Interstitial lung disease diagnosis method and system based on artificial intelligence

Abstract

The invention discloses an interstitial lung disease diagnosis method and system based on artificial intelligence, and the method comprises the steps: carrying out the structural processing of clinical data of a patient, constructing a multi-dimensional hypergraph structure, inputting a variational graph auto-encoder model under a constraint condition, carrying out the coding, and introducing a pathological feature constraint to generate a hidden variable; extracting high-order features by using an artificial intelligence multi-dimensional hypergraph neural network, and inputting the high-order features into a diagnosis model based on a deep neural network for optimization training; and finally, processing to-be-detected data according to the same process, calculating a diagnosis probability value, and outputting a result in combination with a threshold value. The system comprises six units including a data structured processing unit, a coding feature extraction unit, a feature aggregation unit, a model training optimization unit, a diagnosis reasoning unit and a result generation unit which are mutually connected and cooperatively work. Through multi-dimensional hypergraph modeling and a pathological constraint mechanism, the problems of poor multi-modal data fusion and lack of pathological feature constraints of a traditional method are effectively solved, and efficient and accurate diagnosis of interstitial lung diseases is realized.

Description

Technical Field

[0001] The present invention relates to the field of interstitial lung disease diagnosis, and in particular to an interstitial lung disease diagnosis method and system based on artificial intelligence. Background Art

[0002] Interstitial lung disease (ILD) is a complex disease characterized by interstitial fibrosis and inflammatory responses. Its clinical diagnosis relies on a comprehensive analysis of the patient's imaging features (such as ground-glass opacities and honeycombing), lung function indicators (such as vital capacity and carbon monoxide diffusion capacity), and pathological and laboratory test data. Traditional diagnostic methods rely heavily on the clinician's experience and suffer from subjectivity, low diagnostic efficiency, and insufficient recognition of early lesions, making them difficult to meet the needs of precision medicine. With the development of artificial intelligence technology, auxiliary diagnostic methods based on machine learning are gradually being applied to the medical field, improving diagnostic efficiency by exploring potential correlations in multi-dimensional data. However, existing technologies still face key technical bottlenecks when processing the complex data related to the diagnosis of ILD.

[0003] Existing technologies suffer from two major drawbacks: First, insufficient multimodal data fusion capabilities. Interstitial lung disease diagnostic data encompasses multiple dimensions, including imaging, pulmonary function, and laboratory tests. The physical meaning, data morphology, and feature correlations of these dimensions vary significantly. Traditional graph neural network models (such as standard graph convolutional networks) can only process binary relationships between nodes and struggle to model high-order, complex relationships between multidimensional feature parameters (such as the synergistic effect between imaging lesion volume and the ventilation / perfusion ratio of pulmonary function). This results in the loss of critical cross-modal pathological correlations during feature aggregation, hindering the diagnostic model's ability to accurately capture disease characteristics. Second, there is a lack of pathological feature constraint mechanisms. Existing AI diagnostic models are typically trained using a data-driven approach and lack explicit modeling of the pathological mechanisms of interstitial lung disease. For example, core pathological features such as lesion distribution uniformity and fiber cord density are not incorporated as prior knowledge into the model learning process. This can lead to deviations from the clinical diagnostic pathological logic when encoding the feature space, resulting in a disconnect between feature representation and actual pathological mechanisms, resulting in insufficient specificity and interpretability of diagnostic results. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides an interstitial lung disease diagnosis method and system based on artificial intelligence.

[0005] The technical solution adopted by the present invention is an artificial intelligence-based interstitial lung disease diagnosis method, comprising the following steps: Step S1: Acquire clinical data of a patient to be diagnosed, perform structured processing on the clinical data, establish a node set containing multi-dimensional features, and construct a multi-dimensional hypergraph structure based on the correlation between the nodes, wherein each node corresponds to a characteristic parameter in the clinical data, and the hyperedge represents the correlation relationship between multiple characteristic parameters; Step S2: Inputting the constructed multi-dimensional hypergraph structure into a constrained variational graph autoencoder model, encoding the node features in the hypergraph structure through a variational inference method, introducing constraints based on the pathological characteristics of interstitial lung disease during the encoding process, limiting the distribution of feature vectors in the encoding space, making the feature vectors in the encoding space conform to the feature mapping rules related to the diagnosis of interstitial lung disease, and generating a latent variable representation containing high-order semantic information; Step S3: Utilize an artificial intelligence multi-dimensional hypergraph neural network to extract high-order features from the latent variable representation. By designing a multi-layer hypergraph convolution operation, the information transfer between nodes and hyperedges is calculated in each layer. Combined with the weight differences of characteristic parameters of different types of clinical data, the features of the nodes in the hypergraph are iteratively updated, and the correlation information of multi-dimensional features is gradually aggregated to form a high-order feature representation that contains both local node features and global structural information. Step S4: Inputting the generated high-order feature representation into a pre-trained interstitial lung disease diagnostic model, wherein the diagnostic model is constructed based on a deep neural network and combined with clinical diagnostic criteria for interstitial lung disease, defining a loss function for the model, and optimizing the model parameters through a backpropagation algorithm so that the model can output a probability distribution related to the diagnosis of interstitial lung disease based on the input high-order feature representation; Step S5: Acquire test data of the patient to be diagnosed, perform structured processing, encoding, and high-order feature extraction on the test data according to the processing methods of steps S1 to S3, obtain a high-order feature representation of the test data, input the high-order feature representation into the diagnostic model optimized in step S4, and calculate the probability value of the patient to be diagnosed having interstitial lung disease; Step S6: Based on the obtained probability value and the preset diagnostic threshold, a final diagnosis result of interstitial lung disease is generated. If the probability value is greater than or equal to the diagnostic threshold, it is determined to be positive for interstitial lung disease; otherwise, it is determined to be negative.

[0006] Furthermore, in step S1, when constructing a multi-dimensional hypergraph structure, the calculation formula for the hyperedge weight is:

[0007] in, Represents a hyperedge The weight of and is the adjustment coefficient, satisfying They are the lesion volume, density and edge roughness parameters in the imaging examination data; They are vital capacity, carbon monoxide diffusion capacity and ventilation / perfusion ratio parameters in pulmonary function test data; is the imaging feature correlation function based on the Gaussian kernel function, is the correlation function of lung function characteristics based on linear regression.

[0008] Furthermore, in step S2, the encoding process formula of the variational graph autoencoder model under the constraint condition is:

[0009] in, is the variational approximate posterior distribution, represents the multi-dimensional hypergraph structure constructed in step S1, is a node set, For nodes The latent variable representation of and are the mean and variance of the encoding distribution, respectively, and are represented by the neural network Calculated, is the identity matrix; the constraints are:

[0010] in, is the expectation operation, is the interstitial lung disease feature prior vector, including the pathological characteristic parameters of lesion distribution uniformity and fiber cord density, is the preset feature constraint threshold, Represents a vector dot product operation.

[0011] Furthermore, in step S3, the hypergraph convolution operation formula of the artificial intelligence multi-dimensional hypergraph neural network is:

[0012] in, 、 Indicates the layer, +1 layer node The eigenvector of To include nodes The hyperedge set of For super edge The number of nodes included, is the weight of edge e, For the The trainable parameter matrix of the layer, is the activation function; the multi-dimensional feature weight difference is processed by node feature normalization, and the normalization formula is:

[0013] in, is the normalized node feature vector, is the original eigenvector, express norm, Prevent division by zero for very small positive constants.

[0014] Furthermore, in step S4, the loss function formula of the diagnosis model is:

[0015] in, is the loss function, is the number of training samples, For samples The true label is 1 for positive interstitial lung disease and 0 for negative. is the positive probability predicted by the model, is the weight parameter matrix of the model, is the regularization coefficient; the clinical diagnostic standard constraint is achieved by introducing a pathological feature penalty term into the loss function, and the penalty term formula is:

[0016] in, is the penalty item, The last layer node of the model The eigenvector of It is a reference vector of typical pathological features of interstitial lung disease, including the range of ground-glass opacity and honeycomb lung diameter parameters.

[0017] Furthermore, in step S5, during the high-order feature extraction process of the test data, a dynamic feature selection mechanism is introduced to calculate the importance weight of each dimensional feature according to the following formula:

[0018] in, For nodes The feature importance weights of is the parameter vector of the attention mechanism, Represents vector concatenation operation, is a global feature vector obtained by averaging all node features; the importance weight is used to adjust the contribution of test data features when inputting the diagnostic model.

[0019] Furthermore, the step S3 specifically includes: S3-1: Initialize node features of the latent variable representation obtained, normalize features of different dimensions according to the type of clinical data, eliminate the impact of dimensional differences on feature aggregation, and provide unified scale input features for subsequent hypergraph convolution operations; S3-2: Calculate the weight of each hyperedge in the hypergraph, combine the correlation of pathological features of interstitial lung disease, quantify the cross-modal feature associations between imaging, lung function, and laboratory test data, and determine the interaction strength of the node features connected by the hyperedge; S3-3: Perform multi-layer hypergraph convolution operations. In each layer, the correlation information of multi-dimensional features is gradually aggregated through information transmission between nodes and hyperedges. The node features are continuously integrated with the contextual information of adjacent nodes and hyperedges to form high-order features. S3-4: Fuse the features of the multi-layer hypergraph convolution output, combine the feature representations of different layers through weighted averaging, retain the semantic information of different levels, and finally generate a high-order feature representation that contains node local features and global structural information.

[0020] Furthermore, the step S4 specifically includes: S4-1: Build the network architecture of the deep neural network, determine the number of layers, the number of neurons in each layer, and the type of activation function; S4-2: Initialize the network weight parameters using a random initialization method combined with prior knowledge of interstitial lung disease characteristics to set the distribution range of the initial weights and provide a reasonable initial state for model training; S4-3: Input the high-order feature representation of the training data into the network, calculate the forward propagation output of the network, obtain the predicted probability distribution of the training samples, and calculate the loss function value by comparing it with the true label; S4-4: Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network weight parameters, use the optimization algorithm to update the weight parameters, and iterate the training until the loss function converges to obtain the optimized interstitial lung disease diagnostic model.

[0021] Furthermore, the step S5 specifically includes: S5-1: Structural processing of the test data of the patients to be diagnosed, converting the unstructured clinical data into a node set containing multi-dimensional features. The feature representation of the test data is consistent with that of the training data. S5-2: Input the processed test data into the constrained variational graph autoencoder model, perform feature encoding on the test data according to the trained encoding parameters, and generate a latent variable representation of the test data; S5-3: Input the latent variable representation of the test data into the artificial intelligence multi-dimensional hypergraph neural network, and extract high-order features through the trained hypergraph convolution parameters to obtain a high-order feature representation that contains the correlation information of the multi-dimensional features of the test data; S5-4: Input the high-order feature representation of the test data into the optimized diagnostic model, perform forward propagation calculations, and obtain the probability value of the patient to be diagnosed having interstitial lung disease.

[0022] An artificial intelligence-based interstitial lung disease diagnostic system, which includes: The clinical data structured processing unit is used to obtain the clinical data of the patient to be diagnosed, perform structured processing, establish a node set containing multi-dimensional features, and construct a multi-dimensional hypergraph structure based on the correlation between nodes; a constrained variational coding feature extraction unit, connected to the clinical data structured processing unit, for inputting a multidimensional hypergraph structure into a variational graph autoencoder model under constraints, and generating a latent variable representation containing high-order semantic information by introducing variational reasoning constrained by interstitial lung disease pathological features; A multi-dimensional hypergraph feature aggregation unit, connected to the constrained variational coding feature extraction unit, is used to perform multi-layer hypergraph convolution operations on latent variable representations using an artificial intelligence multi-dimensional hypergraph neural network, combine multi-dimensional feature weight differences, aggregate node local and global structural information, and generate high-order feature representations; a diagnostic model training optimization unit, connected to the multidimensional hypergraph feature aggregation unit, for constructing a deep neural network as a diagnostic model, defining a loss function in combination with clinical diagnostic criteria, optimizing model parameters through a backpropagation algorithm, and outputting a probability distribution related to the diagnosis of interstitial lung disease; A test data diagnostic reasoning unit is connected to the multidimensional hypergraph feature aggregation unit and the diagnostic model training optimization unit, and is used to perform the same structured processing, encoding, and high-order feature extraction on the test data of the patient to be diagnosed, input the obtained high-order feature representation into the optimized diagnostic model, and calculate the diagnostic probability value; The diagnosis result generating unit is connected to the test data diagnosis reasoning unit and is used to generate the final diagnosis result of interstitial lung disease according to the diagnosis probability value and the preset diagnosis threshold value, thereby completing the automated diagnosis of the patient.

[0023] Beneficial effects: The present invention proposes an artificial intelligence-based interstitial lung disease diagnosis method and system. This method maps multi-dimensional feature parameters such as imaging, lung function, and laboratory tests into hypergraph nodes by constructing a multi-dimensional hypergraph structure, and uses hyperedges to model high-order complex associations between multiple features, breaking through the limitation that traditional graph networks can only handle binary relationships. The hyperedge weight calculation combines the imaging feature correlation function with the lung function feature correlation function to quantify the interaction strength of cross-modal features, so that the artificial intelligence multi-dimensional hypergraph neural network can iteratively aggregate node local features and global structural information in multi-layer convolution operations, effectively capturing the collaborative pathological signals of cross-dimensional parameters such as lesion volume and ventilation / perfusion ratio, and solving the problem of insufficient multi-modal data association modeling. In response to the defect of the lack of pathological feature constraint mechanism, the system introduces an interstitial lung disease feature prior vector in the variational graph autoencoder encoding process, and forces the latent variable representation to maintain semantic consistency with core pathological features such as lesion distribution uniformity and fiber cord density through constraint conditions, ensuring that the feature mapping of the encoding space conforms to the clinical diagnostic logic. At the same time, a pathological feature penalty term is embedded in the loss function, requiring the feature vector of the last layer of the model to approximate the typical pathological feature reference vector, including parameters such as the range of ground-glass opacity and the diameter of honeycomb lungs. The clinical diagnostic criteria are converted into explicit constraints to prevent model learning from deviating from the pathological mechanism, thereby improving the interpretability and diagnostic specificity of feature representation. The present invention realizes high-order correlation modeling of multi-dimensional clinical data through multi-dimensional hypergraph structure and hypergraph convolution operation, effectively integrates cross-modal feature information, and improves the ability to capture early microlesions and complex pathological features; variational reasoning under constraint conditions and pathological feature penalty mechanism make the model learning process closely fit the pathophysiological mechanism of interstitial lung disease, reduce the feature representation bias caused by data drive, and enhance the reliability of diagnostic results. In addition, the system constructs a fully closed-loop intelligent diagnostic system from data input to result generation through automated clinical data structured processing, feature encoding, high-order feature extraction and diagnostic reasoning process, which significantly improves diagnostic efficiency and provides standardized and quantifiable technical support for clinical precision diagnosis, effectively making up for the shortcomings of traditional methods that rely on doctor experience and are highly subjective. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method steps of the present invention; Figure 2 It is a diagram of the system unit composition of the present invention. DETAILED DESCRIPTION

[0025] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] like Figure 1As shown, the interstitial lung disease diagnosis method based on artificial intelligence includes the following steps: Step S1: Acquire clinical data of the patient to be diagnosed, the clinical data including basic patient information, imaging examination data, pulmonary function test data, and laboratory test data; structure the clinical data to establish a node set containing multi-dimensional features; and construct a multi-dimensional hypergraph structure based on the correlation between the nodes, wherein each node corresponds to a characteristic parameter in the clinical data, and the hyperedge represents the complex correlation between multiple characteristic parameters; Specifically, in step S1, clinical data of the patient to be diagnosed is first obtained. This data includes basic patient information (such as age, gender, and medical history), imaging examination data (such as lesion volume, density, and edge roughness in CT images), pulmonary function test data (such as vital capacity, carbon monoxide diffusion capacity, and ventilation / perfusion ratio), and laboratory test data (such as blood routine tests and inflammatory markers). This data is then structured, converting each feature parameter into an independent node, forming a node set containing multidimensional features. A multidimensional hypergraph structure is constructed based on the correlation between nodes, where a hyperedge is defined as an edge connecting three or more nodes. It is used to represent the complex correlations between feature parameters of different dimensions, such as imaging, pulmonary function, and laboratory tests. For example, connecting lesion volume, vital capacity, and C-reactive protein level through a hyperedge reflects the synergistic effect of the three in the pathogenesis of interstitial lung disease. During the construction process, the feature correlation matrix of the clinical data is analyzed to determine the connection method of the hyperedges, ensuring that the hypergraph structure can accurately reflect the inherent correlations of multimodal data.

[0027] This step provides a unified mathematical modeling framework for subsequent feature encoding and high-order feature extraction by constructing a multi-dimensional hypergraph structure. Compared with the graph model that only processes binary relationships in traditional methods, the multi-dimensional hypergraph can directly model the high-order interactions of three or more feature parameters, effectively capturing the complex associations of multi-dimensional data in the diagnosis of interstitial lung diseases (such as the collaborative pathological signals of imaging features and lung function indicators). Structured processing converts unstructured clinical data into a computable set of nodes, enabling subsequent artificial intelligence models to learn based on clear feature representations, avoiding the problem of feature loss due to inconsistent data formats, and laying a data foundation for the entire diagnostic method.

[0028] Step S2: The multidimensional hypergraph structure constructed in step S1 is input into a variational graph autoencoder model under constraints. The node features in the hypergraph structure are encoded using a variational inference method. Constraints based on the pathological characteristics of interstitial lung disease are introduced into the encoding process to restrict the distribution of feature vectors in the encoding space so that they conform to the feature mapping rules related to the diagnosis of interstitial lung disease, thereby generating a latent variable representation containing high-order semantic information. Specifically, step S2 inputs the multidimensional hypergraph structure constructed in step S1 into a constrained variational graph autoencoder model. The model encodes node features in the hypergraph using a variational inference method. Specifically, a neural network is used to calculate the encoding distribution parameters (mean vector μᵥ and variance vector σᵥ²) for each node feature, generating a latent variable representation zᵥ that conforms to a normal distribution. During the encoding process, constraints based on the pathological characteristics of interstitial lung disease are introduced. These constraints are implemented using a predefined feature prior vector p, which contains core parameters reflecting the disease pathology, such as lesion distribution uniformity and fiber strand density. By constraining the dot product of the expected latent variables in the encoding space and the feature prior vector to be no less than a preset threshold τ, the latent variable representation is forced to retain key features relevant to the diagnosis of interstitial lung disease, preventing the loss of important pathological features during the encoding process. The resulting latent variable representation zᵥ not only contains the original node feature information but also incorporates the prior constraints on the pathological mechanism, forming a feature vector with high-level semantic information.

[0029] This step addresses the lack of domain knowledge guidance in traditional feature encoding methods by combining a variational graph autoencoder with pathological feature constraints. Traditional encoders perform feature compression based solely on data-driven methods, which can easily overlook the pathological logic required for clinical diagnosis. However, the constraints here explicitly incorporate the pathological features of interstitial lung disease (such as the density characteristics of fibrous cord shadows) into the encoding process, allowing the latent variable representation to maintain disease-related feature mapping patterns in low-dimensional space. This encoding method not only reduces the redundancy of the feature space but also ensures the semantic directionality of the latent variables, providing more discriminative input for subsequent high-order feature extraction in artificial intelligence multidimensional hypergraph neural networks, fundamentally improving the relevance of feature representation and disease diagnosis.

[0030] Step S3: Utilize an artificial intelligence multi-dimensional hypergraph neural network to extract high-order features from the latent variable representation obtained in step S2. By designing a multi-layer hypergraph convolution operation, the information transfer between nodes and hyperedges is calculated in each layer. Combined with the weight differences of feature parameters of different types of clinical data, the features of the nodes in the hypergraph are iteratively updated, and the correlation information of multi-dimensional features is gradually aggregated to form a high-order feature representation that contains both local node features and global structural information. Specifically, step S3 utilizes an artificial intelligence multidimensional hypergraph neural network to extract high-order features from the latent variable representation obtained in step S2. First, node features of the latent variables are initialized. Features of different dimensions are normalized based on clinical data types (e.g., imaging, pulmonary function, and laboratory tests) to eliminate the impact of dimensional differences on feature aggregation (e.g., unifying lesion volume in cubic centimeters and vital capacity in liters into dimensionless vectors). Subsequently, the weight of each hyperedge in the hypergraph is calculated. Incorporating the correlations between pathological features of interstitial lung disease, a Gaussian kernel function is used to quantify the nonlinear associations of imaging features (e.g., lesion volume, density, and edge roughness), and a linear regression model is used to quantify the linear associations of pulmonary function features (e.g., vital capacity and carbon monoxide diffusion capacity). The weights of the two feature types are balanced by adjusting coefficients α and β (satisfying α + β = 1). In the multi-layer hypergraph convolution operation, each layer updates node features through bidirectional information transfer between nodes and hyperedges: Node features are first aggregated from all nodes connected by the hyperedge, then nonlinearly transformed using a trainable parameter matrix, and then activated to generate new node features. Finally, the features output by multi-layer convolution are weighted averaged and fused to retain semantic information at different levels (such as local lesion features at the low level and global pathological patterns at the high level), forming a high-order feature representation that includes local node features and global structural information.

[0031] This step achieves deep fusion and high-order feature extraction of multi-dimensional clinical data through the design of an artificial intelligence multi-dimensional hypergraph neural network. Traditional graph neural networks can only process binary relationships between nodes and cannot capture the synergistic effects of three or more features. The hypergraph convolution operation can effectively model complex associations between cross-modal features (such as the combined effect of imaging lesion density and lung function ventilation / perfusion ratio) through differential calculation of hyperedge weights. The multi-layer convolution and feature fusion mechanism enables the model to aggregate feature information of different granularities layer by layer, from local single feature details to global multimodal pathological patterns, gradually enhancing the discriminative ability of features. This feature extraction method not only solves the problem of insufficient fusion of multi-dimensional data, but also provides subsequent diagnostic models with high-order feature representations containing rich pathological information, significantly improving the model's ability to capture the complex pathological characteristics of interstitial lung diseases.

[0032] Step S4: Inputting the high-order feature representation generated in step S3 into a pre-trained interstitial lung disease diagnostic model, wherein the diagnostic model is constructed based on a deep neural network and combined with clinical diagnostic criteria for interstitial lung disease, defining a loss function for the model, and optimizing the model parameters through a backpropagation algorithm so that the model can output a probability distribution related to the diagnosis of interstitial lung disease based on the input high-order feature representation; Specifically, step S4 inputs the high-order feature representation generated in step S3 into a pre-built interstitial lung disease diagnostic model based on a deep neural network architecture (such as a multilayer perceptron or residual network). First, the number of network layers, the number of neurons in each layer, and the type of activation function (such as the ReLU function) are determined to ensure that the network can handle the complex nonlinear relationships of high-order features. A random initialization method is used in combination with prior knowledge of interstitial lung disease characteristics (such as the weight distribution of known key pathological features) to set initial weights, providing a reasonable starting point for training. During training, the high-order feature representation is input into the network for forward propagation, and the predicted probability distribution of the training samples (i.e., the probability of being positive for interstitial lung disease) is calculated. By comparing with the true label, the loss function value containing the cross-entropy loss and the L2 regularization term is calculated, where the cross-entropy loss measures the difference between the predicted probability and the true label, and the regularization term prevents overfitting of the model. A pathological feature penalty term is introduced, requiring the node feature vector of the last layer of the model to approximate the typical pathological feature reference vector containing parameters such as the range of ground-glass opacity and honeycomb lung diameter. The gradient of the loss function with respect to the network weight is calculated through the back-propagation algorithm, and the weight parameters are iteratively updated using optimization algorithms such as Adam until the loss function converges, thereby obtaining an optimized model that can accurately output diagnosis-related probability distributions.

[0033] This step addresses the lack of pathological mechanism constraints in traditional AI diagnostic models by incorporating clinical diagnostic criteria into the model training process. Traditional models optimize solely based on the difference between predicted results and true labels, easily overlooking the consistency between feature representations and actual pathological features. However, the pathological feature penalty term here uses typical pathological parameters of interstitial lung disease (such as honeycomb diameter) as a reference standard, forcing the model's learned feature representations to conform to the pathological logic of clinical diagnosis. This dual constraint (prediction loss and pathological penalty) enables the model to not only accurately classify but also ensure the semantic directionality of the feature space, avoiding diagnostic bias caused by over-reliance on data statistics. The trained model can accurately extract core disease-related information from high-order features, providing a reliable classifier for diagnostic reasoning on subsequent test data.

[0034] Step S5: Acquire test data of the patient to be diagnosed, perform structured processing, encoding, and high-order feature extraction on the test data according to the processing methods of steps S1 to S3, obtain a high-order feature representation of the test data, input the high-order feature representation into the diagnostic model optimized in step S4, and calculate the probability value of the patient to be diagnosed having interstitial lung disease; Specifically, step S5 processes the test data from the patient to be diagnosed. First, structured processing is performed according to the method of step S1, converting unstructured data (such as free-text imaging reports) into a node set containing multi-dimensional features. This ensures that the feature representation of the test data is consistent with the training data (e.g., identical node definitions and hyperedge connections). The processed test data is then fed into the constrained variational graph autoencoder model trained in step S2. The learned encoding parameters (mean vector μᵥ and variance vector σᵥ²) are used to encode the test data features, generating a latent variable representation that conforms to the pathological constraints. The latent variables are then fed into the artificial intelligence multi-dimensional hypergraph neural network trained in step S3. Multi-layer feature extraction is performed using the optimized hypergraph convolution parameters (trainable parameter matrix Θ^(l)). A dynamic feature selection mechanism is introduced during this process. The importance weights of each dimensional feature are calculated using an attention mechanism, adjusting the contribution of different features to the high-level feature representation (e.g., enhancing the weight of key pathological features such as the extent of ground-glass opacity). Ultimately, a high-level feature representation containing information about the correlation between the multi-dimensional features of the test data is obtained, which serves as the input to the diagnostic model.

[0035] This step ensures the consistency and reliability of the diagnostic process by strictly aligning the processing flow of training and test data. The dynamic feature selection mechanism can adaptively highlight the parameters most relevant to the diagnosis of interstitial lung disease (such as honeycomb lung characteristics in imaging or reduced carbon monoxide diffusion capacity in pulmonary function) based on the feature distribution of the test data, avoiding diagnostic errors caused by interference from irrelevant features. Compared with the traditional method of processing fixed feature weights, this dynamic adjustment strategy can more accurately capture the differences in pathological characteristics of individual patients and improve the specificity of diagnostic results. The generated high-order feature representation not only retains the original information of the test data, but also incorporates the pathological association patterns learned during the training process, providing high-quality input data for the probability calculation of the diagnostic model.

[0036] Step S6: Based on the probability value obtained in step S5 and the preset diagnostic threshold, a final diagnosis result of interstitial lung disease is generated. If the probability value is greater than or equal to the diagnostic threshold, the test is determined to be positive for interstitial lung disease; otherwise, it is determined to be negative.

[0037] Specifically, step S6 generates a final diagnosis based on the probability value of the patient to be diagnosed suffering from interstitial lung disease calculated in step S5, combined with a preset diagnostic threshold. The diagnostic threshold is determined through clinical expert experience and ROC curve analysis of training data, and is usually set to the optimal critical point (such as 0.7) that balances sensitivity and specificity. If the probability value is greater than or equal to the diagnostic threshold, the patient is judged to be positive for interstitial lung disease, indicating that the patient has a high probability of having the disease; if the probability value is lower than the threshold, the patient is judged to be negative, indicating that the possibility of having the disease is low. This judgment process is entirely based on quantitative probability calculations and preset standards, avoiding the subjectivity and experience-dependence in manual diagnosis, and achieving objectivity and standardization of diagnostic results.

[0038] This step, as the final link in the diagnostic process, converts complex feature analysis into a clear clinical diagnosis by comparing probability values ​​with thresholds. The preset diagnostic threshold combines statistical analysis with clinical practice needs to ensure that the diagnostic results achieve a reasonable balance between sensitivity (the ability to correctly identify patients) and specificity (the ability to correctly exclude non-patients). This data-driven automated judgment method not only improves diagnostic efficiency, but also reduces diagnostic differences between different doctors through standardized decision-making processes, providing clinicians with repeatable and verifiable diagnostic evidence, and promoting the practical application of artificial intelligence technology in the diagnosis of interstitial lung disease.

[0039] Preferably, in step S1, when constructing a multidimensional hypergraph structure, the calculation method of the hyperedge weight satisfies the following relationship:

[0040] in, Represents a hyperedge The weight of and is the adjustment coefficient, satisfying They are the lesion volume, density and edge roughness parameters in the imaging examination data; They are vital capacity, carbon monoxide diffusion capacity and ventilation / perfusion ratio parameters in pulmonary function test data; is the imaging feature correlation function based on the Gaussian kernel function, is the correlation function of lung function characteristics based on linear regression.

[0041] Specifically, when constructing the multidimensional hypergraph structure in step S1, the calculation of hyperedge weights needs to comprehensively consider the degree of correlation between the two types of characteristic parameters: imaging and lung function. In specific implementation, the influence of different modal features is balanced by adjusting the coefficients α and β (the sum of the two is 1), where α corresponds to the imaging feature correlation function based on the Gaussian kernel function, which is used to quantify the interaction strength of nonlinear features such as lesion volume, density, and edge roughness; β corresponds to the lung function feature correlation function based on linear regression, which is used to measure the degree of correlation between linear features such as vital capacity, carbon monoxide diffusion capacity, and ventilation / perfusion ratio. This weight calculation method differentially models the complex texture features of imaging and the physiological indicator characteristics of lung function, allowing hyperedges to accurately reflect the synergistic effect of multidimensional features in the pathological mechanism of interstitial lung disease, providing a weighted basis for the subsequent feature aggregation of the hypergraph neural network, solving the problem of ignoring modal differences in the definition of edge weights in traditional hypergraph models, and improving the hypergraph structure's ability to represent complex associations in clinical data.

[0042] Preferably, in step S2, the encoding process of the variational graph autoencoder model under the constraint condition satisfies:

[0043] in, represents the multi-dimensional hypergraph structure constructed in step S1, is a node set, For nodes The latent variable representation of and are the mean and variance of the encoding distribution, respectively, and are represented by the neural network Calculated; the constraints are:

[0044] in, is the interstitial lung disease feature prior vector, including the pathological characteristic parameters of lesion distribution uniformity and fiber cord density, is the preset feature constraint threshold, Represents a vector dot product operation.

[0045] Specifically, during the encoding process, the constrained variational graph autoencoder model in step S2 calculates the encoding distribution parameters (mean vector μᵥ and variance vector σᵥ²) of each node feature through a neural network, generating a latent variable representation zᵥ that conforms to a normal distribution. The constraints introduced in this process are based on the prior vector p of interstitial lung disease features (including core pathological parameters such as lesion distribution uniformity and fiber cord density). By requiring that the dot product of the latent variable expectation and the prior vector be no less than a preset threshold τ, the latent variable representation is forced to retain pathological feature information directly related to disease diagnosis. This constraint mechanism transforms key clinically known pathological features into explicit constraints in the encoding space, avoiding the loss of effective information due to data noise or feature redundancy during variational inference. It ensures that the generated latent variable representation maintains semantic consistency with the interstitial lung disease pathological mechanism in a low-dimensional space, providing input features with clear pathological specificity for subsequent high-level feature extraction.

[0046] Preferably, in step S3, the hypergraph convolution operation of the artificial intelligence multi-dimensional hypergraph neural network satisfies:

[0047] in, Indicates the Layer Node The eigenvector of To include nodes The hyperedge set of For super edge The number of nodes included, For the The trainable parameter matrix of the layer, is the activation function; the multi-dimensional feature weight difference is processed by node feature normalization, and the normalization formula is:

[0048] in, is the normalized node feature vector, express norm, Prevent division by zero for very small positive constants.

[0049] Specifically, the artificial intelligence multi-dimensional hypergraph neural network in step S3 implements iterative updates of node features by designing a multi-layer information transmission mechanism when performing hypergraph convolution operations. In each layer of convolution, the node features first aggregate all node information connected to the hyperedge to which they belong, and generate new features after being transformed by a trainable parameter matrix and processed by an activation function. This process is normalized by the hyperedge weight and the number of nodes to ensure that the information contribution of hyperedges of different sizes is balanced. At the same time, in view of the weight differences of multi-dimensional features, the L2 norm normalization method is used to unify the scale of node features. By adding a very small positive constant to prevent division by zero, features of different dimensions such as imaging, pulmonary function, and laboratory tests are comparable when aggregated. This combination of multi-layer convolution and normalization processing can fuse high-order correlation information of cross-modal features layer by layer, retain the multi-level semantics of local node details and the global structure of the hypergraph, and solve the problem of insufficient feature fusion caused by dimensional differences in traditional graph neural networks when processing multi-dimensional data.

[0050] Preferably, in step S4, the loss function of the diagnostic model is defined as:

[0051] in, is the number of training samples, For samples The true label is 1 for positive interstitial lung disease and 0 for negative. is the positive probability predicted by the model, is the weight parameter matrix of the model, is the regularization coefficient; the clinical diagnostic standard constraint is achieved by introducing a pathological feature penalty term into the loss function, and the penalty term formula is:

[0052] in, The last layer node of the model The eigenvector of It is a reference vector of typical pathological features of interstitial lung disease, including the range of ground-glass opacity and honeycomb lung diameter parameters.

[0053] Specifically, in the diagnostic model training of step S4, the loss function is composed of cross-entropy loss, L2 regularization term and pathological feature penalty term. Cross-entropy loss measures the difference between the model's predicted probability and the true label, and the regularization term prevents overfitting by constraining the norm of the weight parameter matrix, ensuring the model's generalization ability on the training data. In particular, the pathological feature penalty term requires that the node feature vector of the last layer of the model approximate the typical pathological feature reference vector containing parameters such as the range of ground-glass opacity and honeycomb lung diameter, converting the core pathological indicators in the clinical diagnostic criteria into computable optimization targets. This mechanism forces the model to maintain consistency between the feature representation and the typical pathological characteristics of interstitial lung disease while learning the classification boundaries, avoiding deviation from the pathological logic due to over-reliance on data statistics, and improving the interpretability of the model feature space and the clinical reliability of the diagnostic results.

[0054] Preferably, in step S5, during the high-order feature extraction process of the test data, a dynamic feature selection mechanism is introduced to calculate the importance weight of each dimensional feature according to the following formula:

[0055] in, For nodes The feature importance weights of is the parameter vector of the attention mechanism, Represents vector concatenation operation, is a global feature vector obtained by averaging and pooling all node features. The importance weight is used to adjust the contribution of test data features when inputting the diagnostic model, thereby improving the specificity of the diagnostic results.

[0056] Specifically, in the test data processing of step S5, the dynamic feature selection mechanism calculates the importance weight of each dimensional feature through the attention mechanism. In specific implementation, the node feature is spliced ​​with the global feature vector (obtained by averaging all node features through pooling), mapped by the attention parameter vector, and normalized by the Softmax function to obtain the weight coefficient of each node feature. This weight is used to adjust the contribution of each dimensional feature when the test data is input into the diagnostic model, so that the model can adaptively highlight the features most relevant to the diagnosis of interstitial lung disease (such as honeycomb lung features in imaging or reduced carbon monoxide diffusion capacity indicators in pulmonary function) during the reasoning process. This dynamic adjustment strategy targets the individual feature differences of different patients, reduces the interference of irrelevant or minor features, effectively improves the sensitivity of the diagnostic results to disease-specific features, and solves the problem of insufficient diagnostic robustness caused by fixed feature weights in traditional methods.

[0057] Preferably, the step S3 specifically includes: S3-1: Initialize node features of the obtained latent variable representation and normalize features of different dimensions according to the type of clinical data to eliminate the impact of dimensional differences on feature aggregation and provide unified scale input features for subsequent hypergraph convolution operations; S3-2: Calculate the weight of each hyperedge in the hypergraph, combine the correlation of pathological features of interstitial lung disease, quantify the cross-modal feature correlation between imaging, lung function, and laboratory test data, and determine the interaction strength of the node features connected by the hyperedge; S3-3: Perform multi-layer hypergraph convolution operations. In each layer, through information transfer between nodes and hyperedges, the correlation information of multi-dimensional features is gradually aggregated, so that node features are continuously integrated with the contextual information of adjacent nodes and hyperedges to form more discriminative high-order features. S3-4: Fuse the features of the multi-layer hypergraph convolution output, combine the feature representations of different layers through weighted averaging, retain the semantic information of different levels, and finally generate a high-order feature representation that contains node local features and global structural information.

[0058] Specifically, in the step-by-step implementation of step S3, the node features of the latent variables are first initialized, and the features of different dimensions are normalized according to the clinical data type (imaging, pulmonary function, laboratory examination) to eliminate the impact of dimensional differences on subsequent calculations; then, combined with the pathological correlation of interstitial lung disease, the hyperedge weight is determined by quantifying the interaction strength of cross-modal features (such as imaging lesion density and pulmonary function ventilation / perfusion ratio), providing a weighted basis for information transmission; in the multi-layer hypergraph convolution operation, each layer gradually aggregates multi-dimensional correlation information through bidirectional information transmission between nodes and hyperedges, so that the node features are continuously integrated with contextual semantics to form more discriminative high-order features; finally, the output features of each layer are fused by weighted averaging, retaining multi-level information from low-level local details to high-level global patterns, ensuring that the generated high-order feature representation contains both node specificity and hypergraph structure correlation, providing rich input features for the diagnostic model.

[0059] Preferably, the step S4 specifically includes: S4-1: Build the network architecture of the deep neural network, determine the number of layers, the number of neurons in each layer, and the type of activation function, to ensure that the network structure can effectively handle the complex nonlinear relationships represented by the high-order features generated in step S3; S4-2: Initialize the network weight parameters using a random initialization method combined with prior knowledge of interstitial lung disease characteristics to set the distribution range of the initial weights and provide a reasonable initial state for model training; S4-3: Input the high-order feature representation of the training data into the network, calculate the forward propagation output of the network, obtain the predicted probability distribution of the training samples, and calculate the loss function value by comparing it with the true label; S4-4: Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network weight parameters, use an optimization algorithm (such as the Adam algorithm) to update the weight parameters, and iterate the training until the loss function converges to obtain the optimized interstitial lung disease diagnostic model.

[0060] Specifically, in the step-by-step implementation of step S4, first, a deep neural network architecture is constructed according to the complexity of the high-order features, and the number of layers, the number of neurons and the activation function (such as ReLU) are determined to adapt to the requirements of nonlinear feature transformation; the weight parameter initialization combines the random initialization method with the prior knowledge of interstitial lung disease characteristics to set a reasonable initial distribution range to avoid training falling into local optimality; during the forward propagation process, the high-order features are input into the network to calculate the predicted probability, and the loss function value including cross entropy and regularization terms is calculated by comparing with the true label; the backpropagation algorithm calculates the gradient and uses the optimization algorithm (such as Adam) to iteratively update the weights until the loss function converges, ensuring that the model can accurately capture the mapping relationship between high-order features and disease labels, forming a diagnostic model with strong generalization ability, and solving the problem of insufficient classification accuracy of traditional neural networks due to unreasonable initial parameters or insufficient training.

[0061] Preferably, the step S5 specifically includes: S5-1: Perform the same structural processing as step S1 on the test data of the patient to be diagnosed, converting the unstructured clinical data into a node set containing multi-dimensional features, ensuring that the feature representation of the test data is consistent with that of the training data; S5-2: Input the processed test data into the variational graph autoencoder model under the constraints of step S2, perform feature encoding on the test data according to the trained encoding parameters, and generate a latent variable representation of the test data; S5-3: Input the latent variable representation of the test data into the artificial intelligence multi-dimensional hypergraph neural network in step S3, perform high-order feature extraction using the trained hypergraph convolution parameters, and obtain a high-order feature representation containing the correlation information of the multi-dimensional features of the test data; S5-4: Input the high-order feature representation of the test data into the diagnostic model optimized in step S4, perform forward propagation calculations, and obtain the probability value of the patient to be diagnosed having interstitial lung disease, providing a data basis for the subsequent generation of diagnostic results.

[0062] Specifically, in the step-by-step implementation of step S5, the test data of the patient to be diagnosed is first structured to be consistent with the training data. Unstructured information (such as imaging report text) is converted into a node set in a unified format to ensure consistency in feature representation. The test data is then encoded using the trained variational graph autoencoder parameters to generate a latent variable representation that conforms to pathological constraints, avoiding feature shifts caused by differences in encoding parameters. High-order features are extracted using the trained hypergraph neural network parameters, retaining the cross-modal association patterns learned during training. Finally, the high-order features are input into the optimized diagnostic model, and the probability of illness is calculated through forward propagation, providing a quantitative basis for generating diagnostic results. This process strictly aligns the processing logic of training and test data, ensuring the standardization of the diagnostic process and the reliability of the results, and avoiding diagnostic bias caused by inconsistent data processing procedures.

[0063] like Figure 2 As shown, the interstitial lung disease diagnosis system based on artificial intelligence includes: The clinical data structured processing unit is used to obtain the clinical data of the patient to be diagnosed, perform structured processing on it, establish a node set containing multi-dimensional features, and construct a multi-dimensional hypergraph structure based on the correlation between the nodes; a constrained variational coding feature extraction unit, connected to the clinical data structured processing unit, for inputting a multidimensional hypergraph structure into a variational graph autoencoder model under constraints, and generating a latent variable representation containing high-order semantic information by introducing variational reasoning constrained by interstitial lung disease pathological features; A multi-dimensional hypergraph feature aggregation unit, connected to the constrained variational coding feature extraction unit, is used to perform multi-layer hypergraph convolution operations on latent variable representations using an artificial intelligence multi-dimensional hypergraph neural network, combine multi-dimensional feature weight differences, aggregate node local and global structural information, and generate high-order feature representations; a diagnostic model training optimization unit, connected to the multidimensional hypergraph feature aggregation unit, for constructing a deep neural network as a diagnostic model, defining a loss function in combination with clinical diagnostic criteria, and optimizing model parameters through a backpropagation algorithm so that the model can output a probability distribution related to the diagnosis of interstitial lung disease; A test data diagnostic reasoning unit is connected to the multidimensional hypergraph feature aggregation unit and the diagnostic model training optimization unit, and is used to perform the same structured processing, encoding, and high-order feature extraction on the test data of the patient to be diagnosed, input the obtained high-order feature representation into the optimized diagnostic model, and calculate the diagnostic probability value; The diagnosis result generating unit is connected to the test data diagnosis reasoning unit and is used to generate the final diagnosis result of interstitial lung disease according to the diagnosis probability value and the preset diagnosis threshold value, thereby completing the automated diagnosis of the patient.

[0064] The artificial intelligence-based interstitial lung disease diagnosis method and system proposed in the present invention fundamentally solves the core defects of traditional methods in multimodal data fusion and pathological mechanism modeling through innovative multi-dimensional hypergraph modeling and variational reasoning technology under constraint conditions, forming an intelligent diagnostic system that has both high-order feature association capture capabilities and pathological logic consistency.

[0065] This method improves multimodal data fusion capabilities by overcoming the limitations of traditional graph models, which can only handle binary relationships. By constructing a multidimensional hypergraph structure, it converts multidimensional feature parameters such as imaging, lung function, and laboratory tests into hypergraph nodes. Hyperedges are then used to explicitly model complex relationships between three or more feature parameters (e.g., the synergistic effect of lesion volume, density, and ventilation / perfusion ratio). The calculation of hyperedge weights combines a Gaussian kernel function with a linear regression model to quantify the nonlinear correlations between imaging features and the linear correlations between lung function features, respectively, enabling the hypergraph to accurately capture the strength of cross-modal data interactions. The artificial intelligence multidimensional hypergraph neural network uses multi-layer hypergraph convolution operations to enable bidirectional information transfer between nodes and hyperedges at each layer, iteratively aggregating the correlation information of multi-dimensional features. This process not only considers the weight differences between different data types but also preserves multi-level semantic information from local details to global structure through multi-layer feature fusion. This effectively addresses the problem of cross-modal feature loss caused by the simplistic data association modeling in traditional methods, significantly improving the ability to capture the complex pathological features of interstitial lung diseases.

[0066] To address the lack of pathological feature constraints in existing technologies, this system incorporates prior knowledge based on interstitial lung disease pathological characteristics into the encoding process of the variational graph autoencoder. By defining a feature prior vector (including core pathological parameters such as lesion distribution uniformity and fiber cord density), this system enforces semantic consistency between the latent variable representation and the pathological features in the form of expectation constraints, ensuring that the feature mapping in the encoding space conforms to the pathological logic of clinical diagnosis. During the training of the diagnostic model, the loss function further embeds a pathological feature penalty term, requiring the node feature vectors in the last layer of the model to approximate a typical pathological feature reference vector containing parameters such as ground-glass opacity extent and honeycombing diameter. This dual constraint mechanism transforms clinical diagnostic criteria into a computable model optimization objective, avoiding the feature representation bias caused by the lack of pathological priors in data-driven models and ensuring that the model learning process closely aligns with the pathophysiological mechanisms of interstitial lung disease. Furthermore, a dynamic feature selection mechanism, introduced during the testing phase, adaptively adjusts the importance of features across different dimensions through an attention mechanism, further strengthening the contribution of key pathological features to the diagnostic results and improving diagnostic specificity and interpretability.

[0067] The significant advantages of this method and system are also reflected in the full automation of the diagnostic process and the synergy of the technical solutions. The organic collaboration of six functional units, including the clinical data structured processing unit, the constrained variational coding feature extraction unit, and the multi-dimensional hypergraph feature aggregation unit, has built a closed-loop intelligent diagnostic system from data input to result generation. The multi-dimensional hypergraph structure provides a unified modeling framework for multimodal data. The variational reasoning under constraints and the hypergraph neural network form a "coding-aggregation-optimization" technical chain, which not only ensures the pathological relevance of feature representation, but also realizes the deep mining of high-order data associations. Compared with the traditional diagnostic method that relies on the doctor's experience, the system significantly reduces the subjectivity of diagnosis through standardized feature processing procedures and quantitative diagnostic logic, improves the accuracy and efficiency of early lesion identification, and provides innovative and practical technical support for the precision medicine of interstitial lung diseases.

[0068] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0069] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based diagnostic method for interstitial lung disease, characterized in that: The following steps are involved: Step S1: Acquire clinical data of a patient to be diagnosed, perform structured processing on the clinical data, establish a node set containing multi-dimensional features, and construct a multi-dimensional hypergraph structure based on the correlation between the nodes, wherein each node corresponds to a characteristic parameter in the clinical data, and the hyperedge represents the correlation relationship between multiple characteristic parameters; Step S2: Inputting the constructed multi-dimensional hypergraph structure into a constrained variational graph autoencoder model, encoding the node features in the hypergraph structure through a variational inference method, introducing constraints based on the pathological characteristics of interstitial lung disease during the encoding process, limiting the distribution of feature vectors in the encoding space, making the feature vectors in the encoding space conform to the feature mapping rules related to the diagnosis of interstitial lung disease, and generating a latent variable representation containing high-order semantic information; Step S3: Utilize an artificial intelligence multi-dimensional hypergraph neural network to extract high-order features from the latent variable representation. By designing a multi-layer hypergraph convolution operation, the information transfer between nodes and hyperedges is calculated in each layer. Combined with the weight differences of characteristic parameters of different types of clinical data, the features of the nodes in the hypergraph are iteratively updated, and the correlation information of multi-dimensional features is gradually aggregated to form a high-order feature representation that contains both local node features and global structural information. Step S4: Inputting the generated high-order feature representation into a pre-trained interstitial lung disease diagnostic model, wherein the diagnostic model is constructed based on a deep neural network and combined with clinical diagnostic criteria for interstitial lung disease, defining a loss function for the model, and optimizing the model parameters through a backpropagation algorithm so that the model can output a probability distribution related to the diagnosis of interstitial lung disease based on the input high-order feature representation; Step S5: Acquire test data of the patient to be diagnosed, perform structured processing, encoding, and high-order feature extraction on the test data according to the processing methods of steps S1 to S3, obtain a high-order feature representation of the test data, input the high-order feature representation into the diagnostic model optimized in step S4, and calculate the probability value of the patient to be diagnosed having interstitial lung disease; Step S6: Based on the obtained probability value and the preset diagnostic threshold, a final diagnosis result of interstitial lung disease is generated. If the probability value is greater than or equal to the diagnostic threshold, it is determined to be positive for interstitial lung disease; otherwise, it is determined to be negative.

2. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 1, characterized in that: In step S1, when constructing a multi-dimensional hypergraph structure, the calculation formula for the hyperedge weight is: in, Represents a hyperedge The weight of and is the adjustment coefficient, satisfying They are the lesion volume, density and edge roughness parameters in the imaging examination data; They are vital capacity, carbon monoxide diffusion capacity and ventilation / perfusion ratio parameters in pulmonary function test data; is the imaging feature correlation function based on the Gaussian kernel function, is the correlation function of lung function characteristics based on linear regression.

3. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 2, characterized in that: In step S2, the encoding process formula of the variational graph autoencoder model under the constraint condition is: in, is the variational approximate posterior distribution, represents the multi-dimensional hypergraph structure constructed in step S1, is a node set, For nodes The latent variable representation of and are the mean and variance of the encoding distribution, respectively, and are represented by the neural network Calculated, is the identity matrix; the constraints are: in, is the expectation operation, is the interstitial lung disease feature prior vector, including the pathological characteristic parameters of lesion distribution uniformity and fiber cord density, is the preset feature constraint threshold, Represents a vector dot product operation.

4. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 3, characterized in that: In step S3, the hypergraph convolution operation formula of the artificial intelligence multi-dimensional hypergraph neural network is: in, 、 Indicates the layer, +1 layer node The eigenvector of To include nodes The hyperedge set of For super edge The number of nodes included, is the weight of edge e, For the The trainable parameter matrix of the layer, is the activation function; the multi-dimensional feature weight difference is processed by node feature normalization, and the normalization formula is: in, is the normalized node feature vector, is the original eigenvector, express norm, Prevent division by zero for very small positive constants.

5. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 4, characterized in that: In step S4, the loss function formula of the diagnosis model is: in, is the loss function, is the number of training samples, For samples The true label is 1 for positive interstitial lung disease and 0 for negative. is the positive probability predicted by the model, is the weight parameter matrix of the model, is the regularization coefficient; the clinical diagnostic standard constraint is achieved by introducing a pathological feature penalty term into the loss function, and the penalty term formula is: in, is the penalty item, The last layer of nodes in the model The eigenvector of It is a reference vector of typical pathological features of interstitial lung disease, including the range of ground-glass opacity and honeycomb lung diameter parameters.

6. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 5, characterized in that: In step S5, during the high-order feature extraction process of the test data, a dynamic feature selection mechanism is introduced to calculate the importance weight of each dimension feature according to the following formula: in, For nodes The feature importance weights of is the parameter vector of the attention mechanism, Represents vector concatenation operation, is a global feature vector obtained by averaging all node features; the importance weight is used to adjust the contribution of test data features when inputting the diagnostic model.

7. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 1, characterized in that: The step S3 specifically includes: S3-1: Initialize node features of the latent variable representation obtained, normalize features of different dimensions according to the type of clinical data, eliminate the impact of dimensional differences on feature aggregation, and provide unified scale input features for subsequent hypergraph convolution operations; S3-2: Calculate the weight of each hyperedge in the hypergraph, combine the correlation of pathological features of interstitial lung disease, quantify the cross-modal feature associations between imaging, lung function, and laboratory test data, and determine the interaction strength of the node features connected by the hyperedge; S3-3: Perform multi-layer hypergraph convolution operations. In each layer, the correlation information of multi-dimensional features is gradually aggregated through information transmission between nodes and hyperedges. The node features are continuously integrated with the contextual information of adjacent nodes and hyperedges to form high-order features. S3-4: Fuse the features of the multi-layer hypergraph convolution output, combine the feature representations of different layers through weighted averaging, retain the semantic information of different levels, and finally generate a high-order feature representation that contains node local features and global structural information.

8. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 1, characterized in that: The step S4 specifically includes: S4-1: Build the network architecture of the deep neural network, determine the number of layers, the number of neurons in each layer, and the type of activation function; S4-2: Initialize the network weight parameters using a random initialization method combined with prior knowledge of interstitial lung disease characteristics to set the distribution range of the initial weights and provide a reasonable initial state for model training; S4-3: Input the high-order feature representation of the training data into the network, calculate the forward propagation output of the network, obtain the predicted probability distribution of the training samples, and calculate the loss function value by comparing it with the true label; S4-4: Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the network weight parameters, use the optimization algorithm to update the weight parameters, and iterate the training until the loss function converges to obtain the optimized interstitial lung disease diagnostic model.

9. The artificial intelligence-based interstitial lung disease diagnosis method according to claim 1, characterized in that: The step S5 specifically includes: S5-1: Structural processing of the test data of the patients to be diagnosed, converting the unstructured clinical data into a node set containing multi-dimensional features. The feature representation of the test data is consistent with that of the training data. S5-2: Input the processed test data into the constrained variational graph autoencoder model, perform feature encoding on the test data according to the trained encoding parameters, and generate a latent variable representation of the test data; S5-3: Input the latent variable representation of the test data into the artificial intelligence multi-dimensional hypergraph neural network, and extract high-order features through the trained hypergraph convolution parameters to obtain a high-order feature representation that contains the correlation information of the multi-dimensional features of the test data; S5-4: Input the high-order feature representation of the test data into the optimized diagnostic model, perform forward propagation calculations, and obtain the probability value of the patient to be diagnosed having interstitial lung disease.

10. An artificial intelligence-based interstitial lung disease diagnosis system, characterized by: include: The clinical data structured processing unit is used to obtain the clinical data of the patient to be diagnosed, perform structured processing, establish a node set containing multi-dimensional features, and construct a multi-dimensional hypergraph structure based on the correlation between nodes; a constrained variational coding feature extraction unit, connected to the clinical data structured processing unit, for inputting a multidimensional hypergraph structure into a variational graph autoencoder model under constraints, and generating a latent variable representation containing high-order semantic information by introducing variational reasoning constrained by interstitial lung disease pathological features; A multi-dimensional hypergraph feature aggregation unit, connected to the constrained variational coding feature extraction unit, is used to perform multi-layer hypergraph convolution operations on latent variable representations using an artificial intelligence multi-dimensional hypergraph neural network, combine multi-dimensional feature weight differences, aggregate node local and global structural information, and generate high-order feature representations; a diagnostic model training optimization unit, connected to the multidimensional hypergraph feature aggregation unit, for constructing a deep neural network as a diagnostic model, defining a loss function in combination with clinical diagnostic criteria, optimizing model parameters through a backpropagation algorithm, and outputting a probability distribution related to the diagnosis of interstitial lung disease; A test data diagnostic reasoning unit is connected to the multidimensional hypergraph feature aggregation unit and the diagnostic model training optimization unit, and is used to perform the same structured processing, encoding, and high-order feature extraction on the test data of the patient to be diagnosed, input the obtained high-order feature representation into the optimized diagnostic model, and calculate the diagnostic probability value; The diagnosis result generating unit is connected to the test data diagnosis reasoning unit and is used to generate the final diagnosis result of interstitial lung disease according to the diagnosis probability value and the preset diagnosis threshold value, thereby completing the automated diagnosis of the patient.

Images