Attention deficit hyperactivity disorder subtype recognition and evolution analysis method based on comparative learning

Abstract

The invention provides an attention deficit hyperactivity disorder (ADHD) subtype recognition and evolution analysis method based on comparative learning, which comprises the following steps: acquiring brain functional magnetic resonance imaging data of an ADHD patient and a healthy person, and extracting low-frequency amplitude and functional connection multi-modal data from the brain functional magnetic resonance imaging data; inputting the multi-modal data into a feature coding module to generate potential feature representation of each modal; inputting the potential feature representation into a multi-view learning module, generating a projection semantic feature and a shared semantic feature, and minimizing the difference between the projection semantic feature and the shared semantic feature; performing hierarchical multi-agent regularization constraint on the shared semantic features; inputting the shared semantic features into a classifier for classification and recognition; and carrying out dimension reduction on the shared semantic features to obtain feature evolution trajectories from healthy people to different subtypes. According to the method, healthy people and disease subtypes can be identified, the characteristic dynamic evolution process from the healthy people to the subtypes is modeled, the neurological function heterogeneity of the disease is revealed, and a basis is provided for personalized precise diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image data processing and auxiliary diagnosis of neurological diseases, and particularly relates to a subtype recognition and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning. BACKGROUND

[0002] Attention Deficit Hyperactivity Disorder (ADHD) is a common childhood neurodevelopmental disorder, and its core features are difficulties in attention maintenance and behavioral control. In clinical practice, ADHD is usually divided into three subtypes according to the main symptoms: attention deficit (ADHD-I), hyperactivity and impulsivity (ADHD-HI), and mixed type (ADHD-C). Therefore, analyzing the heterogeneity between different subtypes of ADHD has important research significance for promoting personalized diagnosis and precise treatment.

[0003] In recent years, there has been some exploration of ADHD subtype recognition in existing research. For example, patent [1] (Mu Shuhua, Cheng Qianxi. A method and device for distinguishing different subtypes of attention deficit hyperactivity disorder [P]. Guangdong Province: CN120661137A, 2025-09-19.) uses visual and auditory behavior data to distinguish ADHD subtypes; patent [2] (Jiang Xinlong, Chen Yiqiang, Zheng Yi, Qiao Yuxi, Huang Wuliang, Zhou Yuming. Method and system for identifying attention deficit hyperactivity disorder subtypes [P]. Beijing: CN113160967A, 2021-07-23.) uses a trained graph neural network model to classify ADHD subtypes based on speech information. However, these methods are mostly limited to static subtype classification and lack the depiction of the evolution process from healthy individuals to different subtypes, and also fail to reveal the evolution rules between subtypes.

[0004] Currently, the evolution analysis of the heterogeneity of mental illness subtypes has been attempted on other mental disorders. For example, the literature [3] (Jiang Y, Wang J, Zhou E, et al. Neuroimaging biomarkers define neurophysiological subtypes with distinct trajectories in schizophrenia [J]. Nature Mental Health, 2023, 1 (3): 186-199.) based on cross-sectional functional magnetic resonance imaging (fMRI) data, the model can define neurophysiological subtypes with different trajectories in schizophrenia. This method shows that understanding the heterogeneity of the phenotype and the time dimension at the same time helps to more fully characterize the disease characteristics. It is worth noting that although this model can effectively track the dynamic evolution of the disease in brain structure, its applicability in the study of ADHD subtypes is limited because ADHD is a neurodevelopmental disorder and does not accompany significant brain structure abnormalities.

[0005] To solve the above problems, the present application extracts cross-modal shared semantic information from multiple modalities of fMRI data and guides feature evolution through a hierarchical multi-agent strategy, which can model the dynamic evolution process of ADHD subtypes while identifying them, thereby more fully revealing the neural functional heterogeneity of ADHD. SUMMARY

[0006] The purpose of the present application is to solve the problem that the prior art only makes static classification of attention deficit hyperactivity disorder subtypes and healthy people, and cannot reveal the feature evolution of healthy people and each subtype, as well as between each subtype. In order to achieve the above purpose, the present application provides an attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrast learning.

[0007] The embodiment of the present application adopts an attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrast learning, which comprises the following steps:

[0008] Step S1: data acquisition and preprocessing: obtaining a brain functional magnetic resonance imaging (functional magnetic resonance imaging, fMRI) data set containing attention deficit hyperactivity disorder subtype patient individuals and healthy control individuals, wherein each individual in the data set is regarded as a sample; extracting at least multiple modalities of data including amplitude of low frequency fluctuations (Amplitude of Low Frequency Fluctuations, ALFF) and functional connectivity (Functional Connectivity, FC) from the fMRI data of the data set;

[0009] Step S2: latent feature representation generation: input the multi-modal data into a feature encoding module to generate a latent feature representation corresponding to each modality;

[0010] Step S3: multi-view semantic alignment: project the latent feature representation of each modality to a shared semantic space through a multi-view learning module to obtain a projected semantic feature of each modality, simultaneously generate a shared semantic feature in the shared semantic space, and obtain a distribution-aligned shared semantic feature by minimizing the difference between the projected semantic feature and the shared semantic feature;

[0011] Step S4: hierarchical multi-agent regularization: adopt a hierarchical multi-agent learning strategy to regularize and constrain the distribution-aligned shared semantic feature in the shared semantic space, so that the distribution-aligned shared semantic features of individuals of the same category converge to the category center, and different categories and different subtypes remain separated;

[0012] Step S5: classification and identification: based on the distribution-aligned shared semantic feature after regularization constraint, use a classifier to classify and identify attention deficit hyperactivity disorder subtypes and healthy individuals;

[0013] Step S6: evolution trajectory visualization: dimension reduction is performed on the distribution-aligned shared semantic feature to visualize the latent feature evolution trajectory from healthy controls to different attention deficit hyperactivity disorder subtypes.

[0014] In a feasible implementation manner of the above method, the feature encoding module includes an encoder, a sampling unit, and a decoder, wherein:

[0015] The encoder is configured to convert the input multi-modal data into a mean value and a variance corresponding to the distribution of each modality data;

[0016] The sampling unit samples the mean value and the variance to generate a latent feature representation, and minimizes the difference between the mean value and the variance and the parameters of a standard Gaussian distribution through a KLD (Kullback-Leibler divergence) divergence loss function, so that the latent feature representation is subject to a smooth and continuous distribution as a whole;

[0017] The decoder is configured to reconstruct the latent feature representation into the input multi-modal data, and output the reconstructed multi-modal data, and minimize the difference between the reconstructed multi-modal data and the input multi-modal data through a MSE (Mean Squared Error) loss function.

[0018] In a feasible implementation manner of the above method, the multi-view learning module includes a projector and a shared semantic encoder, wherein:

[0019] The projector is configured to map the latent feature representation of each modality into a projected semantic feature;

[0020] The shared semantic encoder is configured to generate a shared semantic feature in the shared semantic space, and the process of generating the shared semantic feature comprises:

[0021] An initialization step: initializing a shared semantic feature vector;

[0022] An encoding step: inputting the initialized shared semantic feature into the shared semantic encoder to generate the mean and variance of the feature distribution;

[0023] A constraint step: using a KLD divergence loss function to minimize the difference between the mean and variance and the standard Gaussian distribution parameters to constrain and update the shared semantic feature; at the same time, minimizing the difference between the projected semantic feature and the shared semantic feature to obtain a distribution-aligned shared semantic feature;

[0024] By iteratively performing the encoding step and the constraint step, the shared semantic feature is finally obtained.

[0025] In a feasible implementation manner of the above method, the hierarchical multi-agent strategy comprises:

[0026] Setting different category levels of agents in the shared semantic space, including main agents representing two large category levels of health and disease and sub-agents representing subtype levels, wherein the main agents include two, which are a main agent representing the central features of healthy human and a main agent representing the central features of attention deficit hyperactivity disorder patient class, and the sub-agents include sub-agents representing the central features of each class of attention deficit hyperactivity disorder subtypes, wherein the each class of subtypes refers to the classes of each different subtype, and the number of sub-agents is equal to the number of subtype classes.

[0027] In a feasible implementation manner of the above method, the hierarchical multi-agent strategy further comprises:

[0028] Setting a contrast loss function to regularize and constrain the shared semantic features in the shared semantic space, so that the shared semantic features of individuals of the same category converge to the category center, and different categories and different subtypes remain separated.

[0029] In a feasible implementation manner of the above method, the contrast loss function L con The sample-agent contrast loss L sp And the agent-agent contrast loss L pp is weighted, that is

[0030] L con = L sp + λLpp

[0031] In the formula, λ is the weighting coefficient that balances the two losses;

[0032] The formula for calculating the sample-proxy contrast loss is as follows:

[0033]

[0034] In the formula z i Let N represent the shared semantic features of the i-th sample, N be the total number of samples, τ1 be the temperature coefficient, d(·) be the distance metric function, and p be the distance metric function. + With p - These represent positive and negative proxies, respectively; the positive proxy corresponds to z. i The negative proxy represents the center of the class to which it belongs, while the negative proxy represents the center of the class to which it does not belong; if z i If the class belongs to the Attention Deficit Hyperactivity Disorder (ADHD) patient category, then the class not belonging to this category is healthy human, and vice versa; if z i If a subtype belongs to a certain category of attention deficit hyperactivity disorder (ADHD), then other categories that do not belong to the ADHD subtype are also considered, and vice versa.

[0035] The formula for calculating the proxy-proxy comparison loss is as follows:

[0036]

[0037] In the formula p i Let p represent the i-th sub-agent, P represent the set of all sub-agents, |P| represent the number of sub-agents, and τ2 be the temperature coefficient; A P represents the principal agent representing the class central characteristics of patients with attention deficit hyperactivity disorder. * Indicates that except for the current sub-agent p i The collection of all other agents, P represents * The kth agent in the process.

[0038] The technical solutions proposed in the embodiments of the present invention can achieve at least the following technical effects:

[0039] The goal is to classify and identify healthy individuals and various subtypes of ADHD. More importantly, this classification can reveal the evolution of characteristics from healthy individuals to various subtypes of ADHD, as well as between different subtypes of ADHD. These characteristics include symptoms, brain function, and brain regions. This evolution of characteristics can help doctors understand the heterogeneity between subtypes and provide a basis for precise and personalized diagnosis and treatment for each subtype. Attached Figure Description

[0040] Figure 1A flow chart of a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0041] Figure 2 A constituent diagram of a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0042] Figure 3 A feature encoder and projector structure corresponding to FC data in a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0043] Figure 4 A feature encoder and projector structure corresponding to ALFF data in a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0044] Figure 5 A shared semantic encoder structure in a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0045] Figure 6 A classifier structure in a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning provided by an embodiment of the application is shown.

[0046] Figure 7 An effect diagram of symptom feature evolution processes of healthy people and attention deficit hyperactivity disorder subtypes of an embodiment of the application is shown.

[0047] Figure 8 An effect diagram of symptom feature correlations of attention deficit hyperactivity disorder subtypes and healthy people of an embodiment of the application is shown.

[0048] Figure 9 An effect diagram of significant functional brain region evolution processes of healthy people and attention deficit hyperactivity disorder subtypes of an embodiment of the application is shown. DETAILED DESCRIPTION

[0049] The application will be further described below in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0050] Referring to Figure 1 The application provides a subtype identification and evolution analysis method for attention deficit hyperactivity disorder based on contrast learning, including the following steps:

[0051] Step S1: data acquisition and preprocessing: obtaining a brain functional magnetic resonance imaging (fMRI) dataset containing patients with attention deficit hyperactivity disorder subtypes and healthy control individuals, wherein each individual in the dataset is regarded as a sample; extracting multi-modal data including at least Amplitude of Low Frequency Fluctuations (ALFF) and Functional Connectivity (FC) from the fMRI data of the dataset;

[0052] Specifically, during the acquisition of brain functional magnetic resonance data, the subject's head movement, external magnetic field, unstable scanning equipment, and machine noise are affected by various factors, which may lead to different brain activation records, thereby affecting the classification accuracy. Therefore, before analyzing the brain functional magnetic resonance data, the data must be preprocessed. The specific process mainly includes:

[0053] (1) Initial data rejection: due to the non-uniformity of the scanner magnetic field and the influence of the subject's initial adaptation stage, the signals of the first 4 time points are not stable and the data reliability is low, so they need to be removed.

[0054] (2) Head motion correction: first, the acquired fMRI image sequence is corrected for head motion to eliminate the effects caused by the subject's head movement during scanning. This is usually achieved by spatially registering each time point image in the sequence.

[0055] (3) Time correction: if there is a time offset or misalignment during acquisition, the time series needs to be corrected to ensure that the data at each time point corresponds to the same time period.

[0056] (4) Spatial smoothing: to reduce noise in the image and improve signal-to-noise ratio. The data needs to be smoothed, which is achieved by weighted averaging the values of each voxel and its surrounding neighborhood in the brain image. Common smoothing methods include Gaussian filtering, which preserves the spatial resolution of the image and reduces noise. The main purpose of spatial smoothing is to enhance the smoothness of the image, making subsequent statistical analysis more reliable.

[0057] (5) Spatial standardization: aligning the scanned brain anatomy with the standard brain image space for comparing data between different subjects or with previous research results. Using the Montreal Neurological Institute (MNI) space for registration at 4x4x4 voxel resolution.

[0058] After removing various noises in the data through the above data processing, the biological signal data of different modalities of each sample is obtained, including low-frequency fluctuation amplitude (ALFF) and functional connectivity (FC), wherein:

[0059] The FC data is constructed in the following manner. First, the pre-processed data is segmented into different brain regions according to the AAL-116 (Automated Anatomical Labeling-116, AAL-116) atlas, and time series data is extracted from each brain region. Generally, each brain region contains multiple voxels, and the average signal is generally used:

[0060]

[0061] wherein, is the average signal of the i-th brain region at time t, X ik is the signal of the k-th voxel in the i-th brain region at time t, N i is the total number of voxels in the i-th brain region. Second, the time series correlation between brain regions is calculated to form the FC matrix R. The correlation is usually calculated using the Pearson correlation coefficient:

[0062]

[0063] wherein r ij represents the Pearson correlation coefficient of the i-th and j-th brain regions, T represents the total length of the time series, and respectively represent the average values of and over the entire time length T. The obtained correlation coefficient r ij is filled into the corresponding position of the FC matrix R, wherein R ij = r ij .

[0064] The ALFF data is obtained in the following manner. The Fourier transform FFT is performed on the extracted average time series of each brain region to obtain the frequency domain representation

[0065]

[0066] wherein X i (f) is the Fourier transform result of the time series of the i-th brain region at frequency f; then the power spectrum of each brain region in a specific low-frequency range (0.01-0.08 Hz) is calculated, and the square root is taken to obtain the ALFF value of the brain region:

[0067]

[0068] wherein ALFF iALFF value of the i-th brain region.

[0069] Step S2: latent feature representation generation: input the multi-modal data into a feature encoding module to generate a latent feature representation corresponding to each modality;

[0070] Specifically, as shown in the figure, Figure 2 the ALFF and FC modal data obtained through step S1 are respectively input into the feature encoding module, and each modality corresponds to a feature encoding module based on a variational autoencoder structure, including an encoder, a sampling unit and a decoder, wherein the FC data feature encoding module structure is as shown in the figure, Figure 3 The encoder consists of 3364x2000 Dense+ReLU activation function+2000x1000 Dense+ReLU activation function+1000x500 Dense+ReLU activation function+500x100 Dense+ReLU activation function+100x30 Dense in turn; the decoder structure is opposite to the encoder structure, and the corresponding decoder structure is 30x100 Dense+ReLU activation function+100x500 Dense+ReLU activation function+500x1000 Dense+ReLU activation function+1000x2000 Dense+ReLU activation function+2000x3364 Dense; the ALFF data feature encoder structure is as shown in the figure, Figure 4 The encoder consists of 3364x2000 Dense+ReLU activation function+2000x1000 Dense+ReLU activation function+1000x500 Dense+ReLU activation function+500x100 Dense+ReLU activation function+100x30 Dense in turn; the decoder structure is opposite to the encoder structure, and the corresponding decoder structure is 30x100 Dense+ReLU activation function+100x500 Dense+ReLU activation function+500x1000 Dense+ReLU activation function+1000x2000 Dense+ReLU activation function+2000x3364 Dense; the ALFF data feature encoder structure is as shown in the figure,

[0071] The encoder is used to convert the input multi-modal data x m，i into the mean value μ m,i and variance σ m，i of the distribution of the corresponding modal data, where m∈{ALFF, FC}, i represents the i-th sample, μ m,i and σ m,i The output dimension is 30x1;

[0072] The sampling unit samples the mean value and variance to generate a latent feature representation Z m,i , and the sampling specific method is:

[0073] z m，i = μ m,i+ ε⊙σ m,i

[0074] where ε denotes random noise sampled from a standard normal distribution, ⊙ is an element-wise multiplication operation, and z is the high-level feature m，i with output dimension 30x1; and the mean and variance are minimized by a Kullback-Leibler divergence (KLD) loss function to minimize the difference between the latent feature representation z m,i and the parameters of a standard Gaussian distribution, resulting in a smooth continuous distribution over the whole space, where the KLD loss function is expressed as:

[0075]

[0076] where N is the total number of samples; the KLD loss function constrains the latent feature to be optimized by backpropagation during training, enabling stable feature learning:

[0077] The decoder is used to reconstruct the input multi-modal data from the latent feature representation z m,i and outputs the reconstructed multi-modal data, with the difference between the reconstructed multi-modal data and the input multi-modal data minimized by a mean squared error (MSE) loss function, where the MSE loss function is expressed as:

[0078]

[0079] where, denotes the reconstructed data corresponding to x m,i , and ||·||2 denotes the 2-norm operation.

[0080] Step S3: Multi-view semantic alignment: The latent feature representation of each modality is projected into a shared semantic space by a multi-view learning module to obtain the projected semantic features of each modality, and a shared semantic feature is generated in the shared semantic space, and the distribution-aligned shared semantic feature is obtained by minimizing the difference between the projected semantic features and the shared semantic feature;

[0081] Specifically, as shown in Figure 2 , the multi-view learning module includes a projector and a shared semantic encoder, where the projector is a single fully connected layer, as shown in Figure 3 and 4 Each modality projector maintains consistency, with a structure of 30x30Dense, ensuring that the high-level features of each modality are mapped to the same dimensional semantic space; the network structure of the shared semantic encoder is as shown in Figure 5The shown is: 30x25Dense+ReLU activation function+25x20Dense+ReLU activation function+20x15Dense, 2-norm regularization constraint is used in each Dense layer, the model complexity is controlled, and overfitting is prevented.

[0082] The projector is used for projecting the latent feature representation of each modality into a shared semantic space, and mapping into a projected semantic feature, and the specific operation is that the latent feature representation z m,i of each modality is input into the projector, and the output is a projected semantic feature z' m，i ;

[0083] The shared semantic encoder is used for generating a shared semantic feature in the shared semantic space, and the process of generating the shared semantic feature includes:

[0084] An initialization step: initializing a shared semantic feature vector z zi for each sample (such as the i-th sample);

[0085] An encoding step: inputting the initialized shared semantic feature z i into the shared semantic encoder to generate the mean μ zi and variance σ zi of the feature distribution;

[0086] A constraint step: using a KLD divergence loss function to minimize the difference between the mean μ zi and variance σ zi and the standard Gaussian distribution parameters, so as to constrain and update the shared semantic feature; the KLD divergence loss function is:

[0087]

[0088] The KLD loss function maximizes the mutual information across modalities while maintaining modality consistency, thereby improving the discriminability of the features. At the same time, the difference between the projected semantic feature z' m，i and the shared semantic feature z i is minimized to obtain a distribution-aligned shared semantic feature, wherein the minimization difference loss function is:

[0089]

[0090] The loss function aligns the distribution of the high-level features z' m,i mapped by each modality and the shared semantic feature z i , so as to constrain the distribution consistency of different modalities in the shared space;

[0091] By iteratively performing the above encoding step and constraint step, the shared semantic feature z is finally obtained i .

[0092] Step S4: Hierarchical multi-agent regularization: using a hierarchical multi-agent learning strategy to constrain the distribution-aligned shared semantic features in the shared semantic space, so that the distribution-aligned shared semantic features of individuals of the same class converge to the class center, and different classes and different subtypes remain separated;

[0093] Specifically, as shown in Figure 2 , the hierarchical multi-agent strategy includes setting different class level agents in the shared semantic space, including main agents representing two large class levels of health (HC) and disease (ADHD) and sub-agents representing sub-type levels, wherein the main agent includes two, which are a main agent p H representing the health human center feature and a main agent p A representing the attention deficit hyperactivity disorder patient class center feature; the sub-agent includes a sub-agent representing the center feature of each class of the attention deficit hyperactivity disorder subtype, and the number of sub-agents is equal to the number of subtype classes. In this embodiment, considering that the ADHD subtypes in the clinic are mainly ADHD-I and ADHD-C subtypes, and the number of ADHD-HI subtype patients is small, and the patients of this subtype have strong self-healing after adulthood, therefore, the subtype classes are divided into ADHD-I and ADHD-C, and their sub-agents are set as p I and p C .

[0094] The hierarchical multi-agent strategy also includes setting a contrast loss function to constrain the shared semantic features in the shared semantic space, so that the shared semantic features of individuals of the same class converge to the class center, and different classes and different subtypes remain separated.

[0095] The contrast loss function L con is composed of a sample-agent contrast loss L sp and an agent-agent contrast loss L pp , that is

[0096] L con = L sp + λL pp

[0097] Wherein λ is the weight coefficient for balancing the two losses;

[0098] The sample-agent contrast loss constrains the shared semantic feature z iThe distance between a given individual and its corresponding primary agent promotes the aggregation of similar individual characteristics towards their respective category centers and keeps them away from dissimilar categories. The calculation formula is as follows:

[0099]

[0100] In the formula, τ1 is the temperature coefficient, d(·) is the distance metric function, and p + With p - These represent positive and negative proxies, respectively; the positive proxy corresponds to z. i The negative proxy represents the center of the class to which it belongs, while the negative proxy represents the center of the class to which it does not belong; if z i If the class belongs to the Attention Deficit Hyperactivity Disorder (ADHD) patient category, then the class not belonging to this category is healthy human, and vice versa; if z i If a sample belongs to a certain subtype of attention deficit hyperactivity disorder, then other subtypes that do not belong to the same subtype of attention deficit hyperactivity disorder, and vice versa; Equation (1) optimizes and forces the shared semantic features of each sample to move closer to the agent of the correct category, while moving away from the agent of the wrong category.

[0101] The agent-agent contrastive loss, by constraining the relative relationships between sub-agents and the main agent, as well as among sub-agents, ensures that different subtypes remain distinguishable within a category, thereby simultaneously enhancing both inter-category separability and intra-category subtype differentiation. Its calculation formula is as follows:

[0102]

[0103] In the formula p i Let P represent the i-th sub-agent, where P = {p I p C} represents the set of all sub-agents, |P|=2 represents the number of sub-agents, and τ2 is the temperature coefficient; P * Indicates that except for the current sub-agent p i The collection of all other agents, P represents * The kth agent in the equation. Equation (2) strengthens the clustering characteristics of the subtype by adjusting the distance between the sub-agent and other agents, and effectively promotes the modeling of the feature evolution between the subtype and the healthy control group.

[0104] Step S5: Classification and Identification: Based on the shared semantic features aligned by the distribution after regularization constraints, a classifier is used to classify and identify the attention deficit hyperactivity disorder subtype and healthy individuals;

[0105] Specifically, the classifier network consists of multiple fully connected layers and stacked residual blocks, used to process the shared semantic features z. i Perform classification. The structure of this classifier is as follows: Figure 6As shown: 30x20Dense+ReLU activation function+20x10Dense; on this basis, two residual blocks are connected in series, each residual block includes two full connection layers (10x10Dense+BatchNormalization+ReLU activation function), and the residual stacking of input and output is realized through a jump connection; the output layer is 10x3Dense, and a Softmax activation function is used to obtain a classification result, that is, the label of the input sample is determined as ADHD-I or ADHD-C or a healthy person. The classifier is optimized by using a cross-entropy loss function, and the expression of the cross-entropy loss function is

[0106]

[0107] Where y c，i represents the true label value of the fth sample belonging to the cth class, and the value is 0 or 1; represents the probability that the ith sample belongs to the cth class; C represents the total number of classes of classification, and in this embodiment, C=3, that is, three classes of healthy people, ADHD-I and ADHD-C.

[0108] Step S6: Evolution trajectory visualization: dimension reduction is performed on the distribution-aligned shared semantic features to visualize the potential feature evolution trajectory from healthy controls to different attention deficit hyperactivity disorder subtypes.

[0109] Specifically, the diffusion mapping method is used to reduce the dimension of the shared semantic features z i . Diffusion mapping converts the 'diffusion distance' between data points into Euclidean distance in a new low-dimensional space by simulating the process of random walk, thereby ingeniously revealing the intrinsic and nonlinear geometric structure of the data. The diffusion mapping dimension reduction step includes:

[0110] (1) Construct a similarity matrix (construct a graph) that describes the local neighborhood relationship of the data

[0111] First, the similarity W i between any two shared semantic features z j is calculated using a Gaussian kernel function (also known as a radial basis function). ij

[0112]

[0113] Where σ is the variance of the Gaussian kernel, used to control the neighborhood width. In the embodiment of the present application, 2σ 2 =30 is taken, which is set to be consistent with the dimension of the shared semantic feature z i , so as to match the distribution range of the feature space and obtain a reasonable similarity description.

[0114] (2) Construct normalized transition matrix (define random walk)

[0115] Transform the similarity matrix into a probability transition matrix. First, calculate the degree matrix (D): the degree matrix D is a diagonal matrix, the elements on the diagonal line D ii =∑ j W ij , represent the total strength of all connections of the i-th node; second, construct the transition matrix (P): normalize the matrix W by degree P=D -1 W, the elements P ij in the matrix P represent the probability of one-step random walk from node i to node j.

[0116] (3) Eigen decomposition

[0117] Eigen decomposition of the transition matrix Pψ=λψ, get the eigenvalue λ0≥λ1≥λ2≥… and the corresponding eigenvectors ψ0, ψ1, ψ2, ….

[0118] (4) Define diffusion coordinates (dimensionality reduction mapping)

[0119] The final dimensionality reduction representation (diffusion mapping) is composed of the first two largest non-trivial eigenvectors. For the i-th shared semantic feature z i , its coordinates in the new 2-dimensional space are Ψ(i) = (λ'1ψ1(i), λ'2ψ2(i)), where ψ k (i) is the i-th component of the k-th eigenvector, t is the "diffusion time", which is a hyperparameter, and can be understood as the number of random walks, increasing t can observe the structure of the data on a larger scale, ignoring more subtle noise.

[0120] After the above dimensionality reduction steps, the potential feature evolution trajectory from healthy people to ADHD-I and ADHD-C subtypes is obtained.

[0121] To further illustrate the beneficial effects of the technical solutions of the present application, the above examples are taken as examples to illustrate from the following aspects.

[0122] First, to explore the evolutionary path from healthy controls (HC) to the attention deficit hyperactivity disorder (ADHD) subtype, the shared semantic feature z needs to capture sufficient classification information. Therefore, subtype classification performance is used to partially verify the effectiveness of this technical solution. Four ADHD datasets (NYU, PU, ​​KKI, and NI) provided by the ADHD-200 consortium (http: / / foon_1000.projects.nitrc.org / indi / adhd200) were used, and 10-fold cross-validation was performed. Statistical classification results were obtained after 50 rounds of computation. Table 1 summarizes the classification performance metrics. The NI dataset was not classified separately due to the imbalanced subtype sample size. As shown in Table 1, the technical solution of this invention achieved an average classification accuracy of 67.6%, and exhibited acceptable accuracy fluctuations across different datasets, indicating good robustness. Table 2 further provides a comparison of classification performance with existing methods, where the accuracy of the listed methods ranges from 60% to 76%. Although the main objective of this invention is to reveal the evolutionary patterns of ADHD subtypes, the classification performance obtained is still comparable to that of existing methods.

[0123] Secondly, we obtained the potential trajectories of trait evolution from healthy individuals to ADHD-I and ADHD-C subtypes, such as... Figure 7 As shown, each point represents a sample, clearly demonstrating the distinct distribution characteristics of the three groups: healthy controls (HC), attention deficit-predominant ADHD (ADHD-I), and mixed ADHD (ADHD-C). Furthermore, it clearly exhibits three different evolutionary trajectory patterns: HC to ADHD-I, HC to ADHD-C, and ADHD-I to ADHD-C. To further verify the correlation between these trajectories and symptom change trends, a gradient color mapping was used to project the three categories of symptom scores (attention deficit score, hyperactivity-impulsivity score, and ADHD prevalence score) onto the evolutionary trajectories, as shown below. Figure 8 The results show that samples closer to the HC group exhibited lower symptom scores, while samples farther from the HC group showed a significant upward trend in scores. This pattern confirms that the evolutionary trajectory obtained by the technical solution of this invention can not only effectively identify healthy individuals and ADHD subtypes, but also that the evolutionary trajectory is consistent with the symptom change trends of each category.

[0124] Furthermore, to further verify the relevance of the obtained evolutionary trajectory to biology, a biomarker analysis was performed on the ALFF values ​​on the PU dataset. Figure 9The evolution trajectories of three biomarkers between healthy controls (HC) and two ADHD subtypes are shown, in which four transition states reflect the dynamic changes of brain activities. Five key brain regions with the most significant ALFF value variations are marked in each trajectory. It can be observed that the ALFF values of most brain regions show a downward trend when transitioning from HC to subtypes, which indicates insufficient activation of the relevant brain regions, which is consistent with the characteristic of ADHD as a neurodevelopmental disorder accompanied by delayed brain development. Specifically, in the process of transforming from HC to attention deficit dominant type (ADHD-I), the ALFF values of the left angular gyrus (ANG.L) and bilateral olfactory cortex (OLF) decrease, which constitutes the neural basis of attention maintenance dysfunction and decreased reward sensitivity, thereby leading to the symptoms of inattention and hypomotivation specific to this subtype. In contrast, the transformation path from HC to mixed type (ADHD-C) shows a decrease in the ALFF values of the left superior medial orbitofrontal cortex (ORBsupmed L) and the posterior cingulate cortex (PCG L), reflecting impaired impulse control and sensory motor regulation functions, which in turn drive the impulsivity and hyperactivity symptoms associated with ADHD-C. During the transformation between ADHD-I and ADHD-C, a more complex change pattern is presented, with the ALFF values of brain regions such as the left amygdala (AMYG L) and the precentral gyrus (PoCG L) increasing, indicating enhanced emotional reactivity and motor control disorders, which promote the co-occurrence of attention deficit and hyperactivity-impulsivity symptoms. These rules confirm that the evolution trajectories obtained by the technical solution of the present application can effectively identify biomarkers for distinguishing ADHD subtypes, and are consistent with existing neurobiological mechanisms.

[0125] The above only describes the preferred embodiments of the present application. It should be noted that for those skilled in the art, without departing from the technical principles of the present application, several improvements and modifications can be made, which should also be considered as the protection scope of the present application.

[0126] Table 1 Classification performance of the embodiments of the present application on different data sets

[0127]

[0128] Table 2 Comparison of subtype classification performance between the embodiments of the present application and other existing methods

[0129]

[0130] [1] Qureshi M N I, Min B, Jo HJ, and Lee B, “Multiclass classification for the differential diagnosis on the ADHD subtypes using recursive feature elimination and hierarchical extreme learning machine: Structural MRI study,” PloS one, vol. 11, n0.8. pp. e0160697, 2016.

[0131] [2] Qureshi M N I, Oh J, Min B, Jo H J, and Lee B, “Multimodal, multi-measure, and multi-class discrimination of ADHD with hierarchical feature extraction and extreme learning machine using structural and functional brain MRI,” Frontiers in human neuroscience, vol. 11, pp. 157, 2017.

[0132] [3] Saha P and Sarkar D, “Characterization and classification of ADHD subtypes: An approach based on the nodal distribution of eigenvector centrality and classification tree model,” Child Psychiatry & Human Development, pp. 1-13, 2022.

Claims

1. A contrastive learning based attention deficit hyperactivity disorder subtype identification and evolution analysis method, characterized in that The method comprises the following steps: Step S1: data acquisition and preprocessing: acquiring a brain functional magnetic resonance imaging (fMRI) dataset containing attention deficit hyperactivity disorder subtype patient individuals and healthy control individuals, wherein each individual in the dataset is regarded as a sample; extracting multi-modal data including at least amplitude of low frequency fluctuations (ALFF) and functional connectivity (FC) from fMRI data of the dataset; Step S2: latent feature representation generation: inputting the multi-modal data into a feature encoding module to generate latent feature representations corresponding to each modality; Step S3: multi-view semantic alignment: projecting the latent feature representations of each modality to a shared semantic space through a multi-view learning module to obtain projected semantic features of each modality, simultaneously generating a shared semantic feature in the shared semantic space, and obtaining a distribution-aligned shared semantic feature by minimizing the difference between the projected semantic features and the shared semantic feature; Step S4: hierarchical multi-agent regularization: using a hierarchical multi-agent learning strategy to regularize and constrain the distribution-aligned shared semantic feature in the shared semantic space, so that the distribution-aligned shared semantic features of individuals of the same category converge to the category center, and different categories and different subtypes remain separated; Step S5: classification and identification: based on the distribution-aligned shared semantic feature after regularization constraint, using a classifier to classify and identify attention deficit hyperactivity disorder subtypes and healthy individuals; Step S6: evolution trajectory visualization: dimensionality reduction is performed on the distribution-aligned shared semantic feature to visualize the latent feature evolution trajectory from healthy controls to different attention deficit hyperactivity disorder subtypes.

2. The attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrastive learning according to claim 1, characterized in that, The feature encoding module comprises an encoder, a sampling unit and a decoder, wherein: The encoder is used to convert the input multi-modal data into the mean and variance corresponding to the distribution of each modality data; The sampling unit samples the mean and variance to generate a latent feature representation, and minimizes the difference between the mean and variance and the standard Gaussian distribution parameters through a KLD (Kullback-Leibler divergence) divergence loss function, so that the latent feature representation as a whole obeys a smooth continuous distribution; The decoder is used to reconstruct the latent feature representation into the input multi-modal data, and outputs the reconstructed multi-modal data, and minimizes the difference between the reconstructed multi-modal data and the input multi-modal data through a MSE (Mean Squared Error) loss function.

3. The attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrastive learning according to claim 1, characterized in that, The multi-view learning module comprises a projector and a shared semantic encoder, wherein: The projector is used to map the latent feature representations of each modality into projected semantic features; The shared semantic encoder is used to project the projected semantic features into a shared semantic space to obtain a shared semantic feature, and minimize the difference between the projected semantic features and the shared semantic feature through a KLD (Kullback-Leibler divergence) divergence loss function. The shared semantic encoder is configured to generate a shared semantic feature in the shared semantic space, and a process of generating the shared semantic feature comprises: an initialization step of initializing a shared semantic feature vector; an encoding step of inputting the initialized shared semantic feature into the shared semantic encoder to generate a mean and a variance of the feature distribution; a constraint step of minimizing a difference between the mean and the variance and a standard Gaussian distribution parameter by using a KLD divergence loss function, to constrain and update the shared semantic feature, and simultaneously minimizing a difference between the projected semantic feature and the shared semantic feature to obtain a distribution-aligned shared semantic feature; the encoding step and the constraint step are iteratively executed to finally obtain the shared semantic feature.

4. The attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrastive learning according to claim 1, characterized in that, The hierarchical multi-agent strategy comprises: setting different category levels of agents in the shared semantic space, including main agents representing two large category levels of health and disease and sub-agents representing subtype levels, wherein the main agents include two main agents representing health human center features and attention deficit hyperactivity disorder patient class center features, respectively, the sub-agents include sub-agents representing attention deficit hyperactivity disorder subtype class center features, the subtype classes refer to the classes of different subtypes, and the number of sub-agents is equal to the number of subtype classes.

5. The attention deficit hyperactivity disorder subtype identification and evolution analysis method based on contrastive learning according to claim 1, characterized in that, The hierarchical multi-agent strategy further comprises: setting a contrast loss function to regularize and constrain the shared semantic features in the shared semantic space, so that the shared semantic features of individuals of the same category converge to the category center, and different categories and different subtypes remain separated.

6. The method of claim 5, wherein, The contrast loss function L con By sample-agent contrast loss L sp And agent-agent contrast loss L pp Weighted composition, namely L con = L sp + λL pp wherein λ is a weight coefficient for balancing the two loss terms; the calculation formula of the sample-agent contrast loss is: In the formula z i Let N represent the shared semantic features of the i-th sample, N be the total number of samples, τ1 be the temperature coefficient, d(·) be the distance metric function, and p be the distance metric function. + With p - These represent positive and negative proxies, respectively; the positive proxy corresponds to z. i The negative proxy represents the center of the class to which it belongs, while the negative proxy represents the center of the class to which it does not belong; if z i If the category belongs to patients with attention deficit hyperactivity disorder, then the category not belonging to is healthy humans, and vice versa; If z i If the subject class is a subtype of attention deficit hyperactivity disorder, then the non-subject class is a subtype of attention deficit hyperactivity disorder other than the subject class, and vice versa. the calculation formula of the agent-agent contrast loss is: where p i represents the ith sub-agent, P represents the set of all sub-agents, |P| represents the number of sub-agents, τ2is a temperature coefficient; p A represents the master agent representing the central features of the class of attention deficit hyperactivity disorder patients, P * represents the set of all agents other than the current sub-agent p i . represents the kth agent in P * .

Images