Developmental language disorder early recognition method and system fused with neural activity features

Abstract

The invention relates to a neural activity feature fused developmental language disorder early recognition method and system, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring an original neurophysiological signal and matched clinical information, and constructing a data set; performing normalization processing on the power feature and the coherence feature of the signal to generate an enhanced feature vector; obtaining an attention weighted feature vector through attention mechanism weighting based on mutual information; performing nonlinear mapping to a high-dimensional space by using a radial basis function to obtain an extended feature vector; the method comprises the following steps: constructing a developmental language disorder early recognition model, introducing a decision function fused with a nonlinear interaction item by adopting brain region level feature aggregation and modulation, clinical factor fusion and gating modulation, and optimizing the model by a binary cross entropy loss function; and after model training is completed by using the training set, inputting a to-be-evaluated tested signal into the model after same preprocessing, and outputting an identification result. According to the invention, the accuracy of early recognition of the developmental language barrier can be improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for early identification of developmental language disorders that integrates neural activity features. Background Technology

[0002] In the field of developmental language disorder research, early identification and intervention are crucial for improving children's prognosis. However, traditional clinical diagnosis mainly relies on standardized language behavior assessments and expert interviews. While these methods are widely used clinically, they have inherent limitations such as high subjectivity, insufficient sensitivity to young or uncooperative children, and difficulty in capturing deep neurobiological abnormalities in the brain. With the development of cognitive neuroscience, it has become increasingly clear that language disorders are often accompanied by systemic abnormalities in brain neural activity patterns, especially the rhythmicity of neural oscillations at rest and the dysfunction of functional connections between brain regions. This provides a theoretical basis for using non-invasive neurophysiological signals, such as electroencephalography (EEG) or magnetoencephalography (MEG), for objective identification of biological markers. Although existing technologies have attempted to apply machine learning to this type of neural signal analysis, a series of key challenges remain. Existing technologies have the following shortcomings: Conventional normalization methods apply uniform linear scaling to all features, ignoring the essential differences in distribution patterns and physiological significance between power features and coherence features. This leads to decreased feature discrimination after normalization and fails to incorporate prior knowledge of brain networks. Conventional feature selection techniques primarily assess linear relationships or single discriminative power between features, making it difficult to effectively quantify the complex nonlinear dependencies between features and class labels. They are prone to overlooking key nonlinear discriminative features and lack the ability to dynamically weight feature importance. Conventional classification models typically concatenate or process brain features with clinical information independently, failing to design effective mechanisms to adaptively modulate the expression of brain functional features by clinical factors. They also cannot achieve adaptive fusion of local brain activity and network connectivity information, resulting in insufficient modeling ability for individual differences and hierarchical brain network structures. Most models rely solely on linear or shallow nonlinear classifiers, lacking mechanisms for explicitly modeling the complex interactions between abnormal patterns in local brain regions and similarity patterns in global brain networks. The decision-making process is often a "black box," making it difficult to explain from a neurophysiological perspective why the model makes specific judgments, resulting in poor interpretability. Summary of the Invention

[0003] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a method for early identification of developmental language disorders that integrates neural activity features, comprising the following steps: S1. Obtain raw neurophysiological signals and associated clinical information, annotate them, and construct a dataset; S2. A normalization strategy is adopted for the power and coherence features in the original neurophysiological signals to retain the inherent physiological correlation and improve the feature discrimination, resulting in a normalized feature vector; the normalized feature vector is nonlinearly modulated by simulating functional interactions between brain regions to obtain an enhanced feature vector. S3. Enhanced feature vectors may contain information that contributes little to the classification task or is redundant. Conventional feature selection techniques fail to fully consider the correlation between features and their complex nonlinear dependencies with class labels, which may lead to insufficient discriminative power of the selected feature subset and loss of key nonlinear discriminative information. An attention mechanism based on mutual information is adopted to dynamically evaluate and weight the importance of each feature in the enhanced feature vector to highlight key discriminative features and suppress redundant information, resulting in attention-weighted feature vectors. The attention-weighted feature vectors are nonlinearly mapped to a high-dimensional space through radial basis functions to enhance the nonlinear separability of features at the classification boundary, resulting in radial basis function extended feature vectors. S4. Construct an early identification model for developmental language disorders. The model captures the complex relationship between brain function abnormalities and language disorders at multiple levels through brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, and decision functions that incorporate nonlinear interaction terms. The binary cross-entropy loss function is used as the optimization objective for model training. The early identification model for developmental language disorders is trained using the training set in the dataset to obtain a well-trained early identification model for developmental language disorders. S5. After processing the raw neurophysiological signals of the subjects to be evaluated, the signals are input into the trained early identification model for developmental language disorders to obtain the identification results.

[0004] Furthermore, the original neurophysiological signals are electroencephalogram (EEG) or magnetoencephalogram (MEG) signals; the accompanying clinical information includes age, sex, and family history of language disorders; the annotation type includes typical developmental and progressive language disorders.

[0005] Furthermore, for the power characteristics, a logarithmic compression normalization method is used to obtain the normalized values ​​of the power characteristics; for the coherence characteristics, a distribution transformation based on the error function is used to obtain the normalized values ​​of the coherence characteristics; the normalized values ​​of the power characteristics and the normalized values ​​of the coherence characteristics are concatenated in their original order to obtain the normalized feature vector.

[0006] Furthermore, a brain network modulation matrix reflecting prior knowledge of brain inter-regional anatomy and functional connectivity is constructed. This matrix is ​​then used to nonlinearly modulate the normalized feature vectors to enhance the interaction information between features related to the brain network topology. Define the brain network modulation matrix as , dimension It is a fixed matrix defined based on prior knowledge, used to encode the expected interaction strength between different features. The element in the i-th row and j-th column Indicates the first The first feature and the second The prior connection strength between features; by introducing a nonlinear interaction term guided by the prior matrix of the brain network into the normalized feature vector, an enhanced feature vector is obtained.

[0007] Furthermore, the kernel density estimation method is used to estimate the joint probability distribution of each feature and the binary classification label and their respective marginal distributions from the training data. The mutual information between each feature and the class label is calculated to quantify the degree of nonlinear statistical dependence between them. The mutual information value is normalized by the Softmax function to obtain the attention weights. The attention weight vector is multiplied element-wise with the enhanced feature vector to highlight important features and suppress secondary or redundant features, resulting in the attention-weighted feature vector. The radial basis function is introduced to map the attention-weighted feature vector to a higher-dimensional nonlinear space to obtain the radial basis function extended feature vector.

[0008] Furthermore, brain region-level feature aggregation and modulation: features reflecting the comprehensive activity level of each brain region and the intensity of its network interaction are extracted and fused from the attention-weighted feature vector. Learnable balancing parameters are introduced to achieve adaptive aggregation, thereby enhancing the model's ability to focus on key brain regions and obtaining aggregated feature values ​​of brain regions. The aggregated feature values ​​of all brain regions are combined in sequence to obtain brain region feature vectors.

[0009] Furthermore, to consider the modulating effect of clinical factors such as age, gender, and family history on brain function, a gating modulation mechanism is adopted to adaptively fuse the clinical factor vector with the brain region feature vector, thereby enhancing the model's ability to model individual differences and obtaining a modulated fusion feature vector.

[0010] Furthermore, to capture the complex interactions between brain region features, radial basis function extension features, and external factors, a nonlinear hybrid expert decision function is constructed based on the modulation fusion feature vector and the radial basis function extension feature vector. This function combines linear terms and second-order interaction terms to enhance the expressive power of the classification boundary and ultimately output the predicted probability.

[0011] Furthermore, by performing a Sigmoid mapping on the classification scores of the fusion function integration index, the predicted probability of developmental language impairment is obtained.

[0012] This invention also provides an early identification system for developmental language disorders that integrates neural activity features, and the method for early identification of developmental language disorders that integrates neural activity features includes: Data acquisition module: used to acquire raw neurophysiological signals and associated clinical information, and to annotate them to build a dataset; Data normalization module: Used to normalize the power and coherence features in the original neurophysiological signals, retaining the inherent physiological correlation and improving feature discrimination, to obtain a normalized feature vector; by simulating functional interactions between brain regions, the normalized feature vector is nonlinearly modulated to obtain an enhanced feature vector; Feature selection and attention weighting module: Enhanced feature vectors may contain information that contributes little to the classification task or is redundant. Conventional feature selection techniques fail to fully consider the correlation between features and their complex nonlinear dependencies with class labels, which may lead to insufficient discriminative power of the selected feature subset and loss of key nonlinear discriminative information. A mutual information-based attention mechanism is adopted to dynamically evaluate and weight the importance of each feature in the enhanced feature vector to highlight key discriminative features and suppress redundant information, resulting in attention-weighted feature vectors. The attention-weighted feature vectors are nonlinearly mapped to a high-dimensional space through radial basis functions to enhance the nonlinear separability of features at the classification boundary, resulting in radial basis function extended feature vectors. Model building module: Used to build an early identification model for developmental language disorders. The model captures the complex relationship between brain function abnormalities and language disorders at multiple levels through brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, and decision functions that incorporate nonlinear interaction terms. The binary cross-entropy loss function is used as the optimization objective for model training. The early identification model for developmental language disorders is trained using the training set in the dataset to obtain a well-trained early identification model for developmental language disorders. Early identification module for developmental language disorders: This module processes the raw neurophysiological signals of the subjects to be evaluated and inputs them into the trained early identification model for developmental language disorders to obtain the identification results.

[0013] The advantages of this invention are: This invention addresses the different statistical characteristics of power and coherence features by employing differentiated normalization strategies based on logarithmic compression and error function transformation, respectively. It also incorporates prior brain network knowledge to construct nonlinear enhancement features, resolving the issue of conventional linear normalization neglecting physiological differences and leading to loss of discriminative information. Furthermore, it utilizes a mutual information-based attention mechanism, leveraging kernel density estimation to quantify the nonlinear statistical dependence between features and classification labels, dynamically generating feature importance weights. This overcomes the limitations of conventional filtering or encapsulation-based feature selection methods, which struggle to capture complex nonlinear relationships and may lose key discriminative patterns. A composite model architecture integrating brain region-level feature aggregation, clinical factor gating modulation, and nonlinear hybrid expert decision-making is employed. Through learnable equilibrium parameters, it adaptively fuses local activity and network connectivity information, and utilizes gating mechanisms to dynamically modulate brain features with clinical information, achieving collaborative modeling of brain functional abnormalities and individual differences. Finally, a second-order interaction term between brain region features and radial basis function extension features is introduced into the decision function, and a functional integration index is used as a decision calibration term. This enables the model to capture the synergistic effects of cross-level features and adjust classification boundaries based on the whole-brain network integration state, enhancing the model's nonlinear expressive power and neurophysiological interpretability. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0015] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart of the steps of the data normalization module in the method of the present invention; Figure 3 This is a flowchart of the feature selection and attention weighting module of the method of the present invention; Figure 4 This is a flowchart illustrating the steps of brain region-level feature aggregation and modulation in the method of the present invention. Figure 5 This is a diagram illustrating the importance of different brain region features in the method of this invention. Figure 6 This study analyzes the importance of different brain region connectivity features in the method of this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0017] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a method for early identification of developmental language disorders that integrates neural activity features, the specific steps of which include: S1, Data Acquisition Module The data acquisition module is responsible for acquiring raw neurophysiological signals and accompanying clinical information for constructing and training early identification models of developmental language disorders. This module uses non-invasive neuroimaging technology to acquire electroencephalogram (EEG) or magnetoencephalography (MEG) signals from subjects in a resting state. The acquisition process must be conducted in a soundproof, electromagnetically shielded laboratory, requiring subjects to remain awake, with their eyes closed, relaxed, and without systematic thought activity.

[0018] The signal acquisition equipment uses a multi-channel EEG cap or magnetoencephalogram (MEG) sensor array that conforms to international standards to ensure coverage of key brain regions, including the prefrontal cortex, temporal lobe (especially language-related brain regions such as Broca's area and Wernicke's area), parietal lobe, and occipital lobe.

[0019] The acquisition parameters are set as follows: the sampling frequency is no less than 500Hz to capture sufficient details of neural oscillations, and an online bandpass filter (e.g., 0.5Hz-70Hz) is applied to suppress power line interference and physiological artifacts (e.g., eye movements, ECG). Each acquisition and recording session typically lasts 5 to 10 minutes to ensure stable resting-state neural activity data.

[0020] In addition, clinical background information for each participant needs to be collected simultaneously, including age, gender, and family history of language disorders.

[0021] The data used needs to be labeled, specifically by manual annotation based on internationally recognized standardized language assessment tools (such as clinical assessment, language scales, diagnostic interviews, etc.) and diagnostic criteria (such as DSM-5). The annotation content is the category label corresponding to each sample, and the annotation method is binary classification, that is, determining whether the subject belongs to "typical development" or "developmental language disorder".

[0022] The training and testing datasets consist of rigorously labeled brain signal data and corresponding clinical information.

[0023] S2, Data Normalization Module The fusion of neural activity feature data includes power features and coherence features. These two types of features have different physical meanings and numerical scales. Conventional normalization methods apply the same linear transformation to all features, ignoring the differences in physiological meaning between features. This may lead to a decrease in the discriminative power of the features after normalization. Furthermore, a single linear normalization cannot model the complex nonlinear interactions between features of different brain regions.

[0024] This invention employs targeted normalization strategies based on the different statistical characteristics and physiological significance of power and coherence features to preserve their inherent physiological relevance and improve feature discriminative power. It introduces a feature enhancement function based on prior knowledge of brain network connectivity, constructing nonlinear enhanced feature vectors by simulating functional interactions between brain regions, such as... Figure 2 As shown, the specific steps are as follows: 1) Differential normalization of multi-source heterogeneous features To address the skewed distribution of power characteristics and the Gaussian distribution of coherence characteristics, logarithmic compression normalization and distribution transformation normalization based on the error function are applied respectively, outputting normalized feature vectors. ; Specifically, the original feature vector is defined as follows: The total feature dimension The first 10 features are power features, and the subsequent 45 features are coherence features; For the power characteristics, a logarithmic compression normalization method is used, which is expressed as follows: ; For coherence characteristics, a normalization method based on the distribution transformation of the error function is adopted, which is expressed as follows: ; in, Indicates the first The original values ​​of each power characteristic; Indicates the first The original values ​​of each coherent feature; Indicates the first Normalized values ​​of each power characteristic, with a numerical range of... The interval, after logarithmic compression, is closer to a uniform distribution; Indicates the first The normalized values ​​of each coherent feature, with a numerical range of [value range missing]. The interval, the transformed value can be understood as the position of the original value under the Gaussian cumulative distribution function; The index representing the power characteristic, with a value range of 100. ; The index representing the coherence feature has a value range of 1. ; The natural logarithm function is used to map power values, which are typically skewed, to a more uniform distribution to reduce data skewness. This represents the minimum significant value of the power characteristic. It is a constant used to prevent division by zero when the original value of the power characteristic is zero, and also adjusts the scaling rate of the logarithmic function. An example value is shown below. ; Represents the first of all samples A set consisting of power characteristic values; This indicates the operation of retrieving the maximum value. Indicates the first training sample among all training samples The maximum value of each power characteristic; This represents the error function, used to map standardized eigenvalues ​​to the form of a Gaussian cumulative distribution function. Indicates the first The mean of each coherent feature on the training dataset is used to characterize the central location of the feature distribution; Indicates the first The standard deviation of a coherent feature on the training dataset is used to characterize the dispersion of the feature distribution; This represents a dimensional set of features, formed by concatenating the normalized values ​​of all features in their original order. The normalized eigenvectors; Represents the normalized eigenvector The first in The last (i.e., the last) feature value.

[0025] It should be noted that power characteristics refer to the estimated power spectral density of EEG or MEG signals in a specific frequency band (such as theta waves, alpha waves, beta waves, etc.), reflecting the intensity or energy level of oscillatory activity of neuronal groups in a specific brain region in the corresponding frequency band. Its value is generally non-negative and often follows a skewed distribution. Coherence characteristics refer to the coherence of signals from two different brain regions in a specific frequency band. It is a statistic calculated in the frequency domain with values ​​between 0 and 1, used to measure the linear dependence or functional connectivity strength of signals from two brain regions at a specific frequency, reflecting the synchronicity or coordination of neural activities between different brain regions.

[0026] It should also be noted that the error function, also known as the probability integral transform or normalization based on the cumulative distribution function, is a part of the standard normal distribution integral. The term essentially calculates the standard normal variable being less than The probability is twice the probability minus 1, therefore... The result of the operation on the term is exactly equivalent to the mean being The standard deviation is The cumulative distribution function value of a Gaussian distribution is a common mathematical method for mapping data that follows (approximately) a Gaussian distribution to the interval [0,1].

[0027] It should also be noted that, The preferred value is In electroencephalogram (EEG) analysis, power spectral density is often expressed as... The unit is used, but the total power or average power of a specific frequency band is often used. It means that it will Set as The magnitude can prevent power values ​​that may be close to zero in actual measurements from causing numerical instability (such as tending towards negative infinity) in logarithmic operations. At the same time, it can act as a scaling factor to adjust the curvature of the logarithmic function, making the normalized data distribution better.

[0028] It should also be noted that the normalization method based on the error function works well when the eigenvalue distribution approximates Gaussian. If the actual data distribution deviates significantly from the Gaussian assumption, a normalization method based on the empirical cumulative distribution function can be considered as an alternative.

[0029] 2) Feature enhancement based on brain network priors A brain network modulation matrix reflecting prior knowledge of brain inter-regional anatomy and functional connectivity is constructed. This matrix is ​​then used to nonlinearly modulate the normalized feature vectors to enhance the interaction information between features related to the brain network topology. Define the brain network modulation matrix as , dimension It is a fixed matrix defined based on prior knowledge, used to encode the expected interaction strength between different features. The element in the i-th row and j-th column Indicates the first The first feature and the second The strength of prior connections between features; In practical implementation, the brain network modulation matrix Based on the definition of brain region anatomical distance and functional correlation, assuming that the first 10 power features correspond to 10 different brain regions in sequence, for power features indexed 1 to 10, if the 10th power feature corresponds to the 10th power feature, then the power feature indexed 1 to 10 corresponds to the power feature indexed 1 to 10. The first feature and the first If each feature originates from a different brain region, then it is defined as follows: ,otherwise For the coherence features of indices 11 to 55, if the first... The first feature and the first If each feature corresponds to the same brain region pair, then ,otherwise .

[0030] in, Indicates the first The first feature and the first Anatomical distances (e.g., in millimeters) between brain regions corresponding to each feature. Indicates the first feature index; This indicates the second feature index.

[0031] Furthermore, while preserving the original normalized features, a nonlinear interaction term guided by the brain network prior matrix is ​​introduced to calculate the enhanced feature vector, expressed as: , In the formula, This represents an enhanced feature vector, with the dimension still being [dimensionality missing]. Interaction terms are introduced to simulate the nonlinear interactions that may exist between features of different brain regions, constrained by anatomical and functional connectivity, thereby enhancing the information in the feature vectors related to the brain network topology; This represents the modulation intensity coefficient, used to control the amplitude of brain network modulation; the preferred value is 0.3. This represents the hyperbolic tangent activation function, which performs a nonlinear transformation on the input features.

[0032] S3, Feature Selection and Attention Weighting Module Enhanced feature vectors may contain information that contributes little to the classification task or is redundant. Conventional feature selection techniques fail to fully consider the correlation between features and their complex nonlinear dependencies with class labels, which may result in insufficient discriminative power of the selected feature subset and loss of key nonlinear discriminative information.

[0033] This invention employs a mutual information-based attention mechanism to dynamically evaluate and weight the importance of each feature, highlighting key discriminative features and suppressing redundant information. Then, a radial basis function is used to nonlinearly map the weighted features to a high-dimensional space, enhancing the nonlinear separability of features at the classification boundary. Figure 3 As shown, the specific steps are as follows: 1) Attention weight calculation based on mutual information The kernel density estimation method is used to estimate the joint probability distribution and marginal distribution of each feature and binary label from the training data. Then, the mutual information between each feature and the class label is calculated to quantify the nonlinear statistical dependence between them. The mutual information values ​​are normalized using the Softmax function to obtain the attention weights, expressed as follows: , In the formula, Indicates the first Attention weights for each feature, with values ​​ranging from Between these values, and the sum of all weights is 1, this weight is used to weight the features in importance in subsequent steps; Represents the natural exponential function; This represents the mutual information scaling parameter, which controls the degree of concentration of attention weight distribution. The larger the value, the more concentrated the weights are on features with high mutual information values. The preferred value is 1. Indicates the first The mutual information value between a feature and the class label is used to quantify the amount of information about the class label carried by that feature. A larger value indicates a stronger non-linear dependency between the feature and the class label. The calculation method is expressed as follows: In practical implementation, for continuous variables It needs to be approximated by discretization (if the features have been discretized) or numerical integration (such as the trapezoidal rule, Monte Carlo integration); In this binary classification task, the category label represents the category label. These represent typical developmental and developmental language disorders, respectively. Represents enhanced feature vectors The Middle Specific observation values ​​of each feature; Representation of features With tags The joint probability density function is obtained from the training data through kernel density estimation (using a Gaussian kernel function); Representation of features The marginal probability density function is obtained by estimating the kernel density from the training data using only features. Estimate the observed values; Indicates label The marginal probability density function for discrete class labels , It is its empirical probability (or prior probability), which can be obtained by directly calculating the frequency of each category of samples from the training data; This represents a logarithmic function, with the default base being the natural constant. Indicates the feature Integrating over the domain of the variable is used to calculate the expectation of a continuous random variable, which is approximated by discretization in practice.

[0034] 2) Feature Importance Weighting The attention weight vector is multiplied element-wise with the enhanced feature vector to highlight important features and suppress secondary or redundant features, as shown below: , In the formula, This represents the attention-weighted feature vector, with dimension . It is a feature representation that has been recalibrated through an attention mechanism, in which features that are strongly correlated with language disorder category discrimination are amplified, while features that are weakly correlated are weakened, thereby improving the discriminative power of the feature set; This represents the attention weight vector, which consists of the attention weights for all features. ; This represents the element-wise multiplication operator.

[0035] 3) Nonlinear characteristic extension based on radial basis functions To improve the nonlinear separability of features, a radial basis function is introduced to map the attention-weighted feature vectors to a higher-dimensional nonlinear space; Specifically, the K-means clustering algorithm is used to determine a set of cluster centers on the attention-weighted feature vectors of the training set. Then, based on each cluster center, the similarity between the attention-weighted feature vectors and each center is calculated, and a set of radial basis functions is constructed for output. In the specific implementation, the number of radial basis functions is set to be... The attention-weighted feature vectors of all samples in the training set are clustered using the K-means algorithm to obtain... Cluster centers ,in Indicates the first The cluster centers were learned on the training set using the K-means algorithm. This represents the cluster center index, with a value range of [value range missing]. ; For attention-weighted feature vectors , its first The output of each radial basis function is expressed as: , In the formula, Indicates the first The output values ​​of the radial basis functions represent the attention-weighted feature vectors. With the Cluster centers The similarity between them, the output value is in Within the interval, the closer to the center, the larger the output value; This represents the L2 norm.

[0036] Indicates the first The width parameter of the radial basis function determines the smoothness of the function's response; specifically, it is set to the width of the first radial basis function. All samples in each cluster are directed to their centers. The median of the Euclidean distance.

[0037] 4) Construction of high-dimensional feature vectors By concatenating the output values ​​of all radial basis functions, we construct the extended eigenvector of the radial basis functions, which is represented as: , In the formula, This represents the radial basis function extended eigenvector, with dimension . This maps the original features to a non-linear, high-dimensional space based on a similarity metric, where each dimension represents the similarity between a sample and a typical pattern learned from the data.

[0038] S4. Construct an early identification model for developmental language disorders. Conventional classification models fail to fully utilize prior knowledge of brain network topology and do not effectively integrate the complex interactions between clinical factors and high-dimensional features, which may lead to insufficient generalization ability and lack of interpretability.

[0039] This invention constructs an early identification model for developmental language disorders. This model, through brain region-level feature aggregation, gating modulation of clinical factors, and a decision function incorporating nonlinear interaction terms, captures the complex correlation between brain functional abnormalities and language disorders at multiple levels. Figure 4 As shown, the specific steps are as follows: 1) Brain region-level feature aggregation and modulation Features reflecting the overall activity level and network interaction intensity of each brain region are extracted and fused from the attention-weighted feature vector. A learnable balancing parameter is introduced to achieve adaptive aggregation, thereby enhancing the model's ability to focus on key brain regions. This is represented as: , In the formula, Indicates the first The aggregated feature values ​​of each brain region range from 1 to 10. , characterizing the The comprehensive neural activity intensity index of each brain region is fused with learnable weights to reflect the intensity of the brain region's own oscillatory activity and the intensity of the functional connectivity between the brain region and other key brain regions, thereby obtaining brain region-level features that include both local activity and network interaction. The hyperbolic tangent activation function is used to introduce nonlinearity, enabling the model to capture the complex nonlinear relationship between power and coherence features during aggregation. The balance parameter representing the power feature and the coherence feature is a trainable scalar parameter used to adaptively adjust the contribution weights of the two types of features in the aggregation. Represents attention-weighted feature vectors The middle corresponds to the first Power characteristic values ​​of individual brain regions Indicates the first Power characteristics of individual brain regions Indices in a vector Used to locate this feature, for the first There are brain regions, ; Indicates the relationship with the first A brain region is a set of brain regions with anatomical or functional connections. This set is predefined based on prior knowledge of brain networks. For example, brain regions that are anatomically adjacent or highly correlated on resting-state functional networks can be defined using a standard brain network atlas (such as the Desikan-Killiany template). Members; Represents attention-weighted feature vectors The middle corresponds to the brain region pair coherence eigenvalues, Indicates by the first The brain regions and the first Connectivity pairs composed of brain regions, index The position of a brain region in a vector is determined by a predefined mapping table of brain regions to coherent feature numbers. Represents a set The cardinality, that is, the number of the th... The total number of brain regions connected to each other; Indicates a connection index for brain regions; This represents the total number of brain regions, with examples of possible values. Corresponding to the division from EEG or MEG signals A key brain region.

[0040] Furthermore, by sequentially constructing the aggregated feature values ​​of all brain regions, a brain region feature vector is obtained. ,Right now , dimension It integrates local activity and network connectivity information of each brain region.

[0041] In one implementation, the definition is... Relying on prior knowledge of brain networks, such as anatomical connectivity probability maps (e.g., white matter fiber connectivity probabilities based on diffusion tensor imaging) or population-level resting-state functional connectivity templates, a connectivity probability threshold can be set for a specific brain region (e.g., the left inferior frontal gyrus, IFG), incorporating all other brain regions with connectivity probabilities to the IFG higher than that threshold. This allows them to leverage group prior knowledge when individualized connection data is lacking.

[0042] It should be noted that in the calculation Attention-weighted feature vectors It is dimensional vector, Directly specify the location, represented as That is, taking the attention-weighted feature vector. The first vector elements ( (from 1 to 10); and, in calculation At that time, a brain region is needed to... To coherence feature index The lookup table, assuming it has Each brain region has a maximum of These 45 unique pairs of brain regions correspond to exactly 45 coherent features, requiring a predefined mapping: ,in For example, regulations , ,..., .but .

[0043] 2) Clinical factor fusion and gating modulation To account for the modulating effects of clinical factors such as age, sex, and family history on brain function, a gating modulation mechanism is employed to adaptively fuse clinical factor vectors with brain region feature vectors, thereby enhancing the model's ability to model individual differences. This is expressed as: , In the formula, This represents the modulation fusion feature vector, with dimension . It simultaneously encodes brain region feature information and clinical factor information, and dynamically adjusts the importance of each dimension through a gating mechanism; The clinical factor vector is defined as follows: , dimension ,Include A standardized clinical factor, including, for example: age (standardized to Z-score), sex (unique heat coding), family history (binarized). This indicates the number of clinical factors, with examples of possible values. ; Represents the clinical factor vector The first in Each factor value, The range of values ​​is ; This represents the weight matrix used to transform brain region features, with dimension 1. , are trainable parameters; This represents the weight matrix used to transform the clinical factor vector, with dimensions of [missing information]. , are trainable parameters; This represents the weight matrix used to generate the gated signal, which transforms brain region features; its dimension is [missing information]. , are trainable parameters; This represents the weight matrix used to generate the transformed clinical factor vector for gating signals, with dimensions of [missing information]. , are trainable parameters; The bias vector representing the fusion of basic values ​​has dimension 1. , are trainable parameters; The bias vector representing the generation of the gated signal has dimension . , are trainable parameters; This represents the Sigmoid activation function, whose output range is... between; It should be noted that, The basic values ​​for generating fusion features are: The gating weights are generated to adaptively determine the degree of preservation of each dimension in the fusion feature based on the current brain region features and external factors. The two are combined through element-wise multiplication. The gating weights dynamically control the "passage" of each dimension in the basic fusion feature. For example, for individuals of certain ages and genders, the gating mechanism may selectively enhance the brain region feature dimensions related to language processing while suppressing other irrelevant dimensions. This allows the model to adaptively and selectively focus on the brain function abnormality patterns most relevant to the current individual based on external clinical information, greatly enhancing the model's personalized modeling ability and interpretability.

[0044] 3) Nonlinear hybrid expert decision function To capture the complex interactions between brain region features, radial basis function extension features, and external factors, a nonlinear hybrid expert decision function is constructed. This function combines linear terms and second-order interaction terms to enhance the expressive power of the classification boundary and ultimately output the predicted probability. The specific calculation is as follows: , In the formula, Represents the classification score, with a value range of 100%. , used to map to classification probabilities through the Sigmoid function; This represents a weight vector associated with the modulation-fusion feature vector, with dimension 1. , are trainable parameters; express transpose; This represents the weight vector associated with the radial basis extended eigenvectors, with dimension 1. , are trainable parameters; express transpose; express transpose; This represents the weight matrix characterizing the second-order interaction, with dimension 1. , are trainable parameters; The bias term of the decision function is a trainable parameter.

[0045] It should be noted that, This approach can capture the coupling effect between brain region activity patterns and global similarity patterns, and model the pairwise interaction between modulated brain region features and global similarity features of radial basis expansion. This is because in developmental language disorders, abnormalities may manifest as dysfunction in specific brain regions (corresponding to modulated fusion feature vectors representing local features). This also manifests as a deviation from the overall pattern of the whole-brain functional network (corresponding to the radial basis function extended feature vector representing global features). Furthermore, these two abnormal patterns may influence each other and jointly determine the disorder phenotype. This is to capture the synergistic or antagonistic effects of such cross-level features, thereby enhancing the nonlinear expressive power of the classification decision boundary and enabling the model to learn more complex classification boundaries that are more consistent with real neurophysiological mechanisms.

[0046] It should also be noted that, The method linearly weights brain region-level features modulated with clinical information to capture core discriminative patterns based on brain regions and clinical factors. The term applies a linear weight to global similarity pattern features (i.e., how close a sample is to multiple typical patterns).

[0047] 4) Decision calibration based on functional integration index By applying a Sigmoid mapping to the classification scores that incorporate the functional integration index, the predicted probability of developmental language impairment is obtained, expressed as: , In the formula, This represents the predicted probability of developmental language disorder, with a value range of [value range missing]. ,when At that time, the sample was determined to belong to the developmental language disorder category, and the corresponding label was... ; This represents the calibration intensity coefficient, used to control the degree of influence of the functional integration index on the final decision, and is a trainable parameter. The functional integration index is a scalar used to quantify the balance between integration and disintegration in the brain's functional networks. It is calculated as follows: ; This represents the weight vector used to weight latent features of brain regions, with dimension 1. ; express transpose; This represents the weight vector used for expanding the features of the weighted radial basis function, with dimension 1. ; Denotes a trainable projection matrix with dimension . Used to extract brain region feature vectors Extract One potential feature related to functional integration; The dimension representing the potential functional integration features extracted from brain region features is a hyperparameter, with an optimal value of 5. express transpose; This represents the modified linear unit activation function, used to introduce nonlinearity while ensuring nonnegativity.

[0048] It should be noted that in the neurophysiological hypothesis of developmental language disorders, abnormal integration of brain functional networks (i.e., dysfunction of functional connectivity between different brain regions) is one of the core pathological mechanisms. By fusing the local activity patterns encoded by brain region feature vectors with the global similarity patterns encoded by radial basis function extension feature vectors, a comprehensive index reflecting the degree of integration of individual brain functions is constructed and used as a calibration term for the decision function. This enables the model to adaptively adjust the classification boundary based on the overall functional state of the neural network, thereby enhancing the neurophysiological interpretability of the decision-making process.

[0049] 5) Define the loss function The binary cross-entropy loss function is used as the optimization objective for model training. It is used to quantify the difference between the model's predicted probability of developmental language disorder and the actual clinical diagnostic label. Its function is to measure the degree of prediction error of the model under the current parameters and to provide a clear gradient direction for updating the model parameters.

[0050] By minimizing the binary cross-entropy loss, the model is driven to adjust its internal trainable parameters so that the predicted probability of samples in the training data that belong to developmental language disorders is as close to 1 as possible, while the predicted probability of samples that belong to typical developmental language disorders is as close to 0 as possible.

[0051] It should be noted that the binary cross-entropy loss function naturally matches the Sigmoid output layer, effectively handling the probability output of binary classification tasks, and is a standard and effective method for training such classification models.

[0052] 6) Define the model training method The model is trained using a gradient descent-based optimization algorithm, specifically the mini-batch stochastic gradient descent method. The binary cross-entropy loss function is used as the optimization objective, and the gradient of the loss function with respect to all trainable parameters in the model is calculated using the backpropagation algorithm.

[0053] In each iteration, a small batch of samples is randomly drawn from the training dataset, the average loss of that batch is calculated, and the model parameters are updated accordingly, thereby gradually reducing the overall loss of the model on the entire training set.

[0054] The training process requires the creation of a separate validation dataset to monitor the model's performance on unseen data. Early stopping is used to determine when to stop iterations: the model's performance (e.g., validation loss or accuracy) is continuously evaluated on the validation set. When the validation loss stops decreasing and begins to increase after a predetermined number of iterations, it is considered that the model may be overfitting. At this point, the training process is terminated, and the model parameters corresponding to the iteration with the best performance on the validation set are selected as the final model.

[0055] S5. Early identification of developmental language disorders Once the early identification model for developmental language disorders is trained, it can be deployed for the early identification of developmental language disorders.

[0056] For new subjects to be evaluated, resting-state EEG or MEG signals, along with clinical information such as age, gender, and family history, are first acquired through a data acquisition module. Then, a 55-dimensional feature vector, identical to that used in the training phase, is extracted from the raw signals. This vector contains power features from 10 brain regions and coherence features from 45 brain regions, and a clinical factor vector is constructed. Next, the feature vector of this new sample is sequentially input into a pre-trained data normalization module and a feature selection and attention weighting module for processing, yielding enhanced feature vectors, attention-weighted feature vectors, and radial basis function extended feature vectors, respectively. Finally, these processed feature vectors, along with the clinical factor vector, are input into a pre-trained early identification model for developmental language disorders. The model performs brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, calculates a nonlinear hybrid expert decision function score, and performs decision calibration using a functional integration index. Finally, a predicted probability value between 0 and 1 is output through a sigmoid function. This probability value represents the risk level of the subject having a developmental language disorder.

[0057] Based on a pre-set decision threshold (e.g., 0.5), if the predicted probability is greater than the threshold, the subject is determined to be a high-risk individual for developmental language disorder, and further in-depth clinical evaluation and intervention are recommended; if the predicted probability is less than or equal to the threshold, the subject is determined to be highly likely to have typical developmental language disorder.

[0058] This completes the early and objective identification of the risk of developmental language disorders in individuals.

[0059] In one embodiment, an importance analysis of brain region features for early identification of developmental language disorders is performed. This analysis examines the importance distribution of different brain regions and their connectivity features in the early identification task of developmental language disorders, validating the effective focus of the method on language-related brain regions and providing a basis for the neurophysiological interpretability of the model. The experiment calculates the feature importance scores of 10 key brain regions and the connectivity feature importance of 8 important brain region connectivity pairs. Brain region importance is calculated by combining the contributions of attention weights and brain region-level feature aggregation modules in the model, while connectivity feature importance is assessed based on the contribution of coherence features in the final decision.

[0060] like Figure 5 As shown in the subplot "Importance Analysis of Different Brain Region Features," the feature importance scores of 10 brain regions are displayed, ranked from lowest to highest importance. The horizontal axis represents the feature importance score, and the vertical axis represents the names of the 10 brain regions. The left temporal lobe and Broca's area have the longest bars, consistent with neuroscience knowledge regarding the localization of brain functions related to language processing. The prefrontal, parietal, and occipital lobes have relatively lower importance but still contribute significantly. The method of this invention can effectively capture the neural activity features most relevant to language disorders.

[0061] like Figure 6 As shown in the subplot "Importance Analysis of Connectivity Features in Different Brain Regions," the importance distribution of connectivity features for eight important brain region connectivity pairs is displayed. The horizontal axis represents the names of the eight connectivity pairs, and the vertical axis represents the importance of the connectivity features. Each connection in the figure corresponds to multiple scatter points, representing fluctuations across multiple experiments. The size of the points indicates the magnitude of the importance value, and the color, from cool to warm, indicates the value's intensity. Among the connectivity features, the left temporal lobe-Broca's area, right temporal lobe-Wernicke's area, and Broca's area-Wernicke's area—three connectivity pairs closely related to language function—have the highest importance. The scatter points are mainly concentrated in the region with higher vertical axes, and the size and color of the points also indicate a concentration of high values. Other connectivity pairs, such as those between the left and right prefrontal lobes and between the parietal and occipital lobes, have relatively lower importance. The red solid lines in the figure connect the average values ​​of each connectivity pair, showing the ranking of importance. Experimental results show that the method of this invention can effectively identify the indicative role of abnormal functional connectivity between brain regions in developmental language disorders, especially the dysregulation of connectivity between language-related brain regions.

[0062] Example 2 This embodiment provides an early identification system for developmental language disorders that integrates neural activity features, and executes the early identification method for developmental language disorders that integrates neural activity features described in Embodiment 1, including: Data acquisition module: used to acquire raw neurophysiological signals and associated clinical information, and to annotate them to build a dataset; Data normalization module: Used to normalize the power and coherence features in the original neurophysiological signals to obtain normalized feature vectors; and to obtain enhanced feature vectors by nonlinearly modulating the normalized feature vectors through simulated brain inter-region functional interactions. Feature selection and attention weighting module: Employs a mutual information-based attention mechanism to dynamically evaluate and weight the importance of each feature in the enhanced feature vector, resulting in an attention-weighted feature vector; The attention-weighted feature vector is nonlinearly mapped to a high-dimensional space using radial basis functions, resulting in a radial basis function-extended feature vector; Model building module: Used to build an early identification model for developmental language disorders. The model captures the complex relationship between brain function abnormalities and language disorders at multiple levels through brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, and decision functions that incorporate nonlinear interaction terms. The binary cross-entropy loss function is used as the optimization objective for model training. The early identification model for developmental language disorders is trained using the training set in the dataset to obtain a well-trained early identification model for developmental language disorders. Early identification module for developmental language disorders: This module processes the raw neurophysiological signals of the subjects to be evaluated and inputs them into the trained early identification model for developmental language disorders to obtain the identification results.

[0063] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early identification of developmental language disorders that integrates neural activity features, characterized in that, Includes the following steps: S1. Obtain raw neurophysiological signals and associated clinical information, annotate them, and construct a dataset; S2. The power and coherence features in the original neurophysiological signals are normalized to obtain normalized feature vectors; the normalized feature vectors are nonlinearly modulated by simulating functional interactions between brain regions to obtain enhanced feature vectors. S3. Employ a mutual information-based attention mechanism to dynamically evaluate and weight the importance of each feature in the enhanced feature vector, resulting in an attention-weighted feature vector. Then, use radial basis functions to nonlinearly map the attention-weighted feature vector to a high-dimensional space, resulting in a radial basis function-extended feature vector. S4. Construct an early identification model for developmental language disorders. The model captures the complex relationship between brain function abnormalities and language disorders through brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, and a decision function that incorporates nonlinear interaction terms. The binary cross-entropy loss function is used as the optimization objective for model training. The early identification model for developmental language disorders is trained using the training set in the dataset to obtain a well-trained early identification model for developmental language disorders. S5. After processing the raw neurophysiological signals of the subjects to be evaluated, the signals are input into the trained early identification model for developmental language disorders to obtain the identification results.

2. The method for early identification of developmental language disorders incorporating neural activity features according to claim 1, characterized in that, The original neurophysiological signals are electroencephalogram (EEG) or magnetoencephalogram (MEG) signals; the accompanying clinical information includes age, sex, and family history of language disorders; the annotation type includes typical developmental and progressive language disorders.

3. The method for early identification of developmental language disorders incorporating neural activity features according to claim 2, characterized in that, For power characteristics, a logarithmic compression normalization method is used to obtain normalized values ​​for power characteristics; for coherence characteristics, a distribution transformation based on the error function is used to obtain normalized values ​​for coherence characteristics; the normalized values ​​for power characteristics and coherence characteristics are concatenated in their original order to obtain a normalized feature vector.

4. The method for early identification of developmental language disorders incorporating neural activity features according to claim 3, characterized in that, Construct a brain network modulation matrix and use this matrix to perform nonlinear modulation on the normalized feature vector: Define the brain network modulation matrix as , dimension It is a fixed matrix defined based on prior knowledge, used to encode the expected interaction strength between different features. The element in the i-th row and j-th column Indicates the first The first feature and the second The prior connection strength between features; by introducing a nonlinear interaction term guided by the prior matrix of the brain network into the normalized feature vector, an enhanced feature vector is obtained.

5. The method for early identification of developmental language disorders incorporating neural activity features according to claim 4, characterized in that, The joint probability distribution and marginal distribution of each feature and binary label are estimated from the training data using the kernel density estimation method. The mutual information between each feature and the class label is calculated. The mutual information value is normalized by the Softmax function to obtain the attention weight. The attention weight vector is multiplied element-wise with the enhanced feature vector to obtain the attention weighted feature vector; a radial basis function is introduced to map the attention weighted feature vector to a higher-dimensional nonlinear space to obtain the radial basis function extended feature vector.

6. The method for early identification of developmental language disorders incorporating neural activity features according to claim 5, characterized in that, Brain region-level feature aggregation and modulation: Features reflecting the comprehensive activity level and network interaction intensity of each brain region are extracted and fused from the attention-weighted feature vector. A learnable balancing parameter is introduced to achieve adaptive aggregation, resulting in aggregated feature values ​​of the brain regions. The aggregated feature values ​​of all brain regions are combined in sequence to obtain the brain region feature vector.

7. The method for early identification of developmental language disorders incorporating neural activity features according to claim 6, characterized in that, A gating modulation mechanism is used to adaptively fuse clinical factor vectors and brain region feature vectors to obtain a modulation fusion feature vector.

8. The method for early identification of developmental language disorders incorporating neural activity features according to claim 7, characterized in that, A nonlinear hybrid expert decision function is constructed based on the modulation fusion feature vector and the radial basis function extended feature vector. This function combines linear terms and second-order interaction terms to enhance the expressive power of the classification boundary and finally output the predicted probability.

9. The method for early identification of developmental language disorders incorporating neural activity features according to claim 8, characterized in that, By applying a Sigmoid mapping to the classification scores of the integration index of fusion function, the predicted probability of developmental language impairment is obtained.

10. An early identification system for developmental language disorders that integrates neural activity features, performing the early identification method for developmental language disorders that integrates neural activity features as described in claim 1, characterized in that, include: Data acquisition module: used to acquire raw neurophysiological signals and associated clinical information, and to annotate them to build a dataset; Data normalization module: Used to normalize the power and coherence features in the original neurophysiological signals to obtain normalized feature vectors; and to obtain enhanced feature vectors by nonlinearly modulating the normalized feature vectors through simulated brain inter-region functional interactions. Feature selection and attention weighting module: Employs a mutual information-based attention mechanism to dynamically evaluate and weight the importance of each feature in the enhanced feature vector, resulting in an attention-weighted feature vector; The attention-weighted feature vector is nonlinearly mapped to a high-dimensional space using radial basis functions, resulting in a radial basis function-extended feature vector; Model building module: Used to build an early identification model for developmental language disorders. The model captures the complex relationship between brain function abnormalities and language disorders at multiple levels through brain region-level feature aggregation and modulation, clinical factor fusion and gating modulation, and decision functions that incorporate nonlinear interaction terms. The binary cross-entropy loss function is used as the optimization objective for model training. The early identification model for developmental language disorders is trained using the training set in the dataset to obtain a well-trained early identification model for developmental language disorders. Early identification module for developmental language disorders: This module processes the raw neurophysiological signals of the subjects to be evaluated and inputs them into the trained early identification model for developmental language disorders to obtain the identification results.

Images