formula, is the sparse weight matrix, is the sparseness function, is the weight threshold.
[0211] In one embodiment, the sparsification function It is calculated by the adaptive threshold adjustment method based on L1 regularization and is expressed as:
[0212] ,
[0213] In the formula, is a symbolic function. return The sign of each element in ; represents element-wise multiplication, is the maximum value function; is the scale factor, Calculated as The standard error of the elements in is used to adaptively adjust the threshold so that the sparsification operation can take into account the distribution characteristics of the weights. Preferably, Set to 0.3.
[0214] S3034: Input the preliminarily processed feature data into the extreme learning machine to train the extreme learning machine. The training process is expressed as:
[0215] ,
[0216] ,
[0217] In the formula, is the output of the hidden layer of the extreme learning machine, is the bias of the hidden layer of the Jixian County learning machine, is the weight of the output layer, is the target output, is the transpose of the output of the hidden layer of the extreme learning machine, is the regularization parameter, is the identity matrix, Representation Matrix The inverse of.
[0218] S3035: Automatically adjust the regularization parameters of the memory network and extreme learning machine based on the performance during training to ensure that the model can maintain optimal performance at different stages. The adaptive parameter adjustment method is expressed as:
[0219] ,
[0220] In the formula, is the adjusted regularization parameter, is the learning rate, Performance evaluation function, is the output of the model on the training set, is the performance evaluation function with respect to the parameter Preferably, Set to 0.3.
[0221] Furthermore, the gradient The calculation method is expressed as:
[0222] ,
[0223] In the formula, is the performance evaluation function Model Output The partial derivative of is the model output parameter Preferably, the performance evaluation function The root mean square loss function is used.
[0224] S3036: Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 800 times.
[0225] As a refinement of the above embodiment, the step S4 uses the trained medical monitoring data analysis and auxiliary classification model to process new medical monitoring data and obtain the final diagnosis output. In one embodiment, the collected raw data is input into the trained feature extraction and feature dimension reduction model for feature processing, and further, the processed features are input into the classifier model for classifier training to obtain the classification result. In this embodiment, the classification categories include: normal, early warning, early stage of disease, mid-stage of disease, and late stage of disease, a total of 5 categories.
[0226] It should be noted that: this embodiment adopts the adaptive wolf pack optimization algorithm to make the feature extraction process more efficient, optimize the network parameters, and improve the accuracy of feature expression and the generalization performance of the model.
[0227] The adopted mask-enhanced autoencoder neural network selectively hides the input data, forcing the model to learn more compact and representative feature representations, thereby improving the quality of data compression and reconstruction.
[0228] The extreme learning machine with adaptive memory network architecture can quickly adapt to changes in input data by dynamically adjusting memory states and weights, thereby improving classification accuracy and robustness.
[0229] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A medical monitoring data expansion method based on artificial intelligence, characterized in that: include: Collect real physiological and biochemical data; Establishing a generative adversarial network based on adaptive feature manifold projection, wherein the generative adversarial network based on adaptive feature manifold projection includes a generator and a discriminator, and enhances the generated data through a dynamic loss function adjustment mechanism; Use the collected physiological and biochemical data to train a generative adversarial network based on adaptive feature manifold projection, and expand the dataset through the trained generative adversarial network based on adaptive feature manifold projection; The training process of the generative adversarial network based on adaptive feature manifold projection is as follows: Initialize the network parameters of the generator and discriminator, and define the weights of the feature manifold projection module of the generator; Pre-train the discriminator and use real datasets to optimize its parameters; In the adversarial training phase, the data generated by the generator is adjusted through the feature manifold projection module; The generated data is fed into the discriminator, and the loss functions of the generator and the discriminator are calculated at the same time, and the loss functions of the generator and the discriminator are adjusted according to the feature manifold projection error; The discriminator is optimized based on the loss function of synthetic data and real data, and the parameters of the discriminator are updated; After each round of training, the weight and structure of the loss function are dynamically adjusted according to the similarity of the feature manifold between the generated data and the real data; Enhance the generated data using an adaptive feature weight adjustment mechanism; Repeat the above steps until the preset stop iteration condition is met; The specific formula for adjusting the data generated by the generator through the feature manifold projection module is as follows: , , , , In the formula, is random noise, The current batch synthetic data, For the generator parameter, is the characteristic manifold projection function, For the The generated data, is the weight of the feature manifold projection module; is the bias of the feature manifold projection module, is the LeakyReLU nonlinear transformation function; The loss functions of the generator and discriminator are adjusted according to the feature manifold projection error as follows: , , In the formula, represents the loss function of the generator, are the discriminator parameters, represents the loss function of the discriminator, is the regularization parameter, is the L2 norm; Dynamic Adjustment The calculation method is as follows: , In the formula, and They are vectors and No. elements; According to the similarity of the feature manifold between the generated data and the real data, the weight and structure of the loss function are dynamically adjusted. This is achieved by dynamically adjusting the regularization parameter based on the similarity of the feature manifold. The formula is as follows: , , , In the formula, To dynamically adjust the factor, its update rule is as follows: , In the formula, is the learning rate parameter, is the dynamic adjustment factor before updating, is the updated dynamic adjustment factor, Indicates The feature manifold similarity of the samples.
2. The method for medical monitoring data expansion based on artificial intelligence according to claim 1, characterized in that: The feature weight vector is updated by dynamic adjustment, and the discriminant generates data according to the discriminant result. , if a feature has a high error rate in the discriminator, increase the weight of the feature; On the contrary, it decreases; the feature weight vector The adjustment method is expressed as: , In the formula, is the weight vector The weight of the feature, is the weight vector before adjustment. The weight of the feature, is the first The weight of each feature; is the learning rate parameter; Denotes the discriminator loss function About The partial derivative of a feature.
3. An artificial intelligence-based medical monitoring data auxiliary classification method, characterized in that: include: Collect physiological and biochemical data and annotate the collected data; Expanding the data set by the artificial intelligence-based medical monitoring data expansion method described in any one of claims 1-2; A medical monitoring data analysis and auxiliary classification model is established, including a feature extraction module, a feature dimension reduction module and a classifier module. The expanded data set is processed by the feature extraction module, the feature dimension reduction module and the classifier module in sequence; the feature extraction module is a neural network based on an adaptive wolf pack optimization algorithm, the feature dimension reduction module is an autoencoder neural network based on mask enhancement, and the classifier module is an extreme learning machine algorithm based on an adaptive memory network architecture; The expanded data set is used to train the medical monitoring data analysis and auxiliary diagnosis model, and the collected physiological and biochemical data to be classified are input into the trained medical monitoring data analysis and auxiliary classification model for classification.
4. The artificial intelligence-based medical monitoring data auxiliary classification method according to claim 3 is characterized in that: The neural network training process based on the adaptive wolf pack optimization algorithm is as follows: Initializing parameters of a neural network, wherein the parameters of the neural network include weights and biases; Based on the adaptive wolf pack optimization algorithm, the parameters of the neural network are adjusted by simulating the social behavior of the wolf pack. Each wolf represents a set of potential solutions of the neural network. In each algorithm iteration, the search strategy is adjusted according to the optimization results of the previous generation, and the weights and biases of the neural network are updated; Repeat the above steps until the preset stop iteration condition is met.
5. The artificial intelligence-based medical monitoring data auxiliary classification method according to claim 3 is characterized in that: The training process of the autoencoder neural network based on mask enhancement is as follows: Initialize the parameters of the autoencoder network, including the weights and bias parameters of the autoencoder network, According to the framework of the greedy algorithm, set the independent training cycle and learning rate for each layer; At the beginning of each training cycle, for each input vector, a mask matrix is randomly generated; The input data is sent to the encoder after masking, and the encoder compresses the data into a low-dimensional feature representation after processing through an activation function; The low-dimensional features are fed into the decoder, which attempts to reconstruct the original input data. By calculating the loss function, the back propagation algorithm updates the parameters in the network. The greedy algorithm is used to optimize the parameters of each layer, and the Adam optimizer is used to update the parameters; After all layers are trained, the models of each layer are integrated to form a complete autoencoder model for overall fine-tuning. Repeat the above steps until the preset stop iteration condition is met.
6. The artificial intelligence-based medical monitoring data auxiliary classification method according to claim 3 is characterized in that: The training process of the extreme learning machine algorithm based on the adaptive memory network architecture is as follows: Initialize the memory unit, transform the input data into features through the adaptive memory network, and update the memory state based on the relationship between the current data and the historical data; The data after feature conversion is subjected to sparse processing of the weight matrix, and the sparsely processed feature data is input into the extreme learning machine to train the extreme learning machine; Automatically adjust regularization parameters of memory networks and extreme learning machines based on performance during training; Repeat the above steps until the preset stop iteration condition is met.
