A system for identifying comorbidities of neurodevelopmental disorders based on residual graph neural networks

Abstract

A comorbidity identification system for neurodevelopmental disorders based on residual graph neural networks. This system belongs to the field of comorbidity identification technology for neurodevelopmental disorders, specifically involving comorbidity identification based on neural networks. The system includes: a data acquisition module for acquiring resting-state fMRI images; a data preprocessing module for extracting standardized time series of 116 brain regions from the resting-state fMRI images; a multi-band segmentation and feature extraction module for performing band filtering on the standardized time series of each brain region to obtain filtered time series for each band; within each band, constructing a binary topological matrix based on the PLV matrix and calculating PLV feature vectors based on the PLV matrix; a comorbidity identification module for neurodevelopmental disorders for analyzing brain functional connectivity maps using a residual graph neural network model to obtain diagnostic information on comorbidities of neurodevelopmental disorders; and an interpretable output module that outputs abnormal brain regions using low-frequency amplitude fluctuations as a reference indicator.

Description

Technical Field

[0001] This invention belongs to the field of comorbidity identification technology for neurodevelopmental disorders, specifically relating to the field of comorbidity identification technology based on neural networks. Background Technology

[0002] Current methods for the automatic identification and mechanism modeling of neurodevelopmental disorders (such as autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD) mostly employ brain functional connectivity analysis based on resting-state functional magnetic resonance imaging (rs-fMRI), including the following mainstream technical approaches:

[0003] Traditional classification models (such as SVM and RF) often construct feature vectors based on fully connected matrices or statistical features, which are then input into traditional machine learning classifiers. While these methods are simple to implement, they lack the ability to model complex interactive structures between brain regions and struggle to capture dynamic network changes and frequency sensitivity.

[0004] Convolutional neural networks (such as BrainNetCNN) model the connection matrix using specific topological convolutions, but their structure is usually fixed, lacks the ability to model node semantics, and is difficult to integrate local brain region features and frequency band characteristics, resulting in a "black box" problem and a lack of biological interpretability.

[0005] Graph neural network methods (such as BrainGNN, MHNet, and HAGCN): In recent years, brain map modeling methods based on GNNs have emerged, treating ROIs as nodes in a graph and connecting them functionally as edges, thus capturing structural dependencies. However, most existing methods have the following shortcomings:

[0006] Frequency insensitive: It does not incorporate frequency domain features such as slow wave bands, and therefore cannot reflect changes in brain functional state at different frequencies;

[0007] Weak dynamic mechanism modeling: lack of modeling of the nonlinear dynamic characteristics of the original BOLD time series, such as LLE;

[0008] The connection mechanism is unclear: most mapping methods use static linear indices such as Pearson correlation, ignoring more physiologically significant connection mechanisms such as the phase synchronization of neural activity;

[0009] Weak interpretability: Most models do not provide an explanation for the contribution of frequency bands / brain regions / connections to classification results, making it difficult to conduct mechanistic analysis and clinical translation. Summary of the Invention

[0010] To address the key issues of existing technologies, such as frequency domain insensitivity, insufficient modeling of brain region dynamic features, lack of physiological interpretation of functional connectivity, and strong black box nature of the models, this invention provides a neurodevelopmental disorder comorbidity identification system based on residual graph neural networks.

[0011] The system includes:

[0012] Data acquisition module: Acquires resting-state fMRI images;

[0013] Data preprocessing module: Based on the AAL brain region template, normalized time series of 116 brain regions from resting-state fMRI images were extracted. ;use This represents any one of the time series;

[0014] Multi-band segmentation and feature extraction module: The standardized time series of each brain region is filtered in four frequency bands to obtain the filtered time series of each frequency band, and the node feature matrix of the time series is extracted; within each frequency band, PLV calculation is performed to obtain the PLV matrix, and a binary topological matrix is ​​constructed based on the PLV matrix and the PLV feature vector is calculated based on the PLV matrix; the node feature matrix, the binary topological matrix and the PLV feature vector are combined to obtain the brain functional connectivity map;

[0015] Neurodevelopmental Disorder Comorbidity Identification Module: Through a residual graph neural network model, the module analyzes brain functional connectivity maps to obtain diagnostic information on neurodevelopmental disorder comorbidities.

[0016] Interpretable output module: Uses low-frequency amplitude fluctuations as a reference indicator to output abnormal brain regions.

[0017] Furthermore, in the data preprocessing module, before generating the standardized time series, data preprocessing is performed, specifically: sequentially performing skull dissection, head motion correction, spatial smoothing through Gaussian filtering, delinearization using FSL bandpass filtering, mean function image calculation, and standard spatial alignment using the mean function image.

[0018] Furthermore, the four frequency bands are: SLOW2 (0.198-0.25Hz), SLOW3 (0.073-0.198Hz), SLOW4 (0.027-0.073Hz) and SLOW5 (0.01-0.027Hz).

[0019] The time series after filtering for each frequency band is: ,in Indexes representing brain regions The index represents the frequency band, and the node feature matrix includes statistical features and nonlinear features.

[0020] Furthermore, the statistical characteristics include the mean. Standard deviation skewness and kurtosis :

[0021] ;in, This indicates the number of time points contained in the time series of each brain region within a specific frequency band. Index representing a point in time;

[0022] ; ;

[0023] .

[0024] Furthermore, the method for obtaining the nonlinear feature is as follows:

[0025] S51, Local Neighborhood Construction: In Time Series In the diagram, the signal value at each time point is... The K-nearest neighbor algorithm is used for each time point. Find the 5 most similar nearest neighbors, and measure the similarity using Euclidean distance: ,in, Indicates time point Similar time point index, Indicates a point in time and time point The Euclidean distance is calculated; the five points with the smallest distance are selected as the nearest neighbors, denoted as . , constituting a time point The local neighborhood;

[0026] S52. Reconstructing Weights: For each time point... Assuming its signal value is reconstructed from a linear combination of its nearest neighbors, i.e. ,in Indicates a point in time At time point To minimize the reconstruction error, we need to determine the contribution weights and solve the following optimization problem:

[0027] ;

[0028] The constraints are ,in Indicates a point in time The set of reconstructed weights for all topological points. During the solution process, the local neighborhood signal is first centered, i.e., subtracted... , to obtain the matrix Then, the weights are solved using the least squares method. :

[0029] ;in To avoid matrix singularity, a regularization term is used, where 1 represents a vector consisting entirely of 1s. It is the identity matrix;

[0030] S53, Feature Quantization: Through Get each The nonlinear characteristics in each frequency band, where, This represents the variance function.

[0031] Furthermore, the PLV calculation is as follows:

[0032] S61, Hilbert Transform: [This part is incomplete and likely refers to a specific transformation or process.] Convert to analytic signal ,in Represents the Hilbert transform;

[0033] S62, Phase Extraction: Through Obtain the phase angle of the analytic signal , This indicates the phase angle calculation operation;

[0034] S63, PLV Calculation: Synchronization is quantified by the average cosine of the phase difference, using the following formula:

[0035] ,in The imaginary unit, The larger the value, the more likely it is that the two The stronger the phase synchronization, and Representing two The instantaneous phase angle at a specific time point is used to calculate the PLV matrix.

[0036] Furthermore, the construction of a binary topology matrix based on the PLV matrix is ​​as follows:

[0037] When in the PLV matrix At that time, it was believed that the two There exists an edge; index all edges to form a binary topological matrix. .

[0038] Furthermore, the calculation of PLV feature vectors based on the PLV matrix specifically involves: extracting the upper triangular elements of the PLV matrix for each frequency band to form a one-dimensional feature vector; and concatenating the feature vectors of the PLV matrices for the four frequency bands to form a globally connected feature. , This is the PLV feature vector.

[0039] Furthermore, the workflow in the residual graph neural network model is as follows:

[0040] S91. Local Node Dynamic Capture: Through the first-layer graph convolution operation, the features of each node are weighted and aggregated with the features of its neighboring nodes to obtain the feature matrix. : ,in This represents the activation function. Represents the normalization degree matrix, Represents a binary topological matrix incorporating self-connections. , Represents the node feature matrix, Indicates learnable weights;

[0041] S92, Regional Network Feature Aggregation: Through a second-layer graph convolution operation, the feature extraction scope is expanded to larger brain network units to obtain a feature matrix. : ,in Indicates learnable weights;

[0042] S93. Whole-brain topology pattern mining: Integrating cross-network connectivity features across the whole brain to form a global topology pattern representation and obtain a feature matrix. : ,in Represent learnable weights; utilize the feature matrix All data are pooled using global mean pooling. The features of the nodes are aggregated into a single vector. ;

[0043] S94. Focusing on PLV Phase Synchronization Mode Mining: Through multi-layer nonlinear transformation, the PLV feature vector is mapped to a low-dimensional space while retaining the core mode. : ,in express Activation function For PLV feature vectors, These are learnable weights;

[0044] S95, Cross-modal feature fusion:

[0045] Will and The fused features are obtained through adaptive weight fusion. : ,in For the fusion weights of network learning;

[0046] S96. Classification result output:

[0047] The fused features First, it goes through two layers of fully connected network processing:

[0048] ,

[0049] The first layer uses a weight matrix. The feature dimension is compressed to 128 dimensions while retaining core discriminative information. To prevent overfitting, a function is used in the second layer; the weight matrix is ​​used... Further refine the key relationships between features;

[0050] pass softmax The function maps the processed features to three probability distributions:

[0051] , The weight matrix is ​​output. ,in The probabilities of having only ASD, having both ADHD and ASD, and having only ASD are given, respectively. The category corresponding to the highest probability is taken as the final prediction result.

[0052] Furthermore, the low-frequency amplitude fluctuations are calculated by each The amplitude values ​​of low-frequency oscillation signals in four frequency bands were obtained. The differences in low-frequency fluctuation values ​​of amplitude in each frequency band between the comorbidity group and the single disease group were compared using independent samples testing. Statistically significant abnormal brain regions were screened out and output.

[0053] The beneficial effects of the system described in this invention are as follows: This invention can be used to process resting-state fMRI brain network data, realize the identification of ADHD, ASD and their comorbid individuals, and has the ability to interpret abnormal neural connectivity across frequencies and structures.

[0054] It has the following technological advantages:

[0055] High frequency sensitivity: It comprehensively introduces multiple slow wave frequency bands from SLOW-2 to SLOW-5, and captures frequency-dependent brain functional activity features through filtering and feature fusion, thereby enhancing the model's sensitivity to frequency band-specific abnormalities such as ASD / ADHD.

[0056] Accurate modeling of dynamic features: By integrating Local Linear Embedding (LLE) features, the nonlinear dynamic structure of brain regions is accurately characterized, making up for the limitations of traditional statistical features.

[0057] The connectivity method has stronger physiological significance: using the phase-locked value (PLV) as a brain region connectivity indicator reflects the phase synchronization relationship between neural activities, which is more in line with the actual neurophysiological mechanism.

[0058] Structure-dynamic fusion modeling: By using graph convolutional networks and residual collaborative modeling of global PLV features, and integrating node features (local) and connection features (global), an end-to-end structure-function coupling identification model is established.

[0059] Multidimensional interpretability mechanism analysis: Introducing frequency band contribution visualization, PLV connectivity difference analysis and brain region significance statistics into the model to construct a complete visualization system for neural network abnormality mechanisms, thereby improving the clinical interpretability and translational value of the results. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the workflow of the system described in an embodiment of the present invention.

[0061] Figure 2 This is a diagram showing the differences in abnormal brain regions according to an embodiment of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1

[0064] This embodiment provides a neurodevelopmental disorder comorbidity identification system based on residual graph neural networks. The workflow of the system is as follows: Figure 1 As shown. The system includes:

[0065] Data acquisition module: Acquires resting-state fMRI images;

[0066] Data preprocessing module: Based on the AAL brain region template, normalized time series of 116 brain regions from resting-state fMRI images were extracted. ;use This represents any one of the time series;

[0067] Multi-band segmentation and feature extraction module: The standardized time series of each brain region is filtered in four frequency bands to obtain the filtered time series of each frequency band, and the node feature matrix of the time series is extracted; within each frequency band, PLV calculation is performed to obtain the PLV matrix, and a binary topological matrix is ​​constructed based on the PLV matrix and the PLV feature vector is calculated based on the PLV matrix; the node feature matrix, the binary topological matrix and the PLV feature vector are combined to obtain the brain functional connectivity map;

[0068] Neurodevelopmental Disorder Comorbidity Identification Module: Through a residual graph neural network model, the module analyzes brain functional connectivity maps to obtain diagnostic information on neurodevelopmental disorder comorbidities.

[0069] Interpretable output module: Uses low-frequency amplitude fluctuations as a reference indicator to output abnormal brain regions.

[0070] Example 2

[0071] This embodiment further defines Embodiment 1 and provides a further explanation of the data acquisition module.

[0072] The data acquisition module acquires resting-state fMRI images, with a recommended resolution of at least 3 mm³ and a duration of at least 5 minutes.

[0073] Example 3

[0074] This embodiment further defines Embodiment 1 and provides a further explanation of the data preprocessing module.

[0075] The data preprocessing steps in the data preprocessing module are as follows:

[0076] The raw rs-fMRI data were preprocessed using the FSL (FMRIB Software Library) tool as follows:

[0077] 1. Skull Dissection: Using FSL's `bet` command, the skull is dissected from the original structural or functional images to remove non-brain tissue, preserving clean brain tissue regions for accurate registration and analysis later. 2. Head Motion Correction: Using FSL's `mcflirt` tool, all echo-plane imaging (EPI) volumes are rigidly registered to the first volume to correct for head movements during scanning, reducing motion artifacts. 3. Spatial Smoothing: Gaussian filtering is applied to the motion-corrected data, with a kernel width corresponding to approximately 6mm of FWHM (full width at half maximum), improving the signal-to-noise ratio and smoothing spatial noise. 4. De-linearization: FSL's bandpass filtering tool removes linear trends and high-frequency noise from the time series, reducing slow drift and interference from non-neural signals. 5. Calculation of Mean Functional Image: The de-trended mean functional image is calculated as the basis for subsequent registration. 6. Spatial Registration to Standard Space: Affine registration was performed using FSL's flip function to register the mean function images to the MNI152 standard template (4×4×4 mm resolution). This transformation was then applied to the entire 4D detrended data to achieve standard spatial alignment. 7. ROI Time Series Extraction: Based on the AAL brain region template, Nilearn's NiftiLabelsMasker was used to extract normalized time series data from 116 brain regions in the resting-state fMRI images. ;use This represents any one of the time series.

[0078] Example 4

[0079] This embodiment further defines Embodiment 1 and provides a further explanation of the multi-band segmentation and feature extraction module.

[0080] Brain functional activity exhibits different characteristics at different frequencies, especially slow-wave activity, which is closely related to neurodevelopmental disorders. Therefore, this invention focuses on four key slow-wave frequency bands: SLOW2 (0.198-0.25Hz): associated with rapid cognitive processing; SLOW3 (0.073-0.198Hz): associated with attention regulation; SLOW4 (0.027-0.073Hz): involved in memory and emotional processing; and SLOW5 (0.01-0.027Hz): associated with default mode network function. These frequency band divisions are based on previous neuroimaging studies and can specifically capture abnormal brain function patterns in patients with ASD, ADHD, and comorbidities.

[0081] The following two types of features (statistical features and nonlinear features) are extracted in each frequency band to form a node feature matrix.

[0082] 1. Statistical Feature Extraction

[0083] Time series after filtering for each frequency band Four basic statistics are calculated to characterize the signal from different perspectives:

[0084] mean reflect The overall activity level in this frequency band,

[0085] ,in, This indicates the number of time points contained in the time series of each brain region within a specific frequency band. Index representing a point in time;

[0086] Standard deviation Reflecting the stability of the activity,

[0087] ;

[0088] Skewness It reflects the asymmetry of signal distribution.

[0089] ;

[0090] Kudo The steepness of the response signal distribution.

[0091] ;

[0092] These features can effectively distinguish differences in the intensity and stability of activity in specific brain regions.

[0093] 2. Nonlinear Feature Extraction (LLE Local Linear Embedding)

[0094] LLE features are designed to capture a single The core idea of ​​local dynamic patterns in time series (such as the correlation and stability of short-term fluctuations) is: assuming that time series can be reconstructed in a local range through a linear combination of nearest neighbors, and quantifying the dispersion of local patterns by solving the variance of the reconstruction weights.

[0095] (1) Local neighborhood construction

[0096] Time series after bandpass filtering First, determine the nearest neighbors for each time point. Let the length of the time series be... The signal value at each time point is .

[0097] The K-Nearest Neighbors algorithm is used for each time point. Find the most similar The nearest neighbors. Similarity is measured using Euclidean distance:

[0098] ;in, Indicates time point Similar time point index, Indicates a point in time and time point The Euclidean distance;

[0099] Select the 5 smallest points as the nearest neighbors, denoted as . , constituting a time point The local neighborhood.

[0100] (2) Solving for reconstructed weights

[0101] For each time point Assuming its signal value can be reconstructed from a linear combination of its nearest neighbors, i.e.

[0102] ;

[0103] in Indicates a point in time At time point Given the contribution weights (non-nearest neighbor points have a weight of 0), to minimize the reconstruction error, solve the following optimization problem:

[0104] ;

[0105] The constraints are (Ensure weight normalization to avoid scaling bias), where Indicates a point in time The set of reconstructed weights for all topological points. During the solution process, the local neighborhood signal is first centered, i.e., subtracted... , to obtain the matrix Then, the weights are solved using the least squares method. :

[0106]

[0107] in 1 is a vector consisting entirely of 1s. It is an identity matrix, with all values ​​on the main diagonal being 1 and all others being 0; Trace() indicates (Physical meaning: equivalent to) (The sum of squared Euclidean distances of all column vectors), and finally the weights Normalization is performed to satisfy the constraints.

[0108] (3) LLE Feature Quantization

[0109] Reconstruct weights It reflects the contribution strength of nearest neighbors to time point k, and its variance can quantify the stability of the local reconstruction pattern: if the weight variance is small, it indicates that the contribution distribution of nearest neighbors is stable; if the weight variance is large, it indicates that the contribution of nearest neighbors fluctuates drastically. Therefore, the LLE feature is defined as the variance of the reconstruction weights at all time points:

[0110]

[0111] in For the variance function, ultimately each The LLE characteristic in each frequency band is a scalar value.

[0112] Based on PLV, a functional connection graph is constructed, and the connection features are intended to quantify different... The functional synergy between brain regions is assessed using the phase lock value (PLV) as the core indicator, which reflects the phase synchronization (coordination of neural activity) between brain regions.

[0113] The PLV calculation is as follows:

[0114] (1) Hilbert transform: transforms time series Convert to analytic signal Its imaginary part reflects the instantaneous phase of the signal;

[0115] (2) Phase extraction: The phase of the analytical signal is , representing the phase angle of the signal at each time point;

[0116] (3) PLV Calculation: Synchronization is quantified by the average cosine of the phase difference, and the formula is:

[0117] ;in The imaginary unit, The larger the value, the more likely it is that the two The stronger the phase synchronization, and Representing two The instantaneous phase angle at a specific time point is used to calculate the PLV matrix.

[0118] Constructing a binary topological matrix based on PLV matrix: Constructing a brain functional connectivity map based on PLV, where there are 116 "nodes". "edge" represents Connections between:

[0119] when At that time, it was believed and There exists an edge; the edge has only two possibilities: it exists or it does not exist, which is 0 / 1.

[0120] All edge indices form a binary topological matrix. Record the topology of the graph.

[0121] PLV Feature Vector Extraction: To preserve the intensity information of PLV, the upper triangular elements of the PLV matrix for each frequency band are extracted to form a one-dimensional feature vector. (This is repeated for each of the four frequency bands.) After concatenating the feature vectors, a globally connected feature is formed. , as the global connectivity property of the graph. Where 26680 is composed of upper triangular elements (116) across 4 frequency bands. 115 / 2) obtained.

[0122] After the above steps, each sample (patient) is converted into a brain functional connectivity map.

[0123] Example 5

[0124] This embodiment further defines Embodiment 1 and provides a further explanation of the neurodevelopmental disorder comorbidity identification module.

[0125] MBF-ResGCN deeply mines disease-specific patterns hidden in brain functional connectivity maps by fusing multi-stage feature extraction with neurodynamic features. This model closely revolves around the dynamic characteristics of nodes in brain networks and neurodynamic connectivity patterns centered on PLV, constructing a feature mining framework that fits the characteristics of brain functional data. Specifically, the multi-stage graph feature extraction module employs a progressive feature distillation strategy, utilizing multi-level graph convolution operations to model dynamic patterns of brain regions from local to global perspectives, achieving efficient transfer and expression of deep graph features without relying on traditional residual structures.

[0126] Local node dynamic capture:

[0127] For the input node feature matrix (Contains 116) (Multi-band statistical and non-linear features), dimension 116×20 = 116 brain regions × (4 statistical features + 1 non-linear feature) × 4 frequency bands.

[0128] By performing a convolution operation on the first layer of the graph, the focus is on a single... It relates to local connections with brain regions that are directly connected to it. Its core is through graph topology (binary topological matrices). The features of each node are weighted and aggregated with the features of its neighboring nodes to obtain the feature matrix. :

[0129] ;

[0130] in To incorporate self-connected binary topological matrices (Preserve the dynamic characteristics of the node itself) For learnable weights, For activation function, The normalization matrix (mathematical form: diagonal matrix (off-diagonal elements are 0)) is used to capture the functional coordination patterns of a single brain region and its adjacent regions (such as the local interaction between the prefrontal cortex and the cingulate cortex), laying the foundation for subsequent analysis.

[0131] In a binary topological matrix An extension based on the existing model (incorporating self-connection). exist The addition of "connections between nodes themselves" is based on this: for each (node), The positions corresponding to "its own row and its own column" will be set to 1 (or other weights used to preserve the node's own characteristics) to retain the node's dynamic features; while Only describing the differences Connections between elements (no self-connections, i.e., all elements on the diagonal of the matrix are 0).

[0132] Regional network feature aggregation:

[0133] After the first stage of processing, the feature matrix Local correlation information is already included. The second stage expands the feature extraction scope to larger brain network units through deeper graph convolutions:

[0134] ;

[0135] This stage uses the weight matrix. Learning enhances the capture of features of disease-related brain networks, filters irrelevant connectivity information, and highlights regional functional abnormality patterns.

[0136] Whole-brain topological pattern mining:

[0137] Further expand the receptive field and integrate cross-network connectivity features across the entire brain to form a global topological pattern representation:

[0138] ;

[0139] Finally, global mean pooling is used to aggregate the features of 116 nodes into a single vector. This vector encapsulates the dynamic characteristics and topological information of all brain nodes.

[0140] Focusing on the phase synchronization mode exploration of PLV:

[0141] Phase lock value (PLV), as a core indicator reflecting the synchronicity of neural dynamics between brain regions, contains phase coordination patterns that are key biomarkers for distinguishing three types of diseases. MBF-ResGCN has designed a dedicated processing pathway to deeply mine the dynamic information in PLV:

[0142] High-dimensional PLV feature dimensionality reduction and enhancement: Original PLV features (The concatenation of upper triangular elements from four frequency bands) has extremely high dimensionality and contains a large amount of redundant information. Through multi-layer nonlinear transformation, it is mapped to a lower-dimensional space while preserving the core pattern: ,in As learnable weights, the ReLU activation function is used to filter and highlight the phase synchronization mode that is most critical for classification.

[0143] Band-specific PLV fusion: During processing, the network automatically learns the weights of PLVs in different bands to achieve band-adaptive fusion of neurodynamic features, avoiding the subjectivity of manual band selection.

[0144] graph structure features With neurodynamic characteristics By using adaptive weight fusion, complex relationships that cannot be reflected by a single feature can be captured:

[0145] ;

[0146] in For network learning, the fusion weights and the fused features By combining brain region functional characteristics with neurodynamic connectivity patterns, it provides a discriminative power for subsequent classification that far exceeds that of a single feature.

[0147] The classification results serve as the model's decision-making terminal. Taking the fused features as input, it achieves accurate classification of three types of samples—ASD, ASD and ADHD comorbidity, and ADHD—through concise and efficient transformations. The core is to enhance the discriminative power of features and output interpretable class probabilities.

[0148] Fusion features (Integrating graph structural features and PLV neurodynamic features) First, it is processed through two layers of fully connected networks:

[0149] ;

[0150] The first layer uses a weight matrix. The feature dimension is compressed to 128 dimensions while retaining the core discriminative information; (30% probability) to prevent overfitting; the second layer passes... Further refine the key relationships between features.

[0151] Finally, the processed features are mapped to three probability distributions using the softmax function: Output in The probabilities of having only ASD, having both ADHD and ASD, and having only ASD are given, respectively. The category corresponding to the highest probability is taken as the final prediction result.

[0152] Example 6

[0153] This embodiment further defines Embodiment 1 and provides further explanation of the interpretability output module.

[0154] The following methods can be used to output interpretable results from the model:

[0155] Amplitude Low-Frequency Fluctuations (ALFF) are an effective indicator of the intensity of spontaneous neuronal activity in local areas. They exhibit high sensitivity to cognitive functions (including attention and emotion regulation) across different slow-wave frequency bands (e.g., SLOW-2 to SLOW-5). ALFF is calculated by measuring the mean amplitude of low-frequency oscillations in the SLOW2-SLOW5 frequency band for each region of interest (ROI). Independent samples spectroscopy is used to compare the differences in ALFF values ​​across different frequency bands between the comorbidity group and the single-disease group, identifying statistically significant abnormal brain regions and revealing unique neural markers of local brain activity in comorbid individuals. Figure 2 As shown, the ALFF differences between patients with comorbid ADHD-ASD and those with ADHD alone are illustrated across four frequency bands, as are the ALFF differences between patients with comorbid ADHD-ASD and those with ASD alone across four frequency bands.

[0156] Table 1 shows specific brain function abnormalities associated with comorbid neurodevelopmental disorders. This table systematically reveals the differences in brain function between the comorbid neurodevelopmental disorder group (ASD-ADHD) and the single-disorder group (ADHD or ASD), by quantifying different brain regions ( Abnormal activity intensity in key slow wave bands provides an interpretable biomarker for comorbidity mechanisms, where "+" indicates enhancement and "-" indicates significant attenuation.

[0157] Table 1:

[0158]

[0159] Example 7

[0160] This embodiment is a further explanation of Embodiments 1-6. It describes the specific application of the systems in Embodiments 1-6.

[0161] Step 1: rs-fMRI Data Acquisition

[0162] In this embodiment, subjects meeting the inclusion criteria, including ASD patients, ADHD patients, patients with ASD-ADHD comorbidity, and healthy controls, were recruited. Resting-state functional imaging data were acquired using a 3.0T magnetic resonance imaging system. Scanning parameters were set as follows: repetition time (TR) of 2000 ms, echo time (TE) of 30 ms, flip angle of 90°, 32 slices, spatial resolution of 3×3×3 mm³, and scan duration of 8 to 10 minutes. Subjects remained at rest with their eyes closed during the scan and did not engage in any specific thought activities. Simultaneously, demographic information and clinical diagnoses of all subjects were recorded for subsequent grouping and labeling.

[0163] Step 2: rs-fMRI Data Preprocessing

[0164] After acquisition, the raw functional images are preprocessed using the FSL toolkit according to a standard workflow, which includes:

[0165] Skull dissection: Removal of non-brain tissue to extract pure brain signals;

[0166] Head movement correction: The rigid registration algorithm is used to correct artifacts caused by the subject's head movement and to remove samples with head movement greater than 1.5 mm or rotation angle greater than 2°.

[0167] Spatial standardization: Register functional images to the standard MNI152 template space to ensure spatial consistency;

[0168] Spatial smoothing: A Gaussian kernel with a full width at half height (FWHM) of 6 mm is used for smoothing to improve the signal-to-noise ratio;

[0169] Detrending and filtering: Linear drift removal is performed on the time series, and a bandpass filter (frequency range of 0.01Hz to 0.25Hz) is applied to preserve low-frequency components related to neural activity;

[0170] Noise regression: regression of white matter signals, cerebrospinal fluid signals, and head movement parameters to further remove non-neurogenic signal interference.

[0171] Step 3: Brain Region Division and Feature Extraction

[0172] The brain was divided into 116 functional regions using the AAL3 brain region template. Extract each from the preprocessed functional image The time series of the BOLD signal was analyzed, and features were constructed at the following two levels:

[0173] Node feature extraction: On multiple frequency bands from SLOW2 to SLOW5, for each Calculate the following features:

[0174] Low-frequency amplitude characteristics: Calculating ALFF (Amplitude of Low-Frequency Fluctuation) reflects the level of neural activity in the low-frequency range of this brain region;

[0175] Statistical characteristics: including mean, standard deviation, skewness, and kurtosis, which describe the distribution characteristics of the time series;

[0176] LLE Features: Based on the Locally Linear Embedding (LLE) method, each LLE is constructed using K-nearest neighbors. The local reconstruction neighborhood is calculated, and the variance of the reconstruction weights is used to characterize the nonlinear local structure features.

[0177] Connection feature extraction: for all The combination of phase information extracted based on Hilbert transform and the calculation of phase locking value (PLV) are used to construct a 116×116 connectivity matrix that reflects the phase synchronization of brain regions.

[0178] Average the PLV within each sample group to obtain the stable connection pattern within the group.

[0179] Step 4: Classification using Multi-Band Residual Graphical Neural Network (MBF-ResGCN)

[0180] The extracted node features and the connection matrix constructed using PLV are used as graph structure data input into the MBF-ResGCN model. The model structure includes:

[0181] Three-layer residual graph convolution module: extracts local functional connectivity, regional network features and whole brain topology information layer by layer;

[0182] Frequency band adaptive fusion module: Adaptively weighted integration of multi-band PLV features to reduce redundancy and enhance frequency band discrimination capability;

[0183] Cross-modal feature fusion module: fuses node attributes and structural connectivity information to enhance the model's ability to perceive multimodal brain network features;

[0184] Classification layer: Outputs three diagnostic probabilities (ASD, ADHD, comorbidity) through a fully connected network, and takes the category corresponding to the highest probability as the final diagnosis result.

[0185] Step 5: Model Interpretability Analysis

[0186] To improve the interpretability of the model in terms of neural mechanism analysis, the following analytical methods were implemented:

[0187] Brain region node significance analysis: Independent samples test was used to compare the ROI ALFF values ​​between the comorbidity group and the ASD / ADHD individual group. Brain regions with p<0.05 were selected as potential neural markers and their frequency dependence was marked.

[0188] Functional connectivity difference analysis: targeting For each group, the PLV values ​​were compared to show differences between groups: ADHD vs. comorbidity, comorbidity vs. ASD, and ADHD vs. ASD. Groups with a corrected p < 0.05 were selected. Yes, the visualization plots highlight significantly elevated or decreased connectivity to reveal anomalous connectivity patterns in the functional network structure between different disease groups.

Claims

1. A system for identifying comorbidities of neurodevelopmental disorders based on residual graph neural networks, characterized in that, The system includes: Data acquisition module: Acquires resting-state fMRI images; Data preprocessing module: Based on the AAL brain region template, normalized time series of 116 brain regions from resting-state fMRI images were extracted. ;use This represents any one of the time series; Multi-band segmentation and feature extraction module: The standardized time series of each brain region is filtered in four frequency bands to obtain the filtered time series of each frequency band, and the node feature matrix of the time series is extracted; within each frequency band, PLV calculation is performed to obtain the PLV matrix, and a binary topological matrix is ​​constructed based on the PLV matrix and the PLV feature vector is calculated based on the PLV matrix; the node feature matrix, the binary topological matrix and the PLV feature vector are combined to obtain the brain functional connectivity map; The four frequency bands are: SLOW2 (0.198-0.25Hz), SLOW3 (0.073-0.198Hz), SLOW4 (0.027-0.073Hz) and SLOW5 (0.01-0.027Hz). The time series after filtering for each frequency band is: in Indexes representing brain regions The index represents the frequency band, and the node feature matrix includes statistical features and nonlinear features; The statistical characteristics include the mean. Standard deviation skewness and kurtosis : ;in, This indicates the number of time points contained in the time series of each brain region within a specific frequency band. Index representing a point in time; ； ； ； The method for obtaining the nonlinear feature is as follows: S51, Local Neighborhood Construction: In Time Series In the diagram, the signal value at each time point is... The K-nearest neighbor algorithm is used for each time point. Find the 5 most similar nearest neighbors, and measure the similarity using Euclidean distance: ,in, Indicates time point Similar time point index, Indicates a point in time and time point The Euclidean distance is calculated; the five points with the smallest distance are selected as the nearest neighbors, denoted as . , constituting a time point The local neighborhood; S52. Reconstructing Weights: For each time point... Assuming its signal value is reconstructed from a linear combination of its nearest neighbors, i.e. ,in Indicates a point in time At time point To minimize the reconstruction error, we need to determine the contribution weights and solve the following optimization problem: ； The constraints are ,in Indicates a point in time The set of reconstructed weights for all topological points. During the solution process, the local neighborhood signal is first centered, i.e., subtracted... , to obtain the matrix Then, the weights are solved using the least squares method. : ;in To avoid matrix singularity, a regularization term is used, where 1 represents a vector consisting entirely of 1s. It is the identity matrix; S53, Feature Quantization: Through Get each The nonlinear characteristics in each frequency band, where, Represents the variance function; Neurodevelopmental Disorder Comorbidity Identification Module: Through a residual graph neural network model, the module analyzes brain functional connectivity maps to obtain diagnostic information on neurodevelopmental disorder comorbidities. Interpretable output module: Uses low-frequency amplitude fluctuations as a reference indicator to output abnormal brain regions.

2. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 1, characterized in that, In the data preprocessing module, before generating the standardized time series, data preprocessing is performed in sequence, specifically: skull dissection, head motion correction, spatial smoothing by Gaussian filtering, delinearization by bandpass filtering using FSL, mean function image calculation, and standard spatial alignment using the mean function image.

3. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 2, characterized in that, The PLV calculation is as follows: S61, Hilbert Transform: [This part is incomplete and likely refers to a specific transformation or process.] Convert to analytic signal ,in Represents the Hilbert transform; S62, Phase Extraction: Through Obtain the phase angle of the analytic signal , This indicates the phase angle calculation operation; S63, PLV Calculation: Synchronization is quantified by the average cosine of the phase difference, using the following formula: ,in The imaginary unit, The larger the value, the more likely it is that the two The stronger the phase synchronization, and Representing two The instantaneous phase angle at a specific time point is used to calculate the PLV matrix.

4. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 3, characterized in that, The specific steps for constructing a binary topological matrix based on a PLV matrix are as follows: When in the PLV matrix At that time, it was believed that the two There exists an edge; index all edges to form a binary topological matrix. .

5. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 4, characterized in that, The specific steps for calculating PLV feature vectors based on PLV matrices are as follows: Extract the upper triangular elements of the PLV matrix for each frequency band to form a one-dimensional feature vector. Concatenate the feature vectors from the PLV matrices of the four frequency bands to form a globally connected feature vector. , This is the PLV feature vector.

6. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 5, characterized in that, The workflow in the residual graph neural network model is as follows: S91. Local Node Dynamic Capture: Through the first-layer graph convolution operation, the features of each node are weighted and aggregated with the features of its neighboring nodes to obtain the feature matrix. : ,in This represents the activation function. Represents the normalization degree matrix, Represents a binary topological matrix incorporating self-connections. , Represents the node feature matrix, Indicates learnable weights; S92, Regional Network Feature Aggregation: Through a second-layer graph convolution operation, the feature extraction scope is expanded to larger brain network units to obtain a feature matrix. : ,in Indicates learnable weights; S93. Whole-brain topology pattern mining: Integrating cross-network connectivity features across the whole brain to form a global topology pattern representation and obtain a feature matrix. : ,in Indicates learnable weights; Using the feature matrix All data are pooled using global mean pooling. The features of the nodes are aggregated into a single vector. ; S94. Focusing on PLV Phase Synchronization Mode Mining: Through multi-layer nonlinear transformation, the PLV feature vector is mapped to a low-dimensional space while retaining the core mode. : ,in express Activation function For PLV feature vectors, These are learnable weights; S95, Cross-modal feature fusion: Will and The fused features are obtained through adaptive weight fusion. : ,in For the fusion weights of network learning; S96. Classification result output: The fused features First, it goes through two layers of fully connected network processing: ， The first layer uses a weight matrix. The feature dimension is compressed to 128 dimensions while retaining core discriminative information. To prevent overfitting, a function is used in the second layer; the weight matrix is ​​used... Further refine the key relationships between features; pass softmax The function maps the processed features to three probability distributions: , The weight matrix is ​​output. ,in The probabilities of having only ASD, having both ADHD and ASD, and having only ASD are given, respectively. The category corresponding to the highest probability is taken as the final prediction result.

7. The neurodevelopmental disorder comorbidity identification system based on residual graph neural network according to claim 6, characterized in that, Low-frequency amplitude fluctuations are calculated by each The amplitude values ​​of low-frequency oscillation signals in four frequency bands were obtained. The differences in low-frequency fluctuation values ​​of amplitude in each frequency band between the comorbidity group and the single disease group were compared using independent samples testing. Statistically significant abnormal brain regions were screened out and output.

Images