Autism detection system and method based on multi-frequency contrast learning

Abstract

The invention discloses an autism detection method based on multi-frequency contrast learning. The method comprises the following steps: performing frequency division preprocessing on electroencephalogram signals to obtain electroencephalogram signals of a plurality of frequency bands; constructing a graph network for the electroencephalogram signals of the multiple frequency bands; the graph network is enhanced by adopting a node discarding strategy and an edge disturbance strategy to obtain an enhanced graph network, the graph network of each frequency band is enhanced into a node discarding type network and an edge disturbance type network, the node discarding strategy is used for randomly discarding part of nodes and connection thereof, and the edge disturbance type network is used for randomly discarding part of the nodes and connection thereof. The edge disturbance strategy is used for randomly disturbing the connection strength of the edges; for the enhanced graph network, performing fusion by adopting cross attention and global attention to obtain fused graph network data; and taking the fused graph network data as input, and utilizing a trained multi-frequency contrast learning model to obtain an autism detection result. According to the invention, the expression ability and accuracy of autism detection are improved.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more specifically, to an autism detection system and method based on multi-frequency contrastive learning. Background Technology

[0002] Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by impaired social and interpersonal communication skills, emotional comprehension and expression, and repetitive and stereotyped behaviors in early childhood. Currently, the diagnosis of ASD relies primarily on subjective symptom descriptions. However, the lack of reliable biomarkers and the heterogeneity of symptoms have led to controversy regarding diagnostic criteria. The etiology and pathogenesis of ASD remain unclear, further complicating the understanding of the neural mechanisms behind its clinical features and resulting in a lack of highly reproducible and reliable objective diagnostic methods. Therefore, there is an urgent clinical need for precise and objective diagnostic indicators.

[0003] Neural oscillations are electrical activities of groups of neurons occurring synchronously at specific frequencies, manifesting as different brain electrical rhythms. Electroencephalography (EEG) is a high-temporal-resolution, non-invasive method widely used to record these neural oscillations. Different EEG signal rhythms reflect the electrophysiological characteristics of the brain under various states and activities. For example, delta waves are commonly seen in patients with severe organic brain disease, while beta waves are associated with mental stress or emotional excitement. Studies have shown that individuals with ASD exhibit abnormal oscillations in multiple EEG rhythms. Notably, the power spectral density (PSD) in the alpha band of infants with ASD shows asymmetry between brain hemispheres. Previous studies have found that the PSD distribution in ASD patients follows a "U-shape," with increased power in the theta and gamma bands and decreased power in the alpha band compared to controls. Furthermore, functional connectivity of brain regions differs significantly among ASD patients under different EEG rhythms, exhibiting marked patterns of underconnection and overconnection. Utilizing EEG information from different frequency bands may provide new insights into characteristic indicators for identifying children with ASD.

[0004] With the advancement of artificial intelligence technology, machine learning methods based on neuroimaging data have made significant progress in ASD classification. However, traditional supervised learning methods require extensive background knowledge and manual data annotation, limiting their application. Self-supervised learning (SSL) offers a promising alternative, reducing reliance on labeled datasets by designing pre-tasks to extract useful representations from unlabeled data. SSL also allows models to learn correlations between frequency components, capture complex patterns in spectral signals, enhance generalization and noise robustness, and facilitate cross-task transfer learning. These advantages make SSL particularly suitable for spectral data processing in ASD classification. For example, researchers have proposed a task-agnostic self-supervised ensemble framework for ASD classification. However, existing research mainly focuses on resting-state data, which exhibits significant inter-individual variability and fails to adequately reflect clinical problems and cognitive function. This variability poses a challenge in accurately distinguishing individuals with ASD from those with neurotypical behavior. In contrast, task-related EEG signals can capture an individual's neural activity in a specific task, offering advantages in areas where individuals with ASD often struggle, such as emotion recognition and face recognition. These insights are crucial for studying the neural mechanisms of ASD and developing more effective diagnostic tools.

[0005] Analysis reveals the following main shortcomings in existing technologies:

[0006] 1) Existing methods for assisting in the diagnosis of autism largely rely on modeling using single-band EEG signals. This single-band approach fails to effectively utilize the rich cross-band information in EEG. EEG is a multi-frequency physiological signal containing complex temporal and spectral characteristics. Each band conveys different neural information and reflects diverse brain functions and cognitive activities. Traditional methods ignore the potential correlations between bands, limiting the model's ability to capture cross-band features.

[0007] 2) Traditional EEG signal modeling is not robust to noise and missing data, especially when EEG signals are represented in graph structure. Existing technologies rarely introduce methods specifically for handling graph structure noise, making the model susceptible to interference when faced with complex data.

[0008] 3) Traditional frequency band fusion strategies are prone to information loss in cross-frequency band feature capture due to information integration lag, which limits the generalization ability of the model.

[0009] 4) Existing contrastive learning frameworks, when processing multi-frequency EEG signals, rely on late fusion, making it difficult to capture rich multi-frequency information in the early stages. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide an autism detection system and method based on multi-frequency contrastive learning.

[0011] According to a first aspect of the present invention, an autism detection method based on multi-frequency contrastive learning is provided. The method includes:

[0012] Frequency division preprocessing is performed on the EEG signals to obtain EEG signals in multiple frequency bands;

[0013] A graph network is constructed for the EEG signals of the multiple frequency bands, where nodes represent regions of interest in the brain and edges represent the functional connectivity strength between different regions of interest;

[0014] The graph network is enhanced by employing a node dropping strategy and an edge perturbation strategy to obtain an enhanced graph network. The graph network of each frequency band is enhanced into a node dropping network and an edge perturbation network, respectively. The node dropping strategy is used to randomly drop some nodes and their connections, and the edge perturbation strategy is used to randomly perturb the connection strength of the edges.

[0015] For the enhanced graph network, cross attention and global attention are fused to obtain fused graph network data;

[0016] Using the fused graph network data as input, the autism detection results are obtained by utilizing a trained multi-frequency contrastive learning model.

[0017] According to a second aspect of the present invention, an autism detection system based on multi-frequency contrastive learning is provided. The system includes:

[0018] Signal preprocessing module: used to perform frequency division preprocessing on EEG signals to obtain EEG signals in multiple frequency bands;

[0019] Graph network construction module: used to construct a graph network for the EEG signals of the multiple frequency bands, where nodes represent regions of interest in the brain and edges represent the functional connectivity strength between different regions of interest;

[0020] Graph network optimization module: used to enhance the graph network using a node dropping strategy and an edge perturbation strategy to obtain an enhanced graph network, wherein the graph network of each frequency band is enhanced into a node dropping network and an edge perturbation network respectively. The node dropping strategy is used to randomly drop some nodes and their connections, and the edge perturbation strategy is used to randomly perturb the connection strength of the edges.

[0021] Early fusion module: used to fuse the enhanced graph network using cross attention and global attention to obtain fused graph network data;

[0022] Detection module: Used to obtain autism detection results by taking the fused graph network data as input and utilizing a trained multi-frequency contrastive learning model.

[0023] Compared to existing technologies, the advantages of this invention lie in its provision of a novel self-supervised learning (SSL) framework for ASD diagnosis, which incorporates abnormal patterns from the brain functional networks of children with ASD. By integrating the SSL model with electrophysiological acquisition techniques, it identifies ASD patients using emotion-evoked stimuli. This invention uses multi-frequency oscillatory self-difference as a signal to supervise the dynamic reconstruction and characterization of nodes and edges in the brain functional networks of ASD. The proposed MF-CLM model integrates information before feature extraction, improving feature extraction efficiency and maximizing the differences in connectivity edges between ASD patients and typical developmental (TD) individuals in the evoked functional networks, thus learning highly interpretable features. This interpretable representation of the model provides in-depth insights into the neurodynamics of ASD and offers reasonable targets for subsequent neurological treatments.

[0024] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0026] Figure 1 This is a general framework diagram of an autism detection scheme based on multi-frequency contrastive learning according to an embodiment of the present invention;

[0027] Figure 2 This is a flowchart of an autism detection method based on multi-frequency contrastive learning according to an embodiment of the present invention;

[0028] Figure 3 This is a flowchart of preprocessing electroencephalogram (EEG) signals according to an embodiment of the present invention;

[0029] Figure 4 This is a flowchart of constructing a graph network according to an embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of a multi-frequency contrast learning process according to an embodiment of the present invention;

[0031] Figure 6 This is a schematic diagram illustrating the classification performance of a multi-frequency contrastive learning model according to an embodiment of the present invention;

[0032] Figure 7 This is a schematic diagram of significant biomarkers of ASD and TD under different cross-frequency information according to an embodiment of the present invention. Detailed Implementation

[0033] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0034] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0035] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0036] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0037] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0038] In summary, this invention designs a multi-frequency contrastive learning (MF-CLM) framework to acquire cross-frequency EEG signal features relevant to the diagnosis of autism spectrum disorder (ASD), covering neural information associations and differences across multiple frequency bands. It enhances the robustness of the graph network to data noise and uncertainty through NodeDropping and Edge Perturbation techniques. An early fusion strategy is employed to efficiently integrate multi-frequency features in the early stages of training, utilizing cross-interaction and global attention mechanisms to improve the capture of cross-frequency associations. Through improved multi-frequency contrastive learning, RNN (Recurrent Neural Network) subgraph sampling and GNN (Graph Neural Network) subgraph clustering are introduced to intelligently extract latent variable features across multiple frequency bands, achieving efficient diagnosis of ASD. For example, the provided MF-CLM model includes a source localization module, a frequency-based graph network construction and optimization module, and a multi-frequency contrastive learning module. The MF-CLM model optimizes the enhanced consistency of the same data at different frequencies by using contrastive loss in the latent space, promoting representation learning. See also... Figure 1 As shown, the MF-CLM model generally includes processes such as EEG signal preprocessing, graph network construction and optimization, multi-frequency comparison and visualization diagnosis, and can be used for early detection of intelligence in autistic patients.

[0039] Specifically, see Figure 2 As shown, the autism detection method based on multi-frequency contrastive learning includes the following steps:

[0040] Step S110: Preprocess the EEG signals to obtain EEG signals in multiple frequency bands.

[0041] Electroencephalography (EEG) is a multi-frequency physiological signal with complex temporal and spectral characteristics. Each frequency band carries different neural information, reflecting different brain functions and cognitive activities. Traditional EEG modeling typically focuses on only a single frequency band, limiting the capture of potential correlations between cross-frequency signals. The Multi-Frequency Contrast Learning Model (MF-CLM) proposed in this invention integrates EEG signals from multiple frequency bands, enabling simultaneous capture of correlations and differences between bands and enhancing the model's representational power. By comparing brain functional networks at different frequencies, this model deepens the understanding of neurological disorders such as autism spectrum disorder (ASD), providing a new pathway to improve diagnostic accuracy and optimize treatment interventions.

[0042] For example, in the preprocessing of EEG signals, artifact removal, filtering, and eye-tracking artifact removal are first performed to improve the quality of the EEG signal. Then, a Fast Fourier Transform (FFT) is used to decompose the EEG signal into different frequency bands (e.g., δ, θ, α, β, γ), each reflecting different neural activity characteristics and helping to capture brain functional connectivity at different frequencies. Next, based on the frequency-divided EEG data, the temporal progression of regions of interest (ROIs) in the brain atlas is extracted, and phase-locked value (PLV) is used to measure the phase synchronization between brain regions, constructing functional connectivity maps for each frequency band. In the functional connectivity map, nodes represent brain ROIs, and edges represent the strength of functional connectivity between different ROIs.

[0043] See Figure 3 As shown, in one embodiment, preprocessing the EEG signal includes: importing raw EEG data; calculating all extreme points of the EEG power signal; obtaining the upper and lower envelope curves of the signal using spline interpolation; calculating the mean of the upper and lower envelope curves; subtracting the mean envelope from the EEG power signal to obtain the IMF (Intrinsic Mode Function); determining whether the current IMF component is the highest frequency; if the IMF component is the highest frequency, subtracting the current IMF from the EEG power signal to obtain a new source signal; if the new source signal function is monotonic, the decomposition ends; if the IMF component is not the highest frequency, using the current IMF component as the new source signal.

[0044] By combining frequency division and source tracing, we can not only comprehensively capture brain activity from both temporal and frequency perspectives, but also reveal the interactions between different brain regions across various frequency bands by analyzing the functional connections of nodes and edges. This allows for a more precise analysis of the spatial and frequency characteristics of brain functional networks. This process contributes to a deeper understanding of the brain's neural mechanisms and provides more comprehensive data support for subsequent analysis and diagnosis.

[0045] Step S120: Construct a graph network for EEG signals of multiple frequency bands, and enhance the graph network by using a node drop-off strategy and an edge perturbation strategy to obtain an enhanced graph network, wherein the graph network of each frequency band is enhanced into a node drop-off network and an edge perturbation network.

[0046] See Figure 4 As shown, two optimization techniques, Node Dropping and Edge Perturbation, are introduced during the construction of the Graph Neural Network (GNN) to improve the network's robustness and generalization ability. Node Dropping simulates the information loss that may occur in the functional networks of the brain by randomly dropping some nodes and their connections.

[0047] Specifically, given a graph G = (V, E), where V is the set of nodes and E is the set of edges, the adjacency matrix of the graph is represented by A, and the node feature matrix is ​​X. Node Dropping randomly selects a subset. As the retained set of nodes, its new adjacency matrix A′ and characteristic matrix X′ can be represented as:

[0048] A′=A[V′,V′] (1)

[0049] X′=X[V′] (2)

[0050] In this case, the node update formula for GNN changes from the original:

[0051]

[0052] It becomes:

[0053]

[0054] Where l represents the number of network layers, H′ (l) W represents the updated feature vector of the l-th layer. (l) This represents a learnable weight matrix.

[0055] This method of randomly discarding some nodes allows the model to still work effectively in incomplete network structures, thereby enhancing its robustness to uncertainty and missing data.

[0056] Edge perturbation simulates the uncertainty in connectivity by randomly perturbing the edge connection strength. For example, given a graph G with edge weight matrix W... e Edge Perturbation can randomly perturb the connection strength of edges in two ways.

[0057] The first method is to directly perturb the adjacency matrix A, such that:

[0058] A′=A+ΔA (5)

[0059] Here, ΔA is a matrix for adding or deleting edges, generated randomly.

[0060] The second type is perturbation edge weights:

[0061] W e ′(u,v)=W e (u,v)+∈ u,v (6)

[0062] Where, ∈ u,v It is random noise connected to the edge, W e (u,v) represents the edge weight matrix of the original edge (u,v). After perturbation, the propagation formula of GNN becomes:

[0063]

[0064] By perturbing the edges in this way, the model can better adapt to the uncertainties and noise variations in real-world data.

[0065] By combining Node Dropping and Edge Perturbation, the model becomes more robust to real-world data noise and structural changes, thus enhancing its ability to adapt to different scenarios and conditions. Node Dropping simulates missing nodes, and Edge Perturbation simulates random changes in edge connections. During training, the graph network can learn how to effectively extract and predict features in incomplete or uncertain graph structures, further improving the model's robustness and generalization ability.

[0066] Step S130: For the enhanced graph network, an early fusion strategy is adopted to perform fusion to obtain fused graph network data. The early fusion strategy includes cross-fusion of different frequency bands and fusion of different graph networks within the same frequency band.

[0067] The early fusion strategy integrates multi-dimensional information from different frequency bands and optimization methods in the early stages of training by utilizing cross-attention and global attention mechanisms after the graph network is optimized. This enhances the capture of cross-frequency band features and avoids information loss caused by insufficient fusion in the later stages.

[0068] 1) Cross-integration of different frequency bands

[0069] Cross-attention mechanisms are used to fuse networks from different frequency bands. After optimizing two brain functional networks from different frequency bands (e.g., frequency band 1 and frequency band 2) using Node Dropping and Edge Perturbation respectively, four networks are obtained: a Node Dropping-based network for frequency band 1. Edge Perturbation type And the Node Dropping type of band 2 Edge Perturbation type

[0070] The cross-attention mechanism calculates the feature-fused representation by combining the query vector Q of one network with the key vector K and value vector V of another network. For the Node Dropping type in band 1... and Band 2 NodeDropping type The cross-fusion formula is as follows:

[0071]

[0072] in, W k and W v It is the weight of the corresponding item, d k This indicates the scaling factor. Through this cross-attention mechanism, the model can capture the interdependencies between the Node Dropping type networks in band 1 and band 2. Similarly, for the Edge Perturbation type network in band 1... And the EdgePerturbation type of band 2 Cross-attention fusion can be performed using the same formula.

[0073] After fusing Node Dropping and Edge Perturbation networks across different frequency bands using a cross-attention mechanism, the next step is to superimpose these two fused features. The model further integrates the feature representations generated by the cross-attention mechanism through a global attention mechanism. The formula for calculating global attention is:

[0074]

[0075] Q′, K′, and V′ are derived from the two feature representations after cross-attention fusion. This approach enables the model to capture more complex cross-band dependencies on a global scale.

[0076] 2) Integration of different networks within the frequency band

[0077] Besides cross-band fusion, different optimization methods (Node Dropping and EdgePerturbation) within the same frequency band can also be fused. In frequency band 1, a cross-attention mechanism can be used to fuse Node Dropping networks. With Edge Perturbation type network Fusion, represented as:

[0078]

[0079] in, Indicates a Node Dropping type network With Edge Perturbation type network The feature representation generated by the cross-attention mechanism, d k Indicates the scaling factor.

[0080] This fusion can capture the internal dependencies between NodeDropping and Edge Perturbation networks in Band 1, thereby improving the feature representation capability within Band 1. The same operation is performed on Band 2, combining NodeDropping networks... and Edge Perturbation type The network uses cross-attention to fuse features, resulting in a fused representation of features within frequency band 2.

[0081] By introducing an early fusion strategy that combines cross-attention and global attention mechanisms, the model can simultaneously fuse Node Dropping and Edge Perturbation networks across different frequency bands, capturing multidimensional information both across and within frequency bands. Cross-attention establishes connections between and within frequency bands, while global attention further integrates these connections, improving the understanding and classification accuracy of brain functional networks.

[0082] Step S140: Using the fused graph network data as input, the autism detection results are obtained by utilizing a multi-frequency contrastive learning model.

[0083] Figure 5 This is a schematic diagram of multi-frequency contrastive learning. In one embodiment, an improvement upon the self-supervised contrastive learning framework SimCLR is implemented. Unlike previous late fusion strategies, this invention introduces RNN subgraph sampling and GNN subgraph clustering on top of early fusion, enabling the model to capture more multi-frequency information during contrastive learning. Furthermore, the improved enhancer avoids graph rotation or pruning operations in traditional SimCLR, better preserving network structural features and thus improving the learning effect on latent variables. Multi-frequency contrastive learning compares the differences in brain functional networks at different frequencies, replacing positive / negative samples in contrastive learning with different frequency bands to mine the network's latent feature representations. Brain functional networks include GNNs... δ To G γ Assume G δ As a positive sample, and G βAs a negative sample, its loss function is as follows:

[0084]

[0085] in, This represents the information loss function, used to measure the similarity between positive sample pairs and the dissimilarity between negative sample pairs. Let represent the triplet loss function, used to ensure that the distance between a sample (anchor point) and another sample (positive sample) is less than the distance with a third sample (negative sample). τ represents the temperature parameter, used to control the smoothness of the loss function and affect the scaling of similarity. μ and λ represent the weight parameters of the loss function, and N represents the number of samples. [k≠u] Let l represent an indicator function, and l represent the final loss function, which combines information loss and triplet loss. Margin is a constant greater than 0, and the distance can be calculated using cosine for the d(·) function, μ+λ=1. Then, the t-SNE method is used to obtain pseudo-labels, and the difference values ​​between all individual frequency bands are obtained according to formula (4), and then their joint distribution p is calculated. i，j and low-dimensional space divergence q i，j Where i and j represent the subject's index, the objective function for the pseudo-label is:

[0086]

[0087] Where C represents the KL divergence, used to measure the difference between two probability distributions, and KL(P||Q) represents the KL divergence from distribution P to distribution Q, p i，j q i，j Let q represent the joint distribution of different subjects i and j, where i and j represent the subject's index. δ，β p represents the probability of similarity between samples i and j in probability distribution Q. δ，β y represents the probability of similarity between samples i and j in probability distribution P. i y j This represents a sample individual.

[0088] 1) RNN subgraph sampling

[0089] RNN subgraph sampling takes the early-fused graph network data as input. This early-fused network contains rich, multi-frequency complex information, but directly processing this information is computationally expensive and prone to overfitting. Therefore, RNN subgraph sampling is introduced to reduce the computation of complex networks. Unlike random subgraph sampling, RNN subgraph sampling can effectively capture temporal dependencies and preserve recursive information between frequency bands, making it suitable for processing multi-frequency brain networks with temporal dynamics, thereby improving the model's representational power. LSTM and SMILES are combined for variational autoencoders (VAEs). Specifically, an RNN is a specialized neural network that, for each symbol s... t ∈S compute a hidden state h t It can learn sequence or event dependencies F(·), and the principle is as follows:

[0090] i t =σ(W st s t +b ii +W hi h t-1 +b hi (17)

[0091] f t =σ(W sf s t +b if +W hf h t-1 +b hf (18)

[0092] g t =tanh(W sg s t +b ig +W hg h t-1 +b hg (19)

[0093] o t =σ(W so s t +b io +W ho h t-1 +b ho (20)

[0094] c t =f t ☉c t-1 +i t ☉g t (twenty one)

[0095] h t =o t ☉tanh(c t) (twenty two)

[0096] Where W and b are the weighted exponent and bias, respectively, their subscripts indicate the dimension, and ⊙ is the Hadamard product. σ and tanh are the sigmoid logistic function and the hyperbolic tangent activation function, respectively. To obtain s t c t-1 and c t The information is processed using three gating mechanisms in the LSTM, including the input gating mechanism i. t Forgetting Gating f t and output gating o t .

[0097] 2) GNN subgraph clustering

[0098] To further extract structural and functional information from multi-frequency networks, GNN subgraph clustering is employed. This method, based on the feature similarity of nodes and edges, dynamically learns the potential correlations between multi-frequency bands through a graph neural network, thereby achieving more accurate subgraph clustering. Compared to traditional clustering methods (such as spectral clustering and k-means), GNN subgraph clustering can capture complex node relationships in the topological and feature spaces, making it particularly suitable for processing multi-frequency brain functional networks. GNN encodes graph structural information iteratively to obtain local and global information about the graph. It includes two stages: first, acquiring a neighborhood aggregation strategy to learn node structural information; second, automatically grouping local representations into a graph set. Information transfer between graph sets is as follows:

[0099]

[0100] in, It is the embedding vector of v in the (t+1)th layer. N (v) M is an adjacent vertex of v. t (·) represents the information passed between adjacent vertices, U t (·) represents the update function for the vertex. Its hidden state update rules are as follows:

[0101]

[0102] Among them, M t (·), U t (·) and R t (·) The function that determines the learning parameters by different forms of GNN.

[0103] Finally, the normalized cross-loss function (NT-Xent) is used for classification, and the principle is as follows:

[0104]

[0105] in, zi and z j Feature vectors are extracted from encoders of different sources.

[0106] Step S150: Visualize the autism test results.

[0107] In one embodiment, the diagnosis of autism spectrum disorder (ASD) can be visualized using Class Activation Mapping (CAM) technology to explain the model's decision-making mechanism and identify brain regions that play a key role in classification results. Specifically, CAM generates a heatmap by combining convolutional layer feature maps and classification layer weights from a deep learning model to highlight the regions that are most active in specific categories of ASD.

[0108] 1) Convolutional feature maps and classification weights

[0109] First, the brain functional network based on EEG signals is input and processed by a convolutional neural network (CNN) to obtain the feature maps F of the convolutional layers. k (x), where k represents the k-th feature map. Next, the weights in the classification layer related to the category (ASD or TD) are utilized. The sum is then calculated by weighting the sum with the corresponding feature map.

[0110] 2) Calculation of category activation graph

[0111] Activation map M of category c c The formula for calculating (x) is:

[0112]

[0113] Formula (26) generates an activation map that reflects the degree of activation in brain regions by combining feature maps with classification weights. High-value regions represent brain regions or frequency bands that contribute most to the classification of ASD or TD (typical developmental disorder).

[0114] 3) Heatmap generation and visualization

[0115] The final activation map, after interpolation, generates a heatmap that clearly identifies the brain functional regions or frequency bands that play a key role in the classification process. This heatmap helps locate the brain regions that most contribute to distinguishing ASD patients from healthy individuals with TD, thus providing a more intuitive explanation and support for the diagnosis of ASD.

[0116] Accordingly, the present invention also provides an autism detection system based on multi-frequency contrastive learning, used to implement one or more aspects of the above-mentioned method. For example, the system includes: a signal preprocessing module for frequency-division preprocessing of EEG signals to obtain EEG signals in multiple frequency bands; a graph network construction module for constructing a graph network for the EEG signals in the multiple frequency bands, where nodes represent regions of interest in the brain and edges represent the functional connectivity strength between different regions of interest; a graph network optimization module for enhancing the graph network using a node discarding strategy and an edge perturbation strategy to obtain an enhanced graph network, wherein the graph network for each frequency band is enhanced into a node discarding network and an edge perturbation network, respectively, the node discarding strategy for randomly discarding some nodes and their connections, and the edge perturbation strategy for randomly perturbing the connection strength of edges; an early fusion module for fusing the enhanced graph network using cross-attention and global attention to obtain fused graph network data; and a detection module for using the fused graph network data as input and a trained multi-frequency contrastive learning model to obtain autism detection results. The modules in the system can be implemented using dedicated processors, general-purpose processors, or FPGAs, etc.

[0117] To further verify the effectiveness of this invention, the classification performance of the proposed model was quantitatively evaluated. Four widely accepted metrics were used for comprehensive evaluation: accuracy (ACC), area under the receiver operating characteristic (AUC), sensitivity, and 1-specificity. The results are as follows: Figure 6 As shown, when multi-frequency contrastive learning processes both high-frequency and low-frequency EEG information simultaneously, the overall classification performance is optimal when the model uses θ and γ (the sum of the four metrics is 3.28). The synergistic integration of different frequency bands significantly improves the model's discriminative ability, which is reflected in the superior performance metrics obtained.

[0118] To elucidate the unique characteristics and effectiveness of the model, a quantitative comparative analysis was conducted on existing self-supervised learning (SSL) algorithms and previous supervised models, using the same four metrics: accuracy (ACC), AUC, sensitivity, and 1-specificity. Referring to Tables 1 and 2, the results show that MF-CLM outperforms previous classification models in terms of classification performance and effectiveness, specifically: ACC 84.0%, AUC 81.1%, sensitivity 83.2%, and 1-specificity 78.0%. Compared with other SSL algorithm models, the model of this invention demonstrates superior performance by utilizing a multi-frequency information module and feature extraction strategy. The multi-frequency contrastive learning module allows for simultaneous analysis of different EEG bands, thereby capturing a wider range of patterns across bands. This feature is superior to traditional SSL algorithms that typically focus on single-band features, which are more limited in capturing individual-specific neural activity in ASD. Furthermore, this model integrates a sequential feature fusion mechanism, enhancing its ability to capture time-series and cross-band information by capturing the temporal dependence and contextual information of features, outperforming other existing methods that isolate features or ignore temporal order.

[0119] Table 1: Comparative Analysis of Different Models

[0120]

[0121] Table 2: Classification performance of different models

[0122]

[0123] Furthermore, the MF-CLM designed in this invention is interpretable. Figure 7 These are significant biomarkers for ASD and TD under different cross-frequency information, among which Figure 7 (a) is the significant discriminant in the SMN. Figure 7 (b) is the significant discriminant in FPN. Figure 7 (c) represents the significant discriminative power in VMN. Figure 7 (a) shows that the most significant discriminative power comes from the cross-frequency interactions of the delta and theta networks, primarily concentrated in the sensorimotor network (SMN), which includes the supplementary motor area (SMA), primary motor cortex (PMC), somatosensory cortex (SC), and auditory cortex (AC). Notably, children with autism spectrum disorder (ASD) exhibit higher activation in these regions of interest (ROIs) compared to children with typical development (TD).

[0124] In contrast, Figure 7(b) The frontoparietal network (FPN) comprising the dorsolateral prefrontal cortex (DLPFC), anterior cingulate cortex (ACC), dorsomedial prefrontal cortex (dmPFC), and insula is depicted. This network exhibits significant discriminative power due to cross-frequency interactions between the β and γ networks. In this context, children with ASD show lower activation in the FPN nodes than children with TD. Furthermore, cross-frequency interactions between the θ and β networks produce significant discriminative power in the visual-motor network (VMN), such as... Figure 7 As shown in (c), this network comprises the intraparietal sulcus (IPS), primary motor cortex (PMC), DLPFC, visual cortex (VC), inferior parietal lobule (IPL), and dmPFC. Within this network, the activation patterns of children with ASD exhibit a complex distribution, with enhanced activation in some ROIs and weakened activation in others. By applying Granger causality analysis (GCA) to investigate the interrelationships among these subnetworks, we gained a comprehensive understanding of their causal connections. Figure 7 This indicates that overactivated networks across low frequencies influence underactivated and mixed-frequency networks across high frequencies. Similarly, low-interaction networks across high frequencies affect mixed-frequency networks. These findings reveal information flow and interactions across a wide range of networks. The results demonstrate that MF-CLM effectively captures the intrinsic representation of significant differences in brain activity between individuals with ASD and a typical control group.

[0125] In summary, compared with the prior art, the present invention has the following advantages:

[0126] 1) This invention utilizes the multi-frequency contrastive learning model (MF-CLM) to capture the correlations and differences between frequency bands. By integrating signals from multiple frequency bands, it achieves in-depth capture of the correlations and differences between frequency bands, overcoming the problems of feature loss and inaccuracy in traditional EEG signal processing, and improving the expressive power and accuracy of the autism detection model.

[0127] 2) This invention addresses the noise and uncertainty issues present in graph-structured EEG data by employing Node Dropping and Edge Perturbation techniques to simulate the complexity and potential missing data of real data at the node and edge levels, respectively. This enhances the model's robustness to noise and optimizes the modeling effect on graph-structured data.

[0128] 3) This invention achieves brain network data augmentation through RNN subgraph sampling and GNN subgraph clustering, and captures multi-frequency information of the brain network through an improved contrastive learning framework, thereby enhancing data augmentation. RNNs are used to sample subgraphs of functional networks, capturing cross-frequency recursive dependencies in time series, thus strengthening the modeling of dynamic brain functions. GNNs extract spatial structural features of functional networks through subgraph clustering, revealing local and global connectivity patterns in the brain network. This combined strategy enables multi-frequency contrastive learning to more robustly extract and express complex features of brain functional networks, effectively improving the ability to analyze dynamic and complex brain activities.

[0129] 4) An integrated early fusion strategy is implemented by introducing cross-attention and global attention mechanisms after graph network optimization. This strategy integrates different frequency bands and multi-dimensional optimization methods in the early stages of training, effectively capturing cross-frequency band features and avoiding information loss caused by insufficient fusion in later stages. This ensures that the data has global consistency and high feature representation ability from the beginning. The designed early fusion strategy integrates multi-frequency band features in the early stages of training and uses cross-attention and global attention mechanisms to capture complex cross-frequency band correlations, effectively improving the adaptability to multi-frequency features, enhancing the model's generalization ability, and avoiding the frequency band information loss problem caused by later fusion methods.

[0130] 5) Building upon earlier fusion methods, this invention improves the SimCLR framework by incorporating RNN subgraph sampling and GNN subgraph clustering, enabling the model to capture multi-frequency information more fully during contrastive learning. Furthermore, the designed enhancer avoids graph rotation or pruning operations commonly used in traditional SimCLR, better preserving network structural features and thus improving the learning effect on latent variables and optimizing the overall modeling capability for multi-frequency data.

[0131] 6) Experimental verification shows that this invention achieves high-precision diagnosis of ASD patients by mining neural markers of electrophysiological signals from ASD patients. Furthermore, the results of model learning are interpreted from multiple perspectives.

[0132] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0133] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0134] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0135] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as "C" or similar languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0136] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0137] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0138] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0139] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0140] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. An autism detection method based on multi-frequency contrastive learning, comprising the following steps: Frequency division preprocessing is performed on the EEG signals to obtain EEG signals in multiple frequency bands; A graph network is constructed for the EEG signals of the multiple frequency bands, where nodes represent regions of interest in the brain and edges represent the functional connectivity strength between different regions of interest; The graph network is enhanced by employing a node dropping strategy and an edge perturbation strategy to obtain an enhanced graph network. The graph network of each frequency band is enhanced into a node dropping network and an edge perturbation network, respectively. The node dropping strategy is used to randomly drop some nodes and their connections, and the edge perturbation strategy is used to randomly perturb the connection strength of the edges. For the enhanced graph network, cross attention and global attention are fused to obtain fused graph network data; Using the fused graph network data as input, the autism detection results are obtained by utilizing a trained multi-frequency contrastive learning model.

2. The method according to claim 1, characterized in that, For node drop-off networks in the first frequency band Second-band node drop-off network The cross attention and the global attention are respectively represented as follows: Where Attention(·) represents cross-attention, and GlobalAttention(·) represents global attention. Q′, K′, V′ are derived from the feature representations after cross-attention fusion, d k Indicates the scaling factor.

3. The method according to claim 2, characterized in that, It also includes cross-attention fusion of different networks within the frequency band, represented as: in, This represents the side-perturbation type network in the first frequency band. This represents a node drop-off network in the first frequency band. Side-disturbance type network of the first frequency band The feature representation generated by the cross-attention mechanism.

4. The method according to claim 1, characterized in that, During the training of the multi-frequency contrastive learning model, the overall loss function is set as follows: in: Among them, G δ G is the graph network corresponding to frequency band δ, used as positive samples, and G is the graph network corresponding to frequency band γ, used as negative samples. Represents the information loss function. Let represent the triplet loss function, τ be a coefficient used to control the smoothness of the information loss function, μ and λ be the weight parameters of the corresponding terms, N be the number of samples, 1[k≠i] be the indicator function, margin be a constant greater than 0, and d(·) function indicate that the distance is calculated using the cosine function; During the training of the multi-frequency contrastive learning model, the objective function for the pseudo-labels is set as follows: in: Where C represents the KL divergence, used to measure the difference between two probability distributions, and KL(P||Q) represents the KL divergence from distribution P to distribution Q, p i，j q i，j Let q represent the joint distribution of different subjects i and j, where i and j represent the subject's index. δ，β p represents the probability of similarity between samples i and j in probability distribution Q. δ，β y represents the probability of similarity between samples i and j in probability distribution P. i y j This represents a sample individual.

5. The method according to claim 1, characterized in that, The edge perturbation strategy is to directly perturb the adjacency matrix A, such that: A′=A+ΔA Here, ΔA is a matrix for adding or deleting edges, generated randomly.

6. The method according to claim 1, characterized in that, The edge perturbation strategy is to perturb the edge weights, expressed as: W′ e (u,v)=W e (u,v)+∈ u，v Where, ∈ u，v It is random noise connected to the edge, W e (u, v) represents the edge weight matrix of the original edge (u, u).

7. The method according to claim 1, characterized in that, The fused graph network data further includes: extracting temporal dependencies by sampling subgraphs using a recurrent neural network and learning potential correlations between multiple frequency bands by using graph neural network subgraph clustering based on the feature similarity of nodes and edges, thereby achieving subgraph clustering.

8. The method according to claim 1, characterized in that, Also includes: The autism detection results are visualized based on category activation maps and heatmaps, where the activation map M for category c is shown below. c (x) is represented as: Among them, F k (x) represents the k-th feature map. F represents k The weight of (x).

9. An autism detection system based on multi-frequency contrastive learning, comprising: Signal preprocessing module: used to perform frequency division preprocessing on EEG signals to obtain EEG signals in multiple frequency bands; Graph network construction module: used to construct a graph network for the EEG signals of the multiple frequency bands, where nodes represent regions of interest in the brain and edges represent the functional connectivity strength between different regions of interest; Graph network optimization module: used to enhance the graph network using a node dropping strategy and an edge perturbation strategy to obtain an enhanced graph network, wherein the graph network of each frequency band is enhanced into a node dropping network and an edge perturbation network respectively. The node dropping strategy is used to randomly drop some nodes and their connections, and the edge perturbation strategy is used to randomly perturb the connection strength of the edges. Early fusion module: used to fuse the enhanced graph network using cross attention and global attention to obtain fused graph network data; Detection module: Used to obtain autism detection results by taking the fused graph network data as input and utilizing a trained multi-frequency contrastive learning model.

10. A computer-readable storage medium having a computer program stored thereon, wherein, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 8.

Images