Machine Learning and Data-Driven Risk Assessment System for Orthognathic Surgery

Abstract

The present disclosure provides an orthognathic surgery risk assessment system based on machine learning and data-driven, which relates to the technical fields of medicine and artificial intelligence systems, and includes: a data acquisition module for acquiring orthognathic surgery data of a patient to be evaluated, and a risk assessment system for inputting the preprocessed orthognathic surgery data of the patient into a risk classification model. In the risk classification model, a generative adversarial network based on cycle consistency is used to perform data augmentation on the orthognathic surgery data; a neural network based on dynamic adaptive oscillation is used to extract features from the augmented data; the extracted high-dimensional features are input into an autoencoder based on latent space mapping for feature dimensionality reduction; the dimensionality-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and a final auxiliary classification result of the orthognathic surgery risk assessment is output.

Description

Technical Field

[0001] The present disclosure relates to the technical fields of medicine and artificial intelligence systems, and particularly to an orthognathic surgery risk assessment system based on machine learning and data-driven methods. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] Orthognathic surgery is a complex surgical procedure for correcting facial skeletal deformities, aiming to improve the patient's facial function and appearance while enhancing the quality of life. However, orthognathic surgery involves in-depth adjustment of facial anatomical structures, with high risks and strong individual characteristics. Therefore, a comprehensive and scientific assessment of the patient's risk level is required before the surgery. Effective risk assessment can not only help doctors formulate personalized surgical plans but also reduce surgical risks, improve the postoperative recovery efficiency, and enhance patient satisfaction.

[0004] In recent years, machine learning methods have been gradually applied to the medical field, providing powerful tools for solving complex problems. In orthognathic surgery risk assessment, by establishing a data-driven prediction model, it is expected to extract key information from multi-dimensional data and perform accurate classification.

[0005] The following problems still exist in existing machine learning methods:

[0006] 1) Using a generative adversarial network to augment orthognathic surgery data, but only relying on the adversarial loss, the generated orthognathic surgery data often deviates from the distribution of real orthognathic surgery data. Especially when it comes to complex high-order features, the authenticity of the generated samples cannot be guaranteed.

[0007] 2) When traditional fully connected neural networks extract features, problems such as gradient vanishing or gradient explosion are likely to occur, and the optimization process may fall into local optima, resulting in unstable or incomplete extracted features.

[0008] 3) Existing dimensionality reduction methods are mostly based on fixed transformations or mappings, lacking in-depth mining of the internal structure of orthognathic surgery data. The expressive ability of the features after dimensionality reduction is limited, and it is difficult to fully retain the key information in the original orthognathic surgery data.

[0009] 4) In the classification process, traditional support vector machines cannot effectively handle ambiguous and noisy data, and the classification results are easily affected by outliers and data uncertainties, resulting in a decline in classification performance. Summary of the Invention

[0010] To solve the above problems, the present disclosure proposes an orthognathic surgery risk assessment system based on machine learning and data-driven approach. The system expands orthognathic surgery data through a cycle-consistent generative adversarial network, extracts features of the internal structure and patterns of orthognathic surgery data using a dynamic adaptive oscillation algorithm, dynamically adjusts the feature weights of the internal structure and patterns of orthognathic surgery data through an autoencoder based on latent space mapping to achieve feature dimensionality reduction; and then uses a support vector machine based on fuzzy constraints to process the uncertainties and noises in orthognathic surgery data, realizes risk classification, enhances the robustness of the classifier and improves the classification accuracy.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions:

[0012] An orthognathic surgery risk assessment system based on machine learning and data-driven approach, comprising:

[0013] A data acquisition module, configured to acquire orthognathic surgery data of a patient to be evaluated and preprocess it;

[0014] A risk assessment system, configured to input the preprocessed orthognathic surgery data of the patient into a risk classification model to obtain an auxiliary classification result of orthognathic surgery risk assessment;

[0015] Wherein, in the risk classification model, first, a cycle-consistent generative adversarial network is used to expand the orthognathic surgery data; for the expanded orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then the extracted high-dimensional features are dimensionally reduced using an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features to further capture the internal structure and patterns of the orthognathic surgery data; the dimensionally reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainties and noises in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output.

[0016] According to some embodiments, the present disclosure adopts the following technical solutions:

[0017] A computer program product, comprising a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0018] Acquire orthognathic surgery data of a patient to be evaluated and preprocess it;

[0019] Input the preprocessed orthognathic surgery data of the patient into a risk classification model to obtain an auxiliary classification result of orthognathic surgery risk assessment;

[0020] Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and patterns of orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output.

[0021] According to some embodiments, the present disclosure adopts the following technical solutions:

[0022] A non-transitory computer-readable storage medium for storing computer instructions, which when executed by a processor, implement the following method:

[0023] Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0024] Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result for orthognathic surgery risk assessment;

[0025] Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and patterns of orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output.

[0026] According to some embodiments, the present disclosure adopts the following technical solutions:

[0027] An electronic device, comprising: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory to enable the electronic device to execute the following method:

[0028] Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0029] Input the preprocessed orthognathic surgery data of the patient into a risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment;

[0030] Wherein, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on the orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, and the input high-dimensional features are mapped to a low-dimensional latent space representation. A feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and pattern of the orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of the orthognathic surgery risk assessment is output.

[0031] Compared with the prior art, the beneficial effects of the present disclosure are:

[0032] The machine learning and data-driven orthognathic surgery risk assessment system of the present disclosure uses a generative adversarial network based on cycle consistency for orthognathic surgery data augmentation, and uses cycle consistency loss and low-rank tensor decomposition optimization to solve the problems that the orthognathic surgery data generated by the traditional generative adversarial network cannot fully match the distribution of real orthognathic surgery data and the loss of high-order feature information; the generative adversarial network based on cycle consistency can maintain the consistency between the generated orthognathic surgery data and the distribution of real orthognathic surgery data when augmenting the orthognathic surgery data, and at the same time effectively retains high-order features, improving the authenticity and diversity of the orthognathic surgery data augmentation.

[0033] The machine learning and data-driven orthognathic surgery risk assessment system of the present disclosure uses a neural network based on dynamic adaptive oscillation for feature extraction. By dynamically adjusting the oscillation frequency, phase difference, and amplitude, it solves the problems of gradient disappearance, gradient explosion, and local optimal solutions in the training of fully connected neural networks, improving the stability of feature extraction; the dynamic adaptive oscillation algorithm enables the feature extraction process to be more fully optimized in the high-dimensional parameter space, avoiding unstable training and improving the expression ability of the extracted features.

[0034] The risk assessment system for orthognathic surgery based on machine learning and data-driven of the present disclosure uses an autoencoder based on latent space mapping for feature dimensionality reduction. Through feature adaptive optimization and entropy regularization strategies, it solves the problems in traditional dimensionality reduction algorithms that it is difficult to capture the internal structure of orthognathic surgery data and the feature quality after dimensionality reduction is not high. The autoencoder based on latent space mapping significantly reduces the feature dimension while retaining key information by capturing the internal patterns of orthognathic surgery data and dynamically adjusting feature weights, improving the utilization rate of the features after dimensionality reduction.

[0035] The risk assessment system for orthognathic surgery based on machine learning and data-driven of the present disclosure uses a support vector machine based on fuzzy constraints in the classification stage. Through the fuzzy membership function and the adaptive quantization perturbation mechanism, it solves the problem that the classification performance of traditional classifiers deteriorates when facing noisy and ambiguous orthognathic surgery data. The support vector machine based on fuzzy constraints can accurately handle the uncertainty and noise in orthognathic surgery data during the classification process, enhancing the robustness of the classifier and improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the disclosure. The illustrative embodiments and descriptions thereof of the disclosure are used to explain the disclosure and do not constitute an improper limitation of the disclosure.

[0037] Figure 1 is the training process of the neural network based on dynamic adaptive oscillation for the embodiments of the present disclosure;

[0038] Figure 2 is the flowchart of the training process of the risk classification model for the embodiments of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0041] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] Embodiment 1

[0043] In an embodiment of the present disclosure, a risk assessment system for orthognathic surgery based on machine learning and data-driven is provided, including:

[0044] A data acquisition module, configured to acquire the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0045] A risk assessment system, configured to input the preprocessed orthognathic surgery data of the patient into a risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment;

[0046] Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on the orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and pattern of the orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of the orthognathic surgery risk assessment is output.

[0047] As an embodiment, the training process of the risk classification model in the risk assessment system for orthognathic surgery based on machine learning and data-driven of the present disclosure includes data collection and annotation, data augmentation, training a feature extraction model, training a feature dimension reduction model, and training a classifier model, and then the trained risk classification model is used for orthognathic surgery risk assessment. Specifically, the risk classification model training process is as follows:

[0048] S1: Collection and annotation of orthognathic surgery training data

[0049] In this embodiment, the orthognathic surgery training data is the patient diagnosis and treatment process data before and after surgery, the patient's medical history and postoperative recovery situation, and all orthognathic surgery data are stored in a structured JSON format. The attributes of the orthognathic surgery data include:

[0050] Let Ra be the patient's age, represented as an integer; Da be the gender, represented as a character, where M represents male and F represents female; Pa be the preoperative diagnosis result, given as a list of strings; Sa be the surgical type, identified as various possible surgical categories; Ta be the operation duration, an integer in minutes; Ua be the pain score before and after the operation, using a scoring system from 0 to 10; Va be the postoperative recovery rate, represented in days; Wa be the type of postoperative complications, represented as a list of strings; Xa be the patient satisfaction score, using a star rating system from 1 to 5; Ya be the evaluation of postoperative facial symmetry, using a specific quantitative standard and scoring method.

[0051] Preferably, in practical applications, the attributes of orthognathic surgery data are usually more than 10, and the number of data attributes may reach dozens or even hundreds.

[0052] Furthermore, the collected orthognathic surgery data is labeled. The labeling method in this disclosure is manual labeling, and the labeling categories include "low risk", "medium risk", and "high risk".

[0053] S2: Augmentation of orthognathic surgery training data

[0054] Since the acquisition, labeling, and preprocessing of orthognathic surgery training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model. This disclosure uses a generative adversarial network based on cycle consistency for sample generation, thereby realizing the augmentation of orthognathic surgery data. Based on the traditional generative adversarial network, this disclosure uses a cycle consistency loss function to ensure the consistency of the overall distribution between the generated orthognathic surgery data and the real orthognathic surgery data. The discriminator not only needs to judge the authenticity of the orthognathic surgery data but also needs to calculate the transformation consistency of the orthognathic surgery data before and after the transformation, so as to guide the generator to generate more real orthognathic surgery data, including:

[0055] S201: Initialize the parameters of the generator and discriminator of the generative adversarial network. Let the parameters of the generator and discriminator be and , where are the parameters of the generator, are the parameters of the discriminator. The generator is responsible for generating orthognathic surgery target domain data from the orthognathic surgery source domain data . The discriminator function is used to distinguish between real orthognathic surgery target domain data and generated orthognathic surgery target domain data, expressed as:

[0056] (1)

[0057] (2)

[0058] In the formula, Orthognathic surgery data generated by the generator; The probability output by the discriminator, indicating whether the orthognathic surgery data generated by the generator is real; The generator function; The discriminator function.

[0059] S202: The generator generates forged orthognathic surgery target domain data by learning features from the orthognathic surgery source domain data. During training, the generator continuously adjusts its weights to minimize the difference between the generated orthognathic surgery data and the orthognathic surgery target domain data. The difference between the output of the generator and the real target data is measured by a standard adversarial loss function, expressed as:

[0060] (3)

[0061] In the formula, represents the distribution of the orthognathic surgery source domain data, is the orthognathic surgery source domain data; represents being subject to a specific distribution; is the orthognathic surgery target domain data; represents the expectation; is the adversarial loss function of the generator.

[0062] Furthermore, the output of the discriminator is calculated through a Bayesian inference model, and the calculation method is expressed as:

[0063] (4)

[0064] In the formula, is the Sigmoid activation function.

[0065] Furthermore, to ensure that the conversion of the generated orthognathic surgery data between the target domain and the source domain is consistent, the generator needs to learn an inverse process, that is, to generate orthognathic surgery source domain data from the target domain and perform recovery. This disclosure uses a cycle consistency loss to calculate the error of the generated orthognathic surgery data from the target domain back to the source domain, expressed as:

[0066] (5)

[0067] In the formula, is the cycle consistency loss function; represents the inverse process of the generator, that is, mapping the generated orthognathic surgery data in the target domain back to the source domain; represents the inverse matrix of the generator parameter matrix; represents the norm, used to measure the recovery error.

[0068] Furthermore, the inverse process of the generator can be assumed to be a combination of linear mapping and non-linear adjustment, expressed as:

[0069] (6)

[0070] wherein, and are the learning parameters of the inverse generator, representing the weight matrix and bias in the inverse mapping process respectively.

[0071] Furthermore, the calculation method of the total loss function of the generator is expressed as:

[0072] (7)

[0073] wherein, is the weight hyperparameter of the cycle consistency loss; is the total loss function of the generator; is the low-rank tensor decomposition loss function; is the weight hyperparameter of the low-rank tensor decomposition loss. Preferably, is set to 0.2, is set to 0.3.

[0074] S203: In the traditional generative adversarial network, the learning processes of the generator and the inverse generator often focus on direct linear mapping. When converting the data distribution of orthognathic surgery in the present disclosure, simple linear mapping cannot effectively retain the high-order structural information of orthognathic surgery data, resulting in a decrease in the accuracy of the inverse mapping result. The present disclosure utilizes the generator inverse mapping optimization strategy based on tensor low-rank approximation, enabling the generator to better retain the high-order feature information of orthognathic surgery data during the conversion process of orthognathic surgery data and improving the accuracy of the inverse mapping at the same time. Specifically, the key of tensor low-rank approximation is to decompose the high-dimensional orthognathic surgery data into a low-rank tensor representation, that is, to reconstruct it through the method of low-rank tensor decomposition to obtain a more effective mapping. For the orthognathic surgery data generated in the target domain, the optimization process of the inverse generator is expressed as:

[0075] (8)

[0076] wherein, represents the low-rank tensor decomposition operation, is the rank of the tensor. Low-rank tensor decomposition projects the generated orthognathic surgery data into a lower-dimensional space and retains the main structural information of the orthognathic surgery data; is the eigenvalue of the tensor decomposition, representing the contribution weight of the th dimension; and are the left singular vector and right singular vector of the tensor respectively, representing the base structure of the orthognathic surgery data.

[0077] Furthermore, the goal of optimizing the inverse generator is to minimize the reconstruction error between the generated orthognathic surgery data and the inverse-mapped orthognathic surgery data after low-rank approximation, and to measure the difference between the inverse mapping of the target data after low-rank tensor decomposition and the orthognathic surgery source domain data. The goal is to minimize this error. The calculation method of the low-rank tensor decomposition loss function is expressed as:

[0078] (9)

[0079] In the formula, represents the Frobenius norm, that is, the square root of the sum of the squares of the elements of the matrix, representing the reconstruction error; represents the distribution of the orthognathic surgery target domain data.

[0080] S204: The task of the discriminator is to judge the authenticity of the orthognathic surgery data and evaluate whether the orthognathic surgery data generated by the generator conforms to the characteristics of the orthognathic surgery target domain data. The goal of the discriminator is to minimize its discrimination error for the real orthognathic surgery data and the generated orthognathic surgery data. The adversarial loss function of the discriminator is expressed as:

[0081] (10)

[0082] In the formula, represents the distribution of the orthognathic surgery target domain data; is the adversarial loss function of the discriminator, that is, the total loss function of the discriminator.

[0083] S205: By alternately training the generator and the discriminator, the generator can continuously improve the quality of the generated samples, and the discriminator can improve its ability to judge the authenticity of the orthognathic surgery data. After each round of training, the generator will update its weights according to the feedback of the discriminator, and the discriminator will update its discrimination criteria according to the new samples.

[0084] Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0085] After the orthognathic surgery data augmentation model training is completed, the trained orthognathic surgery data augmentation model is used to increase the number of samples. In one embodiment, if the original collected samples are 800 and the orthognathic surgery data augmentation model generates 200 samples after augmentation, then the augmented orthognathic surgery data set contains 1000 samples.

[0086] S3: Train the feature extraction model

[0087] Training the feature extraction model using the augmented orthognathic surgery data, this disclosure uses a 6-layer fully connected neural network for feature extraction. The number of nodes in each layer of the fully connected layer is 128, that is, this layer contains 128 neurons. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in local optimal solutions may be encountered, affecting the stability of training and the performance of the model. This disclosure uses a neural network based on dynamic adaptive oscillation as the feature extraction model. By simulating the non-linear oscillation behavior in physical phenomena, the algorithm can effectively explore and utilize local extrema in the high-dimensional parameter space, thereby achieving the purpose of optimizing the neural network. As Figure 1 shown, including:

[0088] S301: Initialize the model and adaptive oscillation hyperparameters, initialize the parameters of the neural network, and at the same time initialize the hyperparameters of the dynamic adaptive oscillation, including the initial phase of the oscillation and the initial amplitude of the oscillation. In one embodiment, the initialization method is expressed as:

[0089]

[0090]

[0091]

[0092] (11)

[0093] In the formula, means subject to a specific distribution; means a normal distribution with a mean of 0 and a variance of ; means a normal distribution; means the initialization variance of the neural network; is the initial weight of the neural network, is the initial bias of the neural network; is the initial phase of the oscillation; is the initial amplitude of the oscillation. Preferably, is set to 0.01.

[0094] S302: Calculate the oscillation frequency of each parameter based on the current loss function. The adjustment of the oscillation frequency depends on the local curvature estimation of the loss surface and is adjusted using historical gradient information to make the parameters more

[0095] smooth and enable more effective search in the parameter space. The adjustment method is expressed as:

[0096] (12)

[0097] In the formula, is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; is the oscillation frequency, is the oscillation frequency of the is the weight of the neural network in the is the weight of the neural network in the is the loss function of the neural network; is the partial derivative symbol; is the basic oscillation frequency; is the hyperparameter for adjusting the oscillation responsiveness.

[0098] Preferably, is set to 5, is set to 0.3; the loss function of the neural network adopts cross-entropy loss and is calculated from the output of the last layer of the neural network through a preset Softmax.

[0099] S303: The update of each parameter is simultaneously affected by the phase difference, which is determined by the success rate of the previous update and the interaction strength between parameters. The update of the phase depends on the effect of the previous parameter update and is used to simulate the delay effect of the causal relationship. The update method is expressed as:

[0100] (13)

[0101] In the formula, is the phase of the oscillation, is the phase of the oscillation in the is the phase of the oscillation in the is the phase adjustment factor; is the hyperbolic tangent function, which can limit the adjustment range of the phase and avoid instability of the algorithm caused by excessive adjustment; is the hyperparameter for adjusting the phase sensitivity. Preferably, is set to 0.1, is set to 0.3.

[0102] S304: Automatically adjust the amplitude according to the effects of parameter updates in the past few iterations. If the update of a certain parameter continuously leads to a reduction in loss, increase its amplitude; otherwise, decrease the amplitude. The adaptive adjustment method of the amplitude is expressed as:

[0103] (14)

[0104] In the formula, is the amplitude of the oscillation, is the amplitude of the oscillation in the is the amplitude of the oscillation in the is the amplitude adjustment coefficient. Preferably, is set to 0.95.

[0105] S305: Calculate the updated values of the parameters of each neural network. Specifically, parameter updates are combined with the oscillation behavior. The sine function and cosine function are used to simulate the oscillation behavior, allowing the parameters to perform periodic exploration in the gradient direction of the loss function. The update method is expressed as:

[0106] (15)

[0107] (16)

[0108] In the formula, is the weight of the neural network in the iteration; is the bias of the neural network in the iteration; is the weight of the neural network in the iteration; is the bias of the neural network in the iteration; is the current iteration number of the neural network.

[0109] S306: Repeat the above steps iteratively until the preset stop iteration condition is satisfied, and output the high-dimensional feature vector , which indicates that the model training is completed.

[0110] In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0111] S4: Train the feature dimensionality reduction model

[0112] The high-dimensional features of the orthognathic surgery data after feature extraction are input into a feature dimensionality reduction model for training the feature dimensionality reduction model. The present disclosure uses an autoencoder based on latent space mapping as the feature dimensionality reduction model. The latent space mapping means that the encoder maps the input high-dimensional features to a low-dimensional latent space representation through columnar non-linear transformation. Using a feature adaptive optimization mechanism, the parameters of the encoder will be dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and patterns of the orthognathic surgery data after feature extraction. Specifically, the encoder converts the input high-dimensional features into a low-dimensional latent representation, and the decoder reconstructs the original features from the latent representation. Based on the traditional autoencoder algorithm, the present disclosure adopts a feature weight adaptive adjustment strategy, enabling the model to focus on the features that have a greater impact on the dimensionality reduction result, thereby improving the quality of the orthognathic surgery data after dimensionality reduction. It includes:

[0113] S401: Initialize the parameters of the autoencoder network. Let be the weight matrix of the encoder, be the bias vector of the encoder, be the weight matrix of the decoder, be the bias vector of the decoder. In one embodiment, the way to initialize the parameters of the autoencoder network is random initialization.

[0114] S402: Input the high-dimensional feature vector after feature extraction into the encoder. Let the high-dimensional feature vector input into the encoder be represented as:

[0115] (17)

[0116] In the formula, is the high-dimensional feature vector input into the encoder; is the th sample's feature vector, is the total number of samples.

[0117] Furthermore, the encoder maps the input high-dimensional features to a low-dimensional latent space representation through non-linear transformation. The present disclosure uses a feature adaptive optimization mechanism, and the parameters of the encoder will be dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and patterns of the orthognathic surgery data after feature extraction, which is represented as:

[0118] (18)

[0119] In the formula, is the low-dimensional representation in the latent space; is the mapping function of the encoder; is the adaptive activation function.

[0120] Furthermore, let , the calculation method of the adaptive activation function is expressed as:

[0121] (19)

[0122] In the formula, is the input of the adaptive activation function, is the feature adaptive parameter; is the dynamic weight coefficient, and its value range is [0, 1], which is used to balance the influence of the two activation functions; is the hyperbolic tangent function; is the ReLU activation function; is the adaptive parameter of the ReLU activation function.

[0123] Further, the calculation method of the feature adaptive parameter is expressed as:

[0124] (20)

[0125] In the formula, is the adjustment coefficient of the parameter; is the mean value of the input vector, calculated as , is the dimension of the input vector; is a small constant to avoid division by zero. Preferably, is set to 0.2, is set to 0.001.

[0126] Further, the calculation method of the dynamic weight coefficient is expressed as:

[0127] (21)

[0128] In the formula, is the L2 norm.

[0129] Further, the calculation method of the adaptive parameter of the ReLU activation function is expressed as:

[0130] (22)

[0131] In the formula, is the learning rate of the adaptive parameter; is the control parameter of the adaptive parameter; is the skewness index of the feature. Preferably, is set to 0.2.

[0132] Further, the skewness index of the feature is calculated based on the feature variance statistics and is expressed as:

[0133] (23)

[0134] (24)

[0135] Wherein, is the characteristic variance; is the total number of samples.

[0136] S403: The decoder receives the low-dimensional latent representation from the encoder and attempts to reconstruct the original high-dimensional feature vector through an inverse non-linear transformation, expressed as:

[0137] (25)

[0138] Wherein, is the high-dimensional feature vector reconstructed by the decoder; is the mapping function of the decoder; is the adaptive activation function.

[0139] S404: Compare the reconstructed feature output by the decoder with the original input feature, calculate the reconstruction error between the two, and the reconstruction error reflects the degree of information loss during the dimensionality reduction and reconstruction processes. The calculation method is expressed as:

[0140] (26)

[0141] Wherein, is the reconstruction loss of the autoencoder; is the reconstructed feature vector of the decoder for the th sample; is the regularization parameter; is the regularization term of the autoencoder weights; is the entropy regularization coefficient, which controls the influence of the entropy term on the total loss; is the entropy term of the low-dimensional representation in the latent space.

[0142] Furthermore, the regularization term of the autoencoder weights adopts sparse regularization and feature decoration terms to prevent overfitting, promote the sparsity of the autoencoder weights, and reduce the correlation between the encoder weight vectors. The calculation method is expressed as:

[0143] (27)

[0144] Wherein, represents the Frobenius norm, which is used to prevent overfitting and promote the generalization ability of the model; is the first regularization coefficient of the autoencoder; is the second regularization coefficient of the autoencoder; is the element in the th row and th column of the encoder weight matrix; is the element of the th row of the encoder weight matrix; is the element of the th row of the encoder weight matrix.

[0145] S405: Calculate the entropy term of the low-dimensional representation of the latent space using an entropy-constrained latent space information compression strategy. The calculation method is expressed as:

[0146] (28)

[0147] In the formula, is the value of the th sample in the th dimension of the latent space; is the total number of samples; is the dimension of the low-dimensional representation of the latent space, that is, the dimension of the feature vector after dimensionality reduction; is the probability density function of the value of the th sample in the th dimension of the latent space.

[0148] Furthermore, since the latent representation is the continuous output of the encoder, its probability density function needs to be estimated and calculated using the kernel density estimation method, which is expressed as:

[0149] (29)

[0150] In the formula, is the bandwidth parameter, which controls the smoothness of the kernel function; is the value of the th sample in the th dimension of the latent space; is the kernel function. Preferably, is set to 2.

[0151] Furthermore, the kernel function is calculated using the Gaussian kernel, which is expressed as:

[0152] (30)

[0153] In the formula, represents the input of the function.

[0154] S406: Update the parameters of the encoder and decoder, dynamically adjust the corresponding weights and learning rates, so that the model pays more attention to the accurate reconstruction of important features. The update method of the autoencoder parameters is expressed as:

[0155]

[0156]

[0157]

[0158] (31)

[0159] In the formula, is the adaptive learning rate of the autoencoder; is the parameter update operation; is the gradient of the reconstruction loss of the autoencoder with respect to the encoder weight parameters; is the gradient of the reconstruction loss of the autoencoder with respect to the decoder weight parameters; is the gradient of the reconstruction loss of the autoencoder with respect to the encoder bias parameters; is the gradient of the reconstruction loss of the autoencoder with respect to the decoder bias parameters.

[0160] Furthermore, the calculation method of the adaptive learning rate of the autoencoder is expressed as:

[0161] (32)

[0162] S407: Repeat the above steps iteratively until the preset iteration stop condition is satisfied, which means the model training is completed, and the dimensionality-reduced feature vector is obtained.

[0163] In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0164] S5: Train the classifier model

[0165] Input the feature vector of the orthognathic surgery data after dimensionality reduction into the classifier for classification. In this disclosure, a support vector machine based on fuzzy constraints is used as the classifier model. On the basis of the traditional support vector machine, fuzzy logic is adopted to solve the limitations of the traditional support vector machine algorithm in dealing with uncertainty, noise, and fuzzy data, so that it can classify more accurately when facing the diagnosis task of temporomandibular joint diseases and has strong robustness.

[0166] Structurally, the support vector machine based on fuzzy constraints consists of the following main parts:

[0167] Fuzzification layer: Fuzzify the orthognathic surgery data with reduced features input, so that each data point is no longer a definite value but represented by multiple fuzzy values. Through this layer, the uncertainty and fuzziness in the orthognathic surgery data can be processed;

[0168] Optimization layer: Using the framework of support vector machines based on fuzzy constraints, an optimization model is established. This model obtains a classification hyperplane by maximizing the interval between classes while minimizing the within-class differences. The orthognathic surgery data is fuzzified so that the classification hyperplane can still maintain high accuracy in the presence of noise and uncertainty;

[0169] Decision layer: Based on the hyperplane obtained from the optimization layer, decision-making reasoning is carried out to determine which class a new sample belongs to. Through the fuzzy inference mechanism, the uncertainty of the classification result is processed to give the final classification result.

[0170] Specifically, the training process of the support vector machine based on fuzzy constraints is as follows:

[0171] S501: Fuzzify the orthognathic surgery data after feature dimensionality reduction. The feature values of each sample will be converted into fuzzy values. The degree of fuzzification is controlled by the fuzzy membership function. The purpose of the fuzzification process is to introduce the fuzziness and uncertainty of the input orthognathic surgery data into the model, so as to provide more abundant information for the subsequent training process. Specifically, for each input data point , is the th feature of the orthognathic surgery data input to the support vector machine. Use the fuzzification membership function to convert it into a fuzzy value, representing the fuzziness of each data point, expressed as:

[0172] (33)

[0173] In the formula, is the fuzzy membership function, represents the fuzzy membership of the sample on the th feature; is the th feature of the sample; is the mean of the th feature of the sample; is the fuzzification control parameter; is the exponential function.

[0174] Furthermore, the fuzzification control parameter determines the curve steepness of the fuzzy membership, and is dynamically adjusted according to the feature variance of the sample. The calculation method is expressed as:

[0175] (34)

[0176] In the formula, is the standard deviation of all features of the sample; is a small positive number used to prevent division by zero errors. Preferably, is set to 0.01.

[0177] S502: During the training process of the support vector machine based on fuzzy constraints, by inputting the fuzzified orthognathic surgery data into the support vector machine model based on fuzzy constraints, the goal of the support vector machine algorithm based on fuzzy constraints is to construct a classifier by maximizing the inter-class margin (i.e., the distance between the decision hyperplane and various sample classes) and minimizing the intra-class difference (i.e., the distance of the samples from the hyperplane). During this process, the optimization goal is to improve the robustness of classification by maximizing the inter-class margin and enhance the classification accuracy by minimizing the intra-class difference;

[0178] Specifically, let the training set input to the support vector machine be , be the fuzzified feature of the sample, be the corresponding label. The optimization goal of the support vector machine uses a fuzziness constraint term and a perturbation term to enhance the robustness of the model, expressed as:

[0179] (35)

[0180] The constraint conditions are:

[0181] (36)

[0182] In the formula, is the weight vector of the sample; is the L2 norm; is the bias term of the support vector machine; is the slack variable, used to allow misclassification of some samples; is the regularization parameter of the support vector machine, used to balance the weights between maximizing the margin and minimizing misclassification; is the number of samples input to the support vector machine in the current batch; is the adjustment coefficient of the fuzziness constraint optimization term; is the fuzziness constraint optimization term, representing the uncertainty of the fuzzy membership degree; is the adjustment coefficient of the perturbation intensity; is the random perturbation term.

[0183] S503: Adopt an adaptive quantization perturbation mechanism. According to the distribution of sample features and the degree of noise, dynamically adjust the perturbation intensity, which helps to avoid overfitting and improve the tolerance of the model to outliers. The calculation method of the random perturbation term is expressed as:

[0184] (37)

[0185] In the formula, is the adjustment coefficient of the perturbation intensity; is the perturbation growth rate parameter; is a preset threshold value used to control the triggering condition of the perturbation; is a perturbation intensity adjustment function. Preferably, is set to 2, is set to 0.5.

[0186] Furthermore, a fuzzy constraint optimization strategy is adopted to adaptively adjust the fuzzy membership degree instead of keeping it fixed. The classifier can automatically adjust the fuzzy boundary according to different sample features, enhancing the adaptability of the model in the changing orthognathic surgery data environment. The calculation method of the fuzzy constraint optimization term is expressed as:

[0187] (38)

[0188] In the formula, is the feature dimension of the sample input to the support vector machine; represents the fuzzy membership degree of the sample on the th feature; is the th feature of the sample.

[0189] Furthermore, the perturbation intensity adjustment function is a dynamic function related to the model weights and input features, and dynamically adjusts the perturbation intensity according to the local perturbation and fuzzification of the current sample. The calculation method is expressed as:

[0190] (39)

[0191] In the formula, is a coefficient automatically learned during the training process, representing the adjustment speed of the perturbation intensity.

[0192] S504: After all the training processes are completed, the model will obtain an optimal hyperplane that can handle fuzzy data. Classification is performed on this hyperplane. The model will calculate the membership degrees of the input new samples with various categories and output the final classification results. Specifically, the finally obtained decision hyperplane is , and the obtained classification decision function is:

[0193] (40)

[0194] In the formula, is the classification result of the new sample; is the fuzzified feature of the new sample; is the sign function, which determines the classification result according to the position of the hyperplane; is the random perturbation term; is each input data point; T is the transpose.

[0195] S6: Application of orthognathic surgery risk assessment

[0196] Using the trained risk classification model for orthognathic surgery risk assessment. In one embodiment, first, collect the original orthognathic surgery data and input it into the trained risk classification model, and perform feature processing in the trained feature extraction and feature dimensionality reduction model. Further, input the processed features into the classifier model for classification, and then obtain the classification result. In this embodiment, the classification categories include "low risk", "medium risk", and "high risk".

[0197] Embodiment 2

[0198] In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0199] Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0200] Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment;

[0201] Wherein, in the risk classification model, first, use the generative adversarial network based on cycle consistency to perform data augmentation on the orthognathic surgery data; for the augmented orthognathic surgery data, use the neural network based on dynamic adaptive oscillation to perform feature extraction; then, use the autoencoder based on latent space mapping to perform feature dimensionality reduction on the extracted high-dimensional features, map the input high-dimensional features to a low-dimensional latent space representation, and adopt a feature adaptive optimization mechanism to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and pattern of the orthognathic surgery data; input the dimensionality-reduced feature representation into the support vector machine based on fuzzy constraints, and process the uncertainty and noise in the feature representation of the orthognathic surgery data through the fuzzy membership function and the adaptive quantization perturbation mechanism, and output the final auxiliary classification result of the orthognathic surgery risk assessment.

[0202] Embodiment 3

[0203] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, and the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the following method is implemented:

[0204] Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0205] Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment;

[0206] Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and patterns of orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output.

[0207] Embodiment 4

[0208] In an embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory, so that the electronic device executes the following method:

[0209] Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it;

[0210] Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result of orthognathic surgery risk assessment;

[0211] Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and patterns of orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output.

[0212] This disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in multiple blocks.

[0213] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or steps for implementing the functions specified in multiple blocks.

[0214] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, this is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A risk assessment system for orthognathic surgery based on machine learning and data-driven, characterized in that, Including: A data acquisition module, configured to acquire orthognathic surgery data of a patient to be evaluated and perform preprocessing; A risk assessment system, configured to input the preprocessed orthognathic surgery data of the patient into a risk classification model to obtain an auxiliary classification result of orthognathic surgery risk assessment; Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension by an autoencoder based on latent space mapping, and the input high-dimensional features are mapped to a low-dimensional latent space representation. A feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, further capturing the internal structure and pattern of orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of orthognathic surgery risk assessment is output; Adopting a feature adaptive optimization mechanism, the parameters of the encoder are dynamically adjusted according to the statistical characteristics and correlations of the input features, better capturing the internal structure and pattern of the orthognathic surgery data after feature extraction, expressed as: In the formula, is the low-dimensional representation of the latent space; is the mapping function of the encoder; is the adaptive activation function; Further, let , and the calculation method of the adaptive activation function is expressed as: In the formula, is the input of the adaptive activation function, is the feature adaptive parameter; is the dynamic weight coefficient, and its value range is [0, 1], which is used to balance the influence of the two activation functions; is the hyperbolic tangent function; is the ReLU activation function; is the adaptive parameter of the ReLU activation function; Furthermore, the calculation method of the feature adaptive parameter is expressed as: In the formula, is the adjustment coefficient of the parameter; is the mean of the input vector, calculated as , is the dimension of the input vector; is a small constant to avoid division by zero; is set to 0.2, is set to 0.001; Furthermore, the calculation method of the dynamic weight coefficient is expressed as: In the formula, is the L2 norm; Furthermore, the calculation method of the adaptive parameter of the ReLU activation function is expressed as: wherein, is the learning rate of the adaptive parameter; is the control parameter of the adaptive parameter; is the skewness index of the feature; is set to 0.2; Furthermore, the skewness index of the feature is calculated based on the feature variance statistics, expressed as: In the formula, is the characteristic variance, is the total number of samples; The decoder receives the low-dimensional latent representation from the encoder and attempts to reconstruct the original high-dimensional feature vector through an inverse non-linear transformation, expressed as: wherein, is the high-dimensional feature vector reconstructed by the decoder; is the mapping function of the decoder; is the adaptive activation function; The reconstructed feature output by the decoder is compared with the original input feature, and the reconstruction error between the two is calculated. The reconstruction error reflects the degree of information loss in the dimension reduction and reconstruction processes, and the calculation method is expressed as: In the formula, is the reconstruction loss of the autoencoder; is the reconstruction feature vector of the decoder for the -th sample; is the regularization parameter; is the regularization term of the autoencoder weights; is the entropy regularization coefficient, controlling the influence of the entropy term on the total loss; is the entropy term of the low-dimensional representation in the latent space; Furthermore, the regularization term of the autoencoder weight adopts sparse regularization and feature decoration terms to prevent overfitting, promote the sparsity of the autoencoder weight, and reduce the correlation between encoder weight vectors, and the calculation method is expressed as: In the formula, represents the Frobenius norm, which is used to prevent overfitting and promote the generalization ability of the model; is the first regularization coefficient of the autoencoder; is the second regularization coefficient of the autoencoder; is the element in the -th row and -th column of the encoder weight matrix; is the element in the -th row of the encoder weight matrix; is the element in the -th row of the encoder weight matrix; An entropy term of the low-dimensional representation of the latent space is calculated by adopting a latent space information compression strategy based on entropy constraint, and the calculation method is expressed as: Wherein, is the value of the -th sample in the -th dimension of the latent space; is the total number of samples; is the dimension of the low-dimensional representation of the latent space, that is, the dimension of the feature vector after dimensionality reduction; is the probability density function of the value of the -th sample in the -th dimension of the latent space; Furthermore, since the latent representation is the continuous output of the encoder, its probability density function needs to be estimated, and the kernel density estimation method is used for calculation, expressed as: In the formula, is the bandwidth parameter that controls the smoothness of the kernel function; is the -th sample value in the -th dimension of the latent space; is the kernel function; is set to 2; Furthermore, the kernel function is calculated using a Gaussian kernel, expressed as: In the formula, represents the input of the function; The support vector machine based on fuzzy constraints is adopted as the classifier model. The classifier model includes a fuzzification layer, an optimization layer, and a decision layer. The fuzzification layer performs fuzzification processing on the orthognathic surgery data after the input features are dimensionally reduced, so that each data point is no longer a definite value, but is represented by multiple fuzzy values. The optimization layer uses the framework of the support vector machine based on fuzzy constraints to establish an optimization model. By maximizing the inter-class margin and minimizing the intra-class difference at the same time, a classification hyperplane is obtained. Fuzzifying the orthognathic surgery data enables the classification hyperplane to still maintain high accuracy in the presence of noise and uncertainty. The decision layer makes decision reasoning based on the hyperplane obtained by the optimization layer, judges which category a new sample belongs to, and processes the uncertainty of the classification result through a fuzzy inference mechanism to give the final classification result. Finally, the decision hyperplane obtained is , and the obtained classification decision function is: wherein, is the classification result of the new sample; is the fuzzified feature of the new sample; is the sign function, which determines the classification result according to the position of the hyperplane; is the random perturbation term; is the bias term of the support vector machine; is the weight vector of the sample; is each input data point, and T is the transpose.

2. The orthognathic surgery risk assessment system based on machine learning and data-driven as claimed in claim 1, wherein The generative adversarial network based on cycle consistency is adopted to perform data augmentation on the orthognathic surgery data. The cycle consistency loss function is adopted to ensure the consistency of the overall distribution between the generated orthognathic surgery data and the real orthognathic surgery data, including: 1) Initialize the parameters of the generator and discriminator of the generative adversarial network. The generator is responsible for generating orthognathic surgery target domain data from the orthognathic surgery source domain data, and the discriminator is used to distinguish the real orthognathic surgery target domain data from the generated orthognathic surgery target domain data. 2) The generator learns features from the orthognathic surgery source domain data and generates forged orthognathic surgery target domain data. 3) Use the generator inverse mapping optimization strategy based on tensor low-rank approximation to enable the generator to retain the high-order feature information of the orthognathic surgery data during the orthognathic surgery data conversion process. 4) The discriminator judges the authenticity of the orthognathic surgery data and evaluates whether the orthognathic surgery data generated by the generator conforms to the characteristics of the orthognathic surgery target domain data, so as to minimize its discrimination error between the real orthognathic surgery data and the generated orthognathic surgery data, and construct an adversarial loss function.

3. The orthognathic surgery risk assessment system based on machine learning and data-driven as claimed in claim 1, wherein The neural network based on dynamic adaptive oscillation is adopted for feature extraction, including: The neural network based on dynamic adaptive oscillation is a multi-layer fully connected neural network. The number of nodes in each layer of the fully connected layer in the fully connected neural network is 128. By simulating the non-linear oscillation behavior in physical phenomena, local extrema are effectively extracted and utilized in the high-dimensional parameter space to optimize the neural network.

4. The orthognathic surgery risk assessment system based on machine learning and data driving according to claim 1, characterized in that In the neural network based on dynamic adaptive oscillation, the oscillation frequency of each parameter is calculated based on the current loss function. The adjustment of the oscillation frequency depends on the local curvature estimation of the loss surface and is adjusted using historical gradient information, making the parameter update smoother and enabling more effective search in the parameter space. Combine the oscillation behavior for parameter update, and use the sine function and cosine function to simulate the oscillation behavior, allowing the parameter to perform periodic exploration in the gradient direction of the loss function.

5. The risk assessment system for orthognathic surgery based on machine learning and data-driven according to claim 1, characterized in that An autoencoder based on latent space mapping is used as a feature dimensionality reduction model. The autoencoder converts the input high-dimensional features into low-dimensional latent representations. A feature adaptive optimization mechanism is adopted, and the parameters of the encoder are dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and patterns of the orthognathic surgery data after feature extraction. The decoder reconstructs the original high-dimensional feature vector through an inverse non-linear transformation. A feature weight adaptive adjustment strategy is adopted to focus on the features that have a great impact on the dimensionality reduction result.

6. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, the following method steps are implemented: Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it; Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment; Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to augment the orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension using an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation. A feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features to further capture the internal structure and patterns of the orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result of the orthognathic surgery risk assessment is output; Adopting a feature adaptive optimization mechanism, the parameters of the encoder are dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and patterns of the orthognathic surgery data after feature extraction, which is expressed as: In the formula, is the low-dimensional representation of the latent space; is the mapping function of the encoder; is the adaptive activation function; Furthermore, let , and the calculation method of the adaptive activation function is expressed as: In the formula, is the input of the adaptive activation function, is the feature adaptive parameter; is the dynamic weight coefficient, and its value range is [0, 1], which is used to balance the influence of the two activation functions; is the hyperbolic tangent function; is the ReLU activation function; is the adaptive parameter of the ReLU activation function; Furthermore, the calculation method of the feature adaptive parameter is expressed as: wherein, is the adjustment coefficient of the parameter; is the mean of the input vector, calculated as , is the dimension of the input vector; is a small constant to avoid division by zero; is set to 0.2, is set to 0.001; Furthermore, the calculation method of the dynamic weight coefficient is expressed as: In the formula, is the L2 norm; Furthermore, the calculation method of the adaptive parameter of the ReLU activation function is expressed as: In the formula, is the learning rate of the adaptive parameter; is the control parameter of the adaptive parameter; is the skewness index of the feature; is set to 0.2; Furthermore, the skewness index of the feature is calculated based on the feature variance statistics, which is expressed as: Wherein, is the characteristic variance, is the total number of samples; The decoder receives the low-dimensional latent representation from the encoder and attempts to reconstruct the original high-dimensional feature vector through an inverse non-linear transformation, which is expressed as: wherein, is the high-dimensional feature vector reconstructed by the decoder; is the mapping function of the decoder; is the adaptive activation function; Compare the reconstructed feature output by the decoder with the original input feature, calculate the reconstruction error between the two, and the reconstruction error reflects the degree of information loss in the dimension reduction and reconstruction processes. The calculation method is expressed as: wherein, is the reconstruction loss of the autoencoder; is the reconstructed feature vector of the decoder for the -th sample; is the regularization parameter; is the regularization term of the autoencoder weights; is the entropy regularization coefficient, which controls the influence of the entropy term on the total loss; is the entropy term of the low-dimensional representation in the latent space; Furthermore, the regularization term of the autoencoder weight adopts sparse regularization and feature decoration terms to prevent overfitting, promote the sparsity of the autoencoder weight, and reduce the correlation between the encoder weight vectors. The calculation method is expressed as: In the formula, represents the Frobenius norm, which is used to prevent overfitting and promote the generalization ability of the model; is the first regularization coefficient of the autoencoder; is the second regularization coefficient of the autoencoder; is the -th row and -th column element of the encoder weight matrix; is the -th row element of the encoder weight matrix; is the -th row element of the encoder weight matrix; Adopt a latent space information compression strategy based on entropy constraint to calculate the entropy term of the low-dimensional representation of the latent space. The calculation method is expressed as: Wherein, is the value of the -th sample in the -th dimension of the latent space; is the total number of samples; is the dimension of the low-dimensional representation of the latent space, that is, the dimension of the feature vector after dimensionality reduction; is the probability density function of the value of the -th sample in the -th dimension of the latent space; Furthermore, since the latent representation is the continuous output of the encoder, it is necessary to estimate its probability density function, and the kernel density estimation method is used for calculation, which is expressed as: wherein, is the bandwidth parameter, controlling the smoothness of the kernel function; is the -th sample value in the -th dimension of the latent space; is the kernel function; is set to 2; Further, the kernel function is calculated using a Gaussian kernel and is expressed as: In the formula, represents the input of the function; A support vector machine based on fuzzy constraints is used as the classifier model. The classifier model includes a fuzzification layer, an optimization layer, and a decision layer. The fuzzification layer fuzzifies the orthognathic surgery data after dimensionality reduction of the input features, such that each data point is no longer a definite value but is represented by multiple fuzzy values; the optimization layer uses the framework of a support vector machine based on fuzzy constraints to establish an optimization model, and obtains a classification hyperplane by maximizing the interval between classes while minimizing the within-class difference. Fuzzifying the orthognathic surgery data enables the classification hyperplane to still maintain high accuracy in the presence of noise and uncertainty; the decision layer makes a decision inference based on the hyperplane obtained by the optimization layer to determine which class a new sample belongs to, and processes the uncertainty of the classification result through a fuzzy inference mechanism to give the final classification result; Finally, the decision hyperplane obtained is , and the obtained classification decision function is: wherein, is the classification result of the new sample; is the fuzzified feature of the new sample; is the sign function, which determines the classification result according to the position of the hyperplane; is the random perturbation term; is the bias term of the support vector machine; is the weight vector of the sample; is each input data point, and T is the transpose.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the following method steps are implemented: Obtain the orthognathic surgery data of the patient to be evaluated and perform preprocessing; Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result for orthognathic surgery risk assessment; Wherein, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on the orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then, the extracted high-dimensional features are reduced in dimension using an autoencoder based on latent space mapping, mapping the input high-dimensional features to a low-dimensional latent space representation, and a feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features to further capture the internal structure and patterns of the orthognathic surgery data; the dimension-reduced feature representation is input into a support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through a fuzzy membership function and an adaptive quantization perturbation mechanism, and the final auxiliary classification result for orthognathic surgery risk assessment is output; Adopting a feature adaptive optimization mechanism, the parameters of the encoder are dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and patterns of the orthognathic surgery data after feature extraction, and is expressed as: In the formula, is the low-dimensional representation of the latent space; is the mapping function of the encoder; is the adaptive activation function; Further, let , the calculation method of the adaptive activation function is expressed as: wherein, is the input of the adaptive activation function, is the feature adaptive parameter; is the dynamic weight coefficient, with a value range of [0, 1], used to balance the influence of the two activation functions; is the hyperbolic tangent function; is the ReLU activation function; is the adaptive parameter of the ReLU activation function; Further, the calculation method of the feature adaptive parameter is expressed as: wherein, is the adjustment coefficient of the parameter; is the mean value of the input vector, calculated as , is the dimension of the input vector; is a small constant to avoid division by zero; is set to 0.2, is set to 0.001; Further, the calculation method of the dynamic weight coefficient is expressed as: In the formula, is the L2 norm; Further, the calculation method of the adaptive parameter of the ReLU activation function is expressed as: Wherein, is the learning rate of the adaptive parameter; is the control parameter of the adaptive parameter; is the skewness index of the feature; is set to 0.2; Further, the skewness index of the feature is calculated based on the feature variance statistics and is expressed as: In the formula, is the characteristic variance, is the total number of samples; The decoder receives the low-dimensional latent representation from the encoder and attempts to reconstruct the original high-dimensional feature vector through an inverse non-linear transformation, and is expressed as: In the formula, is the high-dimensional feature vector reconstructed by the decoder; is the mapping function of the decoder; is the adaptive activation function; Compare the reconstructed feature output by the decoder with the original input feature, calculate the reconstruction error between the two, and the reconstruction error reflects the degree of information loss during the dimension reduction and reconstruction processes. The calculation method is expressed as: wherein, is the reconstruction loss of the autoencoder; is the reconstructed feature vector of the decoder for the -th sample; is the regularization parameter; is the regularization term of the autoencoder weights; is the entropy regularization coefficient, which controls the influence of the entropy term on the total loss; is the entropy term of the low-dimensional representation in the latent space; Furthermore, the regularization term of the autoencoder weights adopts sparse regularization and feature decoration terms to prevent overfitting, promote the sparsity of the autoencoder weights, and reduce the correlation between the encoder weight vectors. The calculation method is expressed as: In the formula, represents the Frobenius norm, which is used to prevent overfitting and promote the generalization ability of the model; is the first regularization coefficient of the autoencoder; is the second regularization coefficient of the autoencoder; is the th row and th column element of the encoder weight matrix; is the th row element of the encoder weight matrix; is the th row element of the encoder weight matrix; The entropy term of the low-dimensional representation of the latent space is calculated by adopting a latent space information compression strategy based on entropy constraint. The calculation method is expressed as: Wherein, is the value of the -th sample in the -th dimension of the latent space; is the total number of samples; is the dimension of the low-dimensional representation of the latent space, that is, the dimension of the feature vector after dimensionality reduction; is the probability density function of the value of the -th sample in the -th dimension of the latent space; Furthermore, since the latent representation is the continuous output of the encoder, it is necessary to estimate its probability density function, which is calculated by adopting the kernel density estimation method and is expressed as: In the formula, is the bandwidth parameter that controls the smoothness of the kernel function; is the -th sample value in the -th dimension of the latent space; is the kernel function; is set to 2; Furthermore, the kernel function is calculated by adopting the Gaussian kernel and is expressed as: In the formula, represents the input of the function; Adopt a support vector machine based on fuzzy constraints as the classifier model. The classifier model includes a fuzzification layer, an optimization layer, and a decision layer. The fuzzification layer fuzzifies the orthognathic surgery data after dimensionality reduction of the input features, so that each data point is no longer a definite value but is represented by multiple fuzzy values; the optimization layer uses the framework of the support vector machine based on fuzzy constraints to establish an optimization model, and obtains a classification hyperplane by maximizing the inter-class margin and minimizing the intra-class difference at the same time. Fuzzifying the orthognathic surgery data enables the classification hyperplane to still maintain high accuracy in the presence of noise and uncertainty; the decision layer makes decision reasoning based on the hyperplane obtained by the optimization layer, judges which category the new sample belongs to, and processes the uncertainty of the classification result through the fuzzy inference mechanism to give the final classification result; Finally, the obtained decision hyperplane is , and the obtained classification decision function is: Wherein, is the classification result of the new sample; is the fuzzified feature of the new sample; is the sign function, which determines the classification result according to the position of the hyperplane; is the random perturbation term; is the bias term of the support vector machine; is the weight vector of the sample; is each input data point, and T is the transpose.

8. An electronic device, characterized in that, Including: A processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes the following method steps: Obtain the orthognathic surgery data of the patient to be evaluated and preprocess it; Input the preprocessed orthognathic surgery data of the patient into the risk classification model to obtain an auxiliary classification result of the orthognathic surgery risk assessment; Among them, in the risk classification model, first, a generative adversarial network based on cycle consistency is used to perform data augmentation on the orthognathic surgery data; for the augmented orthognathic surgery data, a neural network based on dynamic adaptive oscillation is used for feature extraction; then the extracted high-dimensional features are reduced in dimension by using an autoencoder based on latent space mapping, and the input high-dimensional features are mapped to a low-dimensional latent space representation. A feature adaptive optimization mechanism is adopted to dynamically adjust the encoder parameters according to the statistical characteristics and correlations of the input features, and further capture the internal structure and pattern of the orthognathic surgery data; the dimension-reduced feature representation is input into the support vector machine based on fuzzy constraints, and the uncertainty and noise in the feature representation of the orthognathic surgery data are processed through the fuzzy membership function and the adaptive quantization perturbation mechanism, and the final auxiliary classification result of the orthognathic surgery risk assessment is output; Adopt a feature adaptive optimization mechanism, and the parameters of the encoder are dynamically adjusted according to the statistical characteristics and correlations of the input features to better capture the internal structure and pattern of the orthognathic surgery data after feature extraction. It is expressed as: In the formula, is the low-dimensional representation of the latent space; is the mapping function of the encoder; is the adaptive activation function; Further, let , and the calculation method of the adaptive activation function is expressed as: Wherein, is the input of the adaptive activation function, is the feature adaptive parameter; is the dynamic weight coefficient, and its value range is [0, 1], which is used to balance the influence of the two activation functions; is the hyperbolic tangent function; is the ReLU activation function; is the adaptive parameter of the ReLU activation function; Furthermore, the calculation method of the feature adaptive parameter is expressed as: In the formula, is the adjustment coefficient of the parameter; is the mean of the input vector, calculated as , is the dimension of the input vector; is a small constant to avoid division by zero; is set to 0.2, is set to 0.001; Further, the calculation method of the dynamic weight coefficient is expressed as: In the formula, is the L2 norm; Further, the calculation method of the adaptive parameter of the ReLU activation function is expressed as: wherein, is the learning rate of the adaptive parameter; is the control parameter of the adaptive parameter; is the skewness index of the feature; is set to 0.2; Further, the skewness index of the feature is calculated based on the feature variance statistics and is expressed as: In the formula, is the characteristic variance, is the total number of samples; The decoder receives the low-dimensional latent representation from the encoder and attempts to reconstruct the original high-dimensional feature vector through an inverse non-linear transformation, which is expressed as: In the formula, is the high-dimensional feature vector reconstructed by the decoder; is the mapping function of the decoder; is the adaptive activation function; The reconstructed feature output by the decoder is compared with the original input feature, and the reconstruction error between the two is calculated. The reconstruction error reflects the degree of information loss during the dimensionality reduction and reconstruction processes, and the calculation method is expressed as: wherein, is the reconstruction loss of the autoencoder; is the reconstructed feature vector of the decoder for the -th sample; is the regularization parameter; is the regularization term of the autoencoder weights; is the entropy regularization coefficient, which controls the influence of the entropy term on the total loss; is the entropy term of the low-dimensional representation in the latent space; Further, the regularization term of the autoencoder weights adopts sparse regularization and feature decoration terms to prevent overfitting, promote the sparsity of the autoencoder weights, and reduce the correlation between the encoder weight vectors. The calculation method is expressed as: Wherein, represents the Frobenius norm, which is used to prevent overfitting and promote the generalization ability of the model; is the first regularization coefficient of the autoencoder; is the second regularization coefficient of the autoencoder; is the th row and th column element of the encoder weight matrix; is the th row element of the encoder weight matrix; is the th row element of the encoder weight matrix; The entropy term of the low-dimensional representation in the latent space is calculated using an entropy-constrained latent space information compression strategy, and the calculation method is expressed as: Wherein, is the value of the -th sample in the -th dimension of the latent space; is the total number of samples; is the dimension of the low-dimensional representation of the latent space, that is, the dimension of the feature vector after dimensionality reduction; is the probability density function of the value of the -th sample in the -th dimension of the latent space; Further, since the latent representation is the continuous output of the encoder, it is necessary to estimate its probability density function, and the kernel density estimation method is used for calculation, which is expressed as: In the formula, is the bandwidth parameter, which controls the smoothness of the kernel function; is the value of the th sample in the th dimension of the latent space; is the kernel function; is set to 2; Further, the kernel function is calculated using a Gaussian kernel, which is expressed as: In the formula, represents the input of the function; A support vector machine based on fuzzy constraints is used as the classifier model. The classifier model includes a fuzzification layer, an optimization layer, and a decision layer. The fuzzification layer fuzzifies the orthognathic surgery data after dimensionality reduction of the input features, so that each data point is no longer a definite value but is represented by multiple fuzzy values; the optimization layer uses the framework of the support vector machine based on fuzzy constraints to establish an optimization model. By maximizing the inter-class margin and minimizing the intra-class difference simultaneously, a classification hyperplane is obtained. Fuzzifying the orthognathic surgery data enables the classification hyperplane to maintain high accuracy in the presence of noise and uncertainty; the decision layer makes a decision inference based on the hyperplane obtained by the optimization layer to determine which class a new sample belongs to. Through the fuzzy inference mechanism, the uncertainty of the classification result is processed to give the final classification result. Finally, the obtained decision hyperplane is , and the obtained classification decision function is: wherein, is the classification result of the new sample; is the fuzzified feature of the new sample; is the sign function, which determines the classification result according to the position of the hyperplane; is the random perturbation term; is the bias term of the support vector machine; is the weight vector of the sample; is each input data point, and T is the transpose.

Images