Attention deficit hyperactivity disorder subtype identification method based on brain network topology hub deviation

Abstract

The invention discloses an attention deficit hyperactivity disorder subtype recognition method based on brain network topology hub deviation, relates to the technical field of medical diagnosis, and has the technical key points that a norm model of a brain structure form similarity network is constructed based on multi-center big data; quantifying a brain network topology hub index as a target phenotype of the norm model; then, performing semi-supervised clustering analysis by utilizing the individual deviation phenotype of the topological hub index so as to divide different biological subtypes and reveal unique clinical and biological characteristics of the biological subtypes; finally, strict cross validation is carried out in an external independent queue, and it is ensured that the potential biological subtypes recognized by the typing model have good generalization and clinical effectiveness.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis technology, and in particular to a method for identifying subtypes of attention deficit hyperactivity disorder based on brain network topology hub deviation. Background Art

[0002] In the diagnosis and treatment of attention deficit hyperactivity disorder (ADHD), accurately identifying biologically homogeneous subtypes is crucial for developing personalized treatment plans and targeted interventions. However, this field currently faces multiple technical challenges. Existing clinical classification methods primarily rely on symptom descriptions at the phenotypic level, overlooking underlying neurobiological mechanisms. This leads to significant heterogeneity within the same diagnostic category, making it difficult to guide personalized treatment. Existing dichotomous analytical approaches struggle to capture complex clinical phenotypic patterns and lack consideration of the range of normal variation, limiting their ability to distinguish pathological from nonpathological variants. Methodologically, previous normative modeling studies have primarily focused on morphological features of local brain regions, lacking a systematic understanding of regional morphological metrics at the brain network level. Furthermore, existing brain network feature extraction methods are incomplete and fail to fully characterize network organization. Furthermore, existing studies commonly suffer from insufficient model validation, lacking rigorous cross-validation and external validation, leaving the generalizability and clinical applicability of their results uncertain. In practical applications, existing algorithmic models are computationally expensive when processing large datasets, making them inefficient in clinical practice.

[0003] To this end, the present invention aims to provide a method for identifying subtypes of attention deficit hyperactivity disorder based on brain network topology hub deviation to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to solve the above problems and provide a method for identifying subtypes of attention deficit hyperactivity disorder based on brain network topological hub deviation. First, the original brain structure image data is quality controlled and standardized preprocessed, and then a brain structure morphological similarity network at the individual level is constructed. Multiple topological hub indicators are used to describe the network characteristics, and the brain network topological hub norm of normally developing children is established. Finally, the disease subtype is identified based on the phenotypic deviation of the network topological hub.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] The present invention provides a method for identifying subtypes of attention deficit hyperactivity disorder based on brain network topology hub deviation, the method comprising the following steps:

[0007] 1. Brain Structural Image Quality Control

[0008] The present invention uses the Computational Anatomy Toolbox (CAT12) to evaluate multiple quality dimensions of all T1-weighted magnetic resonance images, including image noise level, spatial uniformity of signal intensity, degree of head motion artifact, and accuracy of tissue segmentation. These independent quality indicators are weighted and combined to obtain a comprehensive image quality rating (IQR). To ensure the reliability of subsequent analysis, the present invention strictly controls data quality and only retains images with an IQR score below 3.5. This threshold is set based on the standardized scoring system of CAT12 to ensure that all images entering the analysis meet a quality standard of at least 70 points. This strict quality control mechanism lays a reliable data foundation for subsequent network construction and feature extraction.

[0009] 2. Image Data Preprocessing

[0010] For images that pass quality control, the present invention uses the CAT12 and SPM12 toolboxes to perform standardized preprocessing procedures. First, the original T1-weighted image is tissue segmented based on the expectation maximization algorithm, and the brain tissue is accurately divided into three types: gray matter, white matter, and cerebrospinal fluid, and the corresponding probability map is generated. This probability-based segmentation method can effectively process the partial volume effect of tissue boundaries and provide more accurate tissue classification results. Subsequently, the DARTEL algorithm is used for spatial normalization processing. While maintaining local volume information, the individual brain images are accurately aligned to the MNI standard space to ensure the accurate correspondence of anatomical structures between different individuals in subsequent analysis. During the standardization process, the present invention performs modulation processing based on the Jacobian determinant on the image, while considering the effects of linear and nonlinear transformations to maintain voxel-level gray matter volume information. In order to improve the signal-to-noise ratio and compensate for residual anatomical differences between individuals, the modulated gray matter image is spatially smoothed using a Gaussian kernel with an 8mm full width at half maximum (FWHM). Finally, the whole-brain gray matter was divided into 90 anatomical regions based on the automated anatomical labeling (AAL) atlas, and the voxel-level gray matter volume distribution of each region was extracted, which formed the basic data for subsequent morphological similarity network analysis.

[0011] 3. Construction of Morphological Similarity Brain Network

[0012] The present invention uses a morphological similarity measure based on Kullback-Leibler divergence (KLS) to construct brain networks at the individual level. KLS is an asymmetric information measure that measures the difference between two probability distributions. Its mathematical expression is:

[0013] KLS(P||Q)=∫p(x)log(p(x) / q(x))dx

[0014] Where P and Q represent the gray matter volume distribution of the two brain regions, respectively, p(x) and q(x) are the corresponding probability density functions, and the integration range covers the support set of the two distributions. The main advantage of choosing KLS as the similarity metric is that it can capture subtle differences between probability distributions and has high sensitivity to the tail characteristics of the distribution. In order to accurately estimate the gray matter volume distribution of each brain region, the present invention uses the kernel density estimation method to construct the probability density function, which includes 2 7 The sampling points are used to capture the key features of the distribution, reducing computational complexity while maintaining sufficient statistical accuracy. Based on the calculated KLS values, the present invention constructs a 90×90 morphological similarity network matrix for each subject based on the AAL brain atlas to reflect the structural association patterns between brain regions.

[0015] 4. Calculation of Brain Network Topology Hubness Indicators

[0016] The present invention conducts a quantitative analysis of the topological indicators of the gray matter volume covariation network of each subject. In order to comprehensively characterize the hub characteristics of the nodes in the network, the present invention adopts three key topological indicators: degree centrality (DC), nodal efficiency (NE) and participation coefficient (PC). During the calculation process, the present invention first uses the GRETNA toolbox to perform sparse processing on the network to remove possible false connections, and sets the sparsity range to 0.10 to 0.34 with a step size of 0.01. The selection of this range ensures that the network has minimal pseudo-edges while maintaining the characteristics of a small-world network. In order to eliminate the possible impact of arbitrary sparsity threshold selection, the present invention calculates the area under the curve of the topological characteristics of each node, thereby obtaining a network connection group feature description independent of the specific threshold, which more robustly characterizes the topological organization characteristics of the brain network. The mathematical expression of the hub index is as follows:

[0017]

[0018] where a ij represents the connection status between node i and node j [1 when the connection (i, j) exists, otherwise 0], n is the set of all nodes in the network; d ij represents the shortest path length between node i and node j; M is the total number of module communities, k im is the number of edges connecting node i with other nodes belonging to module m, k i is the total number of edges connecting the node.

[0019] 5. Establishment of a normative model of brain networks

[0020] This paper uses the normative model construction method based on the Bayesian linear regression framework in the PCN toolkit, which can effectively depict the normal development trajectory of brain network topology characteristics. The specific implementation process is as follows:

[0021] (1) All brain topology phenotype data were preprocessed with Yeo-Johnson transformation to reduce the skewness of the data distribution. In terms of data set division, a stratified sampling strategy was adopted: 90% of the samples were randomly selected from the typical development control group of each research site as the training data set, and the remaining 10% of the typical development control samples together with all attention deficit hyperactivity disorder cases constituted the test data set, aiming to test whether the attention deficit hyperactivity disorder group showed a more significant degree of deviation compared with the typical development control group.

[0022] (2) During the model construction process, the present invention simultaneously considers the influence of demographic variables (age and gender) and multicenter effects. For the age variable, B-spline basis function (cubic spline function with 5 uniformly distributed nodes) is used for nonlinear fitting to better capture the dynamic changes of developmental trajectory. In order to deal with the possible non-Gaussian distribution characteristics of the response variable, a bijective nonlinear distortion function (Sinarcsinsh) is introduced to transform the non-Gaussian response variable into a Gaussian latent space. The site-related confounding effect is controlled by incorporating the site as a fixed effect variable into the model.

[0023] (3) The Powell algorithm was used to optimize the model’s hyperparameters. A normative model was fitted independently for each topological hubness metric of the 90 brain regions, and the regression expression was as follows:

[0024] Y=∫(X,θ)+E

[0025] where Y is the brain effect size, X is the covariate, θ is the relevant model parameter, and E is the residual.

[0026] The degree to which individuals in the test data set deviate from these normative expectations is quantified using the Z score, which is calculated as follows:

[0027]

[0028] For each subject n and each brain region d,y nd is the observed value, is the model prediction value, represents the prediction noise variance, represents the norm variance.

[0029] 6. Fusion of multidimensional brain network topology indicators

[0030] The present invention uses a method combining multimodal canonical correlation analysis with joint independent component analysis (mCCA+jICA) to perform data-driven fusion analysis on multiple topological features to reveal the covariant multimodal topological patterns. In the multimodal canonical correlation analysis (mCCA) stage, for different modes k, the topological feature X of each mode is k Perform projection and generate two typical variables: mixing coefficient matrix D k and association map C k . After obtaining the association map C k After that, the present invention uses joint independent component analysis (jICA) to process it and extract the largest independent source signals through the whitening matrix W. The optimal number of source signals is determined by the minimum description length criterion. This joint decomposition process produces modality-specific components and corresponding individual-specific mixing coefficients A k The relevant mathematical expressions are as follows:

[0031] mCCA stage: X k =D k C k (mCCA)

[0032] jICA stage: X k =A k S k ,C k =S k W -1 (jICA)

[0033] In order to realize the visualization of the results, the present invention converts the independent source signal S k The data were normalized to Z scores and the threshold of |Z|≥2.0 was used for spatial mapping. In terms of clinical relevance analysis, the mixing coefficient A was compared by two-sample test. k The differences between the groups were analyzed to identify the joint components with between-group differences, and the degree of association with clinical indicators was evaluated by product-moment correlation analysis.

[0034] 7. Biological subtype identification based on semi-supervised learning

[0035] This paper uses the semi-supervised HYDRA (Heterogeneity through discriminant analysis) algorithm to identify biological subtypes. This method constructs a convex polyhedron by establishing multiple hyperplanes that simultaneously enclose the control group distribution and optimally distinguish between the case groups, where each face represents a unique biological subtype. The specific implementation process is as follows:

[0036] (1) The individual deviation maps of the three topological features (i.e., the scores of each brain region) were used as input features, and age, gender, and scanning site were included as covariates in the HYDRA modeling. In terms of algorithm parameter settings, the following optimization configuration was adopted: 10-fold nested cross-validation, the regularization parameter was set to 0.25, L1 regularization was used, random hyperplanes were assigned through a determinant point process, 50 iterations were performed between the estimated hyperplane and the cluster, and 25 steps of cluster consistency test were performed. To evaluate the stability of the HYDRA clustering solution, the adjusted Rand index (ARI) was introduced to quantify the similarity of the clustering results in the cross-validation.

[0037] (2) In terms of robustness verification, the present invention designed a double validation analysis: First, the subtype stability was evaluated by permutation testing, specifically by randomly selecting 45% of the typical development control (TDC) sample as the real TDC group and 55% as the pseudo-patient group, repeating the HYDRA analysis 100 times, and generating a null distribution. These results were then compared with the clustering results of the real attention deficit hyperactivity disorder group: TDC and patient groups of the same size were assigned in the permutation test, and repeated 100 times to ensure consistent sample size. The ARI values ​​obtained in the two cases were compared by a two-sample t-test to determine the statistical significance of the k subtypes. Secondly, the present invention performed 20 random iterations of split-half cross-validation to evaluate the consistency of each subtype in the optimal k value clustering scheme. In this process, the stratification factors of diagnosis, gender, and site were also considered as covariates.

[0038] (3) In order to explore the clinical manifestation characteristics of the identified subtypes, Kruskal-Wallis analysis was used to examine the differences in symptom severity (such as attention deficit and hyperactivity / impulsivity phenotypes) among the subtypes, so as to achieve a dimensional characterization of clinical characteristics.

[0039] 8. External Validation of Biological Subtypes

[0040] To evaluate the generalization ability of the biological subtype clustering model, this paper uses a transfer learning strategy to apply the pre-trained model to an independent validation dataset. The specific implementation process is as follows:

[0041] (1) The T1-weighted magnetic resonance images in the validation dataset are processed in a process consistent with the prior samples until the normative model training stage. Since the validation database is completely absent in the initial training set, the present invention uses a transfer learning method to fine-tune the reference model. Specifically, typical developmental control samples in an independent validation dataset are used as adaptation data to adapt to possible site-specific variations. During the adaptation process, the hyperparameters of the reference model are used as information priors for the new dataset, and the mean and variance are adjusted in the latent Gaussian space, and the adjusted data are then mapped back to the original space. After recalibrating the reference norm range, a de-site effect estimate of the validation sample is generated to characterize the individual deviations of its topological features.

[0042] (2) By using the HYDRA algorithm parameters defined in the initial training phase (weight vector w i and the bias term b i ), reconstruct the polyhedron that best distinguishes the typical development control group from each biological subtype, and apply it to the validation data. The basic principle is to use the retained hyperplane weights and bias terms to calculate the expression score E of the validation cohort i (where X is the feature vector of the validation sample and D is the predicted subtype affiliation), the expression score calculation formula is:

[0043] E i =w i X+b i

[0044] D=argmax(E i )

[0045] (3) To assess the consistency of the validation subtypes with the reference subtypes, pattern similarity was quantified by pairwise product-moment correlation analysis, generating a correlation matrix. A greedy matching approach was used to establish cluster mapping relationships, mapping each validation cluster to the reference cluster with which it showed the highest correlation. Finally, following the analysis method used in the discovery cohort, the clinical feature spectrum and neuroimaging deviation patterns of each subtype in the validation sample were simultaneously examined.

[0046] Compared with the existing technology, this solution has the following beneficial effects:

[0047] Based on multi-center big data, the present invention first constructs a normative model of the brain structural morphological similarity network and quantifies the hubness indicators at the brain region level in the brain network topology as the target phenotype of the normative model; then, semi-supervised cluster analysis is performed using the individual deviation degree of topological hubness indicators to divide different biological subtypes and reveal their unique clinical and biological characteristics; finally, rigorous cross-validation is performed in an external independent cohort to ensure that the potential biological subtypes identified by this classification model have good generalization and clinical relevance. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a method for evaluating attention deficit hyperactivity disorder in an embodiment of the present invention;

[0049] Figure 2 is a flow chart of a subtype identification method based on individual multidimensional brain network topology hub deviation in an embodiment of the present invention;

[0050] Figure 3 3. It is an effect diagram of attention deficit hyperactivity disorder subtype identification in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0052] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0053] Example:

[0054] 1. Brain Structural Image Quality Control

[0055] The present invention uses the Computational Anatomy Toolbox (CAT12) to evaluate multiple quality dimensions of all T1-weighted magnetic resonance images, including image noise level, spatial uniformity of signal intensity, degree of head motion artifact, and accuracy of tissue segmentation. These independent quality indicators are weighted and combined to obtain a comprehensive image quality rating (IQR). To ensure the reliability of subsequent analysis, the present invention strictly controls data quality and only retains images with an IQR score below 3.5. This threshold is set based on the standardized scoring system of CAT12 to ensure that all images entering the analysis meet a quality standard of at least 70 points. This strict quality control mechanism lays a reliable data foundation for subsequent network construction and feature extraction.

[0056] 2. Image Data Preprocessing

[0057] For images that pass quality control, the present invention uses the CAT12 and SPM12 toolboxes to perform standardized preprocessing procedures. First, the original T1-weighted image is tissue segmented based on the expectation maximization algorithm, and the brain tissue is accurately divided into three types: gray matter, white matter, and cerebrospinal fluid, and the corresponding probability map is generated. This probability-based segmentation method can effectively process the partial volume effect of tissue boundaries and provide more accurate tissue classification results. Subsequently, the DARTEL algorithm is used for spatial normalization processing. While maintaining local volume information, the individual brain images are accurately aligned to the MNI standard space to ensure the accurate correspondence of anatomical structures between different individuals in subsequent analysis. During the standardization process, the present invention performs modulation processing based on the Jacobian determinant on the image, while considering the effects of linear and nonlinear transformations to maintain voxel-level gray matter volume information. In order to improve the signal-to-noise ratio and compensate for residual anatomical differences between individuals, the modulated gray matter image is spatially smoothed using a Gaussian kernel with an 8mm full width at half maximum (FWHM). Finally, the whole-brain gray matter was divided into 90 anatomical regions based on the automated anatomical labeling (AAL) atlas, and the voxel-level gray matter volume distribution of each region was extracted, which formed the basic data for subsequent morphological similarity network analysis.

[0058] 3. Construction of Morphological Similarity Brain Network

[0059] The present invention uses a morphological similarity measure based on Kullback-Leibler divergence (KLS) to construct brain networks at the individual level. KLS is an asymmetric information measure that measures the difference between two probability distributions. Its mathematical expression is:

[0060] KLS(P||Q)=∫p(x)log(p(x) / q(x))dx

[0061] Where P and Q represent the gray matter volume distribution of the two brain regions, respectively, p(x) and q(x) are the corresponding probability density functions, and the integration range covers the support set of the two distributions. The main advantage of choosing KLS as the similarity metric is that it can capture subtle differences between probability distributions and has high sensitivity to the tail characteristics of the distribution. In order to accurately estimate the gray matter volume distribution of each brain region, the present invention uses the kernel density estimation method to construct the probability density function, which includes 2 7 The sampling points are used to capture the key features of the distribution, reducing computational complexity while maintaining sufficient statistical accuracy. Based on the calculated KLS values, the present invention constructs a 90×90 morphological similarity network matrix for each subject based on the AAL brain atlas to reflect the structural association patterns between brain regions.

[0062] 4. Calculation of Brain Network Topology Hubness Indicators

[0063] The present invention conducts a quantitative analysis of the topological indicators of the gray matter volume covariation network of each subject. In order to comprehensively characterize the hub characteristics of the nodes in the network, the present invention adopts three key topological indicators: degree centrality (DC), nodal efficiency (NE) and participation coefficient (PC). During the calculation process, the present invention first uses the GRETNA toolbox to perform sparse processing on the network to remove possible false connections, and sets the sparsity range to 0.10 to 0.34 with a step size of 0.01. The selection of this range ensures that the network has minimal pseudo-edges while maintaining the characteristics of a small-world network. In order to eliminate the possible impact of arbitrary sparsity threshold selection, the present invention calculates the area under the curve of the topological characteristics of each node, thereby obtaining a network connection group feature description independent of the specific threshold, which more robustly characterizes the topological organization characteristics of the brain network. The mathematical expression of the hub index is as follows:

[0064]

[0065] where a ij represents the connection status between node i and node j [1 when the connection (i, j) exists, otherwise 0], n is the set of all nodes in the network; d ij represents the shortest path length between node i and node j; M is the total number of module communities, k im is the number of edges connecting node i with other nodes belonging to module m, k i is the total number of edges connecting the node.

[0066] 5. Establishment of a normative model of brain networks

[0067] This paper uses the normative model construction method based on the Bayesian linear regression framework in the PCN toolkit, which can effectively depict the normal development trajectory of brain network topology characteristics. The specific implementation process is as follows:

[0068] (1) All brain topology phenotype data were preprocessed with Yeo-Johnson transformation to reduce the skewness of the data distribution. In terms of data set division, a stratified sampling strategy was adopted: 90% of the samples were randomly selected from the typical development control group of each research site as the training data set, and the remaining 10% of the typical development control samples together with all attention deficit hyperactivity disorder cases constituted the test data set, aiming to test whether the attention deficit hyperactivity disorder group showed a more significant degree of deviation compared with the typical development control group.

[0069] (2) During the model construction process, the present invention simultaneously considers the influence of demographic variables (age and gender) and multicenter effects. For the age variable, B-spline basis function (cubic spline function with 5 uniformly distributed nodes) is used for nonlinear fitting to better capture the dynamic changes of developmental trajectory. In order to deal with the possible non-Gaussian distribution characteristics of the response variable, a bijective nonlinear distortion function (Sinarcsinsh) is introduced to transform the non-Gaussian response variable into a Gaussian latent space. The site-related confounding effect is controlled by incorporating the site as a fixed effect variable into the model.

[0070] (3) The Powell algorithm was used to optimize the model’s hyperparameters. A normative model was fitted independently for each topological hubness metric of the 90 brain regions, and the regression expression was as follows:

[0071] Y=∫(X,θ)+E

[0072] where Y is the brain effect size, X is the covariate, θ is the relevant model parameter, and E is the residual.

[0073] The degree to which individuals in the test data set deviate from these normative expectations is quantified using the Z score, which is calculated as follows:

[0074]

[0075] For each subject n and each brain region d,y nd is the observed value, is the model prediction value, represents the prediction noise variance, represents the norm variance.

[0076] 6. Fusion of multidimensional brain network topology indicators

[0077] The present invention uses a method combining multimodal canonical correlation analysis with joint independent component analysis (mCCA+jICA) to perform data-driven fusion analysis on multiple topological features to reveal the covariant multimodal topological patterns. In the multimodal canonical correlation analysis (mCCA) stage, for different modes k, the topological feature X of each mode is k Perform projection and generate two typical variables: mixing coefficient matrix D k and association map C k . After obtaining the association map C kAfter that, the present invention uses joint independent component analysis (jICA) to process it and extract the largest independent source signals through the whitening matrix W. The optimal number of source signals is determined by the minimum description length criterion. This joint decomposition process produces modality-specific components and corresponding individual-specific mixing coefficients A k The relevant mathematical expressions are as follows:

[0078] mCCA stage: X k =D k C k (mCCA)

[0079] jICA stage: X k =A k S k ,C k =S k W -1 (jICA)

[0080] In order to realize the visualization of the results, the present invention converts the independent source signal S k The data were normalized to Z scores and the threshold of |Z|≥2.0 was used for spatial mapping. In terms of clinical relevance analysis, the mixing coefficient A was compared by two-sample test. k The differences between the groups were analyzed to identify the joint components with between-group differences, and the degree of association with clinical indicators was evaluated by product-moment correlation analysis.

[0081] 7. Biological subtype identification based on semi-supervised learning

[0082] This paper uses the semi-supervised HYDRA (Heterogeneity through discriminant analysis) algorithm to identify biological subtypes. This method constructs a convex polyhedron by establishing multiple hyperplanes that simultaneously enclose the control group distribution and optimally distinguish between the case groups, where each face represents a unique biological subtype. The specific implementation process is as follows:

[0083] (1) The individual deviation maps of the three topological features (i.e., the Z scores of each brain region) were used as input features, and age, gender, and scanning site were included as covariates in the HYDRA modeling. In terms of algorithm parameter settings, the following optimization configuration was adopted: 10-fold nested cross-validation, the regularization parameter was set to 0.25, L1 regularization was used, random hyperplanes were assigned through a determinant point process, 50 iterations were performed between the estimated hyperplane and the cluster, and 25 steps of cluster consistency test were performed. To evaluate the stability of the HYDRA clustering solution, the adjusted Rand index (ARI) was introduced to quantify the similarity of the clustering results in the cross-validation.

[0084] (2) In terms of robustness verification, the present invention designed a double validation analysis: First, the subtype stability was evaluated by permutation testing, specifically by randomly selecting 45% of the typical development control (TDC) sample as the real TDC group and 55% as the pseudo-patient group, repeating the HYDRA analysis 100 times, and generating a null distribution. These results were then compared with the clustering results of the real attention deficit hyperactivity disorder group: TDC and patient groups of the same size were assigned in the permutation test, and repeated 100 times to ensure consistent sample size. The ARI values ​​obtained in the two cases were compared by a two-sample t-test to determine the statistical significance of the k subtypes. Secondly, the present invention performed 20 random iterations of split-half cross-validation to evaluate the consistency of each subtype in the optimal k value clustering scheme. In this process, the stratification factors of diagnosis, gender, and site were also considered as covariates.

[0085] (3) In order to explore the clinical manifestation characteristics of the identified subtypes, Kruskal-Wallis analysis was used to examine the differences in symptom severity (such as attention deficit and hyperactivity / impulsivity phenotypes) among the subtypes, so as to achieve a dimensional characterization of clinical characteristics.

[0086] 8. External Validation of Biological Subtypes

[0087] To evaluate the generalization ability of the biological subtype clustering model, this paper uses a transfer learning strategy to apply the pre-trained model to an independent validation dataset. The specific implementation process is as follows:

[0088] (1) The T1-weighted magnetic resonance images in the validation dataset are processed in a process consistent with the prior samples until the normative model training stage. Since the validation database is completely absent in the initial training set, the present invention uses a transfer learning method to fine-tune the reference model. Specifically, typical developmental control samples in an independent validation dataset are used as adaptation data to adapt to possible site-specific variations. During the adaptation process, the hyperparameters of the reference model are used as information priors for the new dataset, and the mean and variance are adjusted in the latent Gaussian space, and the adjusted data are then mapped back to the original space. After recalibrating the reference norm range, a de-site effect estimate of the validation sample is generated to characterize the individual deviations of its topological features.

[0089] (2) By using the HYDRA algorithm parameters defined in the initial training phase (weight vector w i and the bias term b i ), reconstruct the polyhedron that best distinguishes the typical development control group from each biological subtype, and apply it to the validation data. The basic principle is to use the retained hyperplane weights and bias terms to calculate the expression score E of the validation cohort i(where X is the feature vector of the validation sample and D is the predicted subtype affiliation), the expression score calculation formula is:

[0090] E i =w i X+b i

[0091] D=argmax(E i )

[0092] (3) To assess the consistency of the validation subtypes with the reference subtypes, pattern similarity was quantified by pairwise product-moment correlation analysis, generating a correlation matrix. Cluster mapping relationships were determined using a greedy matching approach, mapping each validation cluster to the reference cluster with which it showed the highest correlation. Finally, the clinical feature profiles and neuroimaging deviation patterns of each subtype in the validation sample were simultaneously examined, following the analysis methods used in the discovery cohort.

[0093] The above specific embodiments are merely explanations of the present invention and are not limitations of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the embodiments as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A method for identifying subtypes of attention deficit hyperactivity disorder based on brain network topological hub deviation, characterized by: The method comprises the following steps: S1. Brain structural image quality control: The CAT12 toolbox was used to evaluate the accuracy of multiple quality dimensions of brain T1-weighted magnetic resonance images, and image quality ratings were obtained through weighted combination. Images with an IQR rating score lower than 3.5 were retained for subsequent analysis. S2. Image data preprocessing: For T1-weighted magnetic resonance images that have passed quality control, a standardized preprocessing process was performed using the CAT12 and SPM12 toolboxes to obtain the probability density distribution of brain gray matter volume; S3. Construction of a morphological similarity brain network matrix: Based on the automatic anatomically labeled brain atlas, the whole-brain image was divided into 90 brain regions. The Kullback-Leibler divergence morphological similarity measure was used to calculate the KL score between each pair of brain regions, thereby obtaining an individual-level brain structural network. Its mathematical expression is: KLS(P‖|Q)=∫p(x)log(p(x) / q(x))dx Where P and Q represent the gray matter volume distribution of the two brain regions, respectively, p(x) and q(x) are the corresponding probability density functions, and the integral range covers the support sets of the two distributions; Based on the calculated KLS values, a morphological similarity network matrix based on the automatic anatomically labeled brain atlas segmentation scheme was constructed; S4. Calculation of brain network topology hubness index: using degree centrality DC at the brain region level i , node efficiency NE i and participation coefficient PC i Three topological indicators characterize the hub characteristics of nodes in the network; The network was sparsified using the GRETNA toolbox, with the sparsity range set to 0.10 to 0.34 and a step size of 0.

01. The area under the curve of the topological characteristics of each node was calculated to obtain the multidimensional topological hub characteristics of the network connection group. The mathematical expressions of the above three hub indicators are as follows: where a ij Represents the connection between node i and node j. It is 1 when the connection (i, j) exists, otherwise it is 0. n is the set of all nodes in the network. d ij represents the shortest path length between node i and node j; M is the total number of module communities, k im is the number of edges connecting node i with other nodes belonging to module m, k i is the total number of edges connecting the node; S5. Establishment of a brain network norm model: A brain network norm model was constructed using the Bayesian linear regression framework in the PCNtoolkit toolkit. This model measures the deviation of individual patient hub characteristics based on the normal development trajectory of brain network topology characteristics. S6. Fusion of multi-dimensional brain network topology indicators: A data-driven fusion analysis of multimodal topological features was performed using a combination of multimodal canonical correlation analysis and joint independent component analysis. In the multimodal canonical correlation analysis stage, for different modes k, the topological features X of each mode are k Perform projection and generate two typical variables: mixing coefficient matrix D k and association map C k , joint independent component analysis was used to analyze C k The maximum independent source signal is extracted by whitening matrix W. The optimal number of source signals is determined by the minimum description length criterion. The joint decomposition process produces modal-specific components and corresponding individual-specific mixing coefficients A. k , the relevant mathematical expressions are as follows: Multimodal canonical correlation analysis stage: X k =D k C k Joint independent component analysis stage: X k =A k S k ,C k =S k W -1 The independent source signal S k The data were standardized to Z scores and spatially mapped using a threshold of |Z| ≥ 2.

0. In terms of clinical relevance analysis, the mixing coefficient A was compared using a two-sample test. k differences between the groups to identify joint components with differences between the groups and to assess the degree of association with clinical indicators by product-moment correlation analysis; S7. Biological subtype identification based on semi-supervised learning: The semi-supervised HYDRA algorithm is used to identify biological subtypes; S8. External validation of biological subtypes: A transfer learning strategy was used to apply the pre-trained model to an independent validation dataset.

2. The method for identifying ADHD subtypes based on brain network topology hub deviations according to claim 1, wherein: The quality indicators include image noise level, spatial uniformity of signal intensity, degree of head motion artifacts, and tissue segmentation.

3. The method for identifying ADHD subtypes based on brain network topology hub deviations according to claim 1, wherein: In step S2, the pre-processing process includes the following steps: S1. Perform tissue segmentation on the original T1-weighted magnetic resonance image based on the expectation maximization algorithm, accurately divide the brain tissue into three types: gray matter, white matter, and cerebrospinal fluid, and generate corresponding probability maps; S2. Spatial normalization was performed using the DARTEL algorithm to preserve local volume information and accurately align individual brain images to the MNI standard space. The normalization process involved modulating the image based on the Jacobian determinant and spatially smoothing the modulated gray matter image using a Gaussian kernel with a full-width at half-maximum of 8 mm. S3. Based on the automatic anatomical labeling atlas, the whole brain gray matter was divided into 90 anatomical regions, and the gray matter volume value at the voxel level of each region was extracted. The Kullback-Leibler divergence morphological similarity measurement method was used to calculate the KL score between each pair of brain regions to obtain the individual morphological similarity network.

4. The method for identifying ADHD subtypes based on brain network topology hub deviations according to claim 1, wherein: The steps of step S5 include the following steps: S501. Perform Yeo-Johnson transformation on the three brain topological hub phenotype data. Randomly select 90% of the samples from the developmental control group of the research site as the training data set, and the remaining 10% of the samples from the developmental control group of the research site together with all ADHD cases constitute the testing data set. S502. In the process of constructing the brain network norm model, B-spline basis function expansion is used for nonlinear fitting; a bijective nonlinear distortion function is used to transform the non-Gaussian response variable into a Gaussian latent space to process the non-Gaussian distribution characteristics of the response variable; Confounding effects were controlled by including site as a fixed-effect variable in the model; S503. The Powell algorithm is used to optimize the hyperparameters of the brain network norm model. A norm model is independently fitted for each topological hubness metric of the 90 brain regions. The regression expression is as follows: Y=∫(X,θ)+E Where Y is the brain effect size, X is the covariate, θ is the relevant model parameter, and E is the residual; the degree of deviation of individuals in the test data set from these normative expectations is quantified by the Z score, which is calculated as follows: For each subject n and each brain region d,y nd is the observed value, is the model prediction value, represents the prediction noise variance, represents the norm variance.

5. The method for identifying ADHD subtypes based on brain network topology hub deviation according to claim 1, wherein: The specific method of step S7 is as follows: S701. The deviation Z scores of the three topological hub features were used as input features, and age, sex, and scanning site were included as covariates in the HYDRA modeling. The algorithm parameters were 10-fold nested cross-validation, the regularization parameter was set to 0.25, L1 regularization was used, and the random hyperplane of the determinant point process was assigned. 50 iterations were performed between the estimated hyperplane and the clusters, and a 25-step cluster consistency test was performed. The adjusted Rand index was introduced to quantify the similarity of the clustering results in the cross-validation. S702. Double validation analysis: Subtype stability was assessed by permutation testing. The null distribution was generated by randomly selecting 45% of the typical developmental control group as the true developmental control group and 55% as the pseudo-patient group, and repeating the HYDRA analysis 100 times. These results were then compared with the clustering results of the true ADHD group. The permutation test was repeated 100 times to ensure consistent sample size by assigning the control and patient groups to the same size. The ARI values ​​obtained in the two conditions were compared by two-sample t-test to determine the statistical significance of the k subtypes. The consistency of the subtypes in the optimal k-value clustering scheme was assessed by performing split-half cross-validation with 20 random iterations. S703. Kruskal-Wallis analysis was used to examine the differences in symptom severity among subtypes and quantitatively characterize the clinical characteristics of core symptoms.

6. The method for identifying ADHD subtypes based on brain network topology hub deviations according to claim 1, wherein: The specific method of step S8 is as follows: S801. The T1-weighted magnetic resonance images in the validation dataset are processed in the same manner as the prior samples until the norm model training phase. The reference model is fine-tuned using a transfer learning method, and samples in the independent validation dataset are used as adaptation data to adapt to site-specific variations. During the adaptation process, the hyperparameters of the reference model are used as information priors for the new dataset, and the mean and variance are adjusted in the latent Gaussian space. The adjusted data are then mapped back to the original space. After recalibrating the reference norm range, a de-site effect estimate of the validation sample is generated to characterize the individual deviations of its topological features. S802, reconstructing a polyhedron that optimally distinguishes the developmental control group from each biological subtype by using the HYDRA algorithm parameters defined in the initial training phase, and applying the polyhedron to the validation data; The expression score E for the validation cohort is calculated using the retained hyperplane weights and bias terms. i , the calculation formula of expression score is: E i =w i X+b i D=argmax(E i ) S803. Quantify pattern similarity through pairwise product-moment correlation analysis, generate a correlation matrix, and verify the consistency of subtypes with reference subtypes; use a greedy matching method to determine the cluster mapping relationship, that is, map each validation cluster to the reference cluster with which it shows the highest correlation; finally, according to the analysis method used in the discovery cohort, simultaneously examine the clinical feature spectrum and neuroimaging deviation pattern of each subtype in the validation sample.

Images