Intelligent auxiliary diagnosis system for autism based on eeg microstate and xgboost

Abstract

The application provides an autism intelligent auxiliary diagnosis system based on electroencephalogram microstate and XGBoost, relates to the technical field of auxiliary diagnosis, and comprises a collection module, a processing module, a selection module and an output module. The collection module is used for acquiring electroencephalogram data corresponding to a training sample set and electroencephalogram data of a to-be-tested object; the processing module is used for extracting potential topography in the training sample set through a representative selection mechanism and adaptively determining the number of microstates to generate a microstate template set; the selection module is used for performing time sequence back fitting on the electroencephalogram data of the to-be-tested object with the microstate template set, extracting time learning features and performing sparse feature selection to obtain a target feature index set; and the output module is used for constructing a classifier with a regular constraint, training a classification model based on the training sample set and generating autism risk discrimination information. The application can automatically extract the space-time features of electroencephalogram signals and perform optimized classification, and effectively improves the accuracy and efficiency of autism diagnosis.

Description

Technical Field

[0001] This application relates to the field of auxiliary diagnostic technology, and more specifically, to an intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost. Background Technology

[0002] Autism Spectrum Disorder (ASD) is a common neurodevelopmental disorder. Its clinical diagnosis primarily relies on behavioral observation and scale assessment, which is heavily influenced by subjective factors, and diagnostic delays often lead to delayed intervention. Existing technologies include EEG-based diagnostic systems that attempt to assist ASD identification using machine learning methods, such as functional connectivity analysis or time-series feature extraction combined with classifiers. While these systems have made some progress in capturing abnormal brain activity, they still have limitations: feature extraction often relies on manual intervention or simple statistics, resulting in insufficient automated processing of the spatiotemporal dynamics of EEG signals; furthermore, the classification optimization mechanism is imperfect, susceptible to noise interference and overfitting, making it difficult to meet clinical needs in terms of diagnostic accuracy and efficiency.

[0003] Therefore, how to automatically extract and optimize the spatiotemporal characteristics of EEG signals in an autism diagnostic system to improve the accuracy and efficiency of diagnosis has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides an intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost, including:

[0005] The acquisition module is used to acquire EEG information, which includes: EEG data corresponding to the training sample set and EEG data of the test subject;

[0006] The processing module is used to select representative moments from the EEG data corresponding to the training sample set through a representative selection mechanism, and extract potential topography at the representative moments; cluster the potential topography and determine the number of microstates through stability evaluation, and generate a microstate template set.

[0007] The selection module is used to perform temporal reconfiguration on the EEG data of the test subject using the microstate template set to obtain the temporal feature vector of the test subject; based on the training sample set, a temporal feature matrix is ​​obtained; for the temporal feature matrix, sparse feature selection is performed to determine the target feature index set, and the target feature index set is applied to the temporal feature vector of the test subject to obtain a feature subset of the test subject;

[0008] The output module is used to construct a classifier with regularization constraints, train a classification model using the training sample set, receive a subset of features of the test object to generate a discrimination score, and output auxiliary diagnostic information to characterize the risk of autism based on the discrimination score.

[0009] Optionally, the step of extracting the potential topography at the representative time includes:

[0010] The measurement channels of the EEG data corresponding to the training sample set are divided into a first channel group and a second channel group.

[0011] The first artifact coefficient is calculated for the channels of the first channel group based on the absolute deviation of the instantaneous amplitude relative to the median of the channel. The second artifact coefficient is calculated for the channels of the second channel group based on the degree of inconsistency of the spatial correlation between the channels. The low artifact factor is calculated by combining the first artifact coefficient and the second artifact coefficient.

[0012] Based on the low artifact factor, the global field strength quantile at each time point, and the spatial correlation average of the potential topography within the first window length at each time point, the representative score for each time point is calculated.

[0013] Non-maximum suppression is performed based on the representative scores to determine multiple representative moments, and potential terrain is extracted at the representative moments to form a terrain sample set.

[0014] Optionally, dividing the measurement channels of the EEG data corresponding to the training sample set into a first channel group and a second channel group includes:

[0015] Based on the EEG data corresponding to the training sample set, the artifact sensitivity index of each measurement channel is calculated, and each measurement channel is divided into a first channel group and a second channel group according to the quantile threshold of the artifact sensitivity index.

[0016] The first artifact coefficient and the second artifact coefficient are calculated for the first channel group and the second channel group respectively; the contribution of the first artifact coefficient and the second artifact coefficient in the potential topography clustering stability evaluation index within the corresponding channel group is determined as the weight; and a weighted geometric mean is performed on the first artifact coefficient and the second artifact coefficient according to the weight to generate a low artifact factor.

[0017] Based on the training sample set, the quantile thresholds used for channel segmentation are determined according to the absolute deviation rules of quantiles and medians within the channel group.

[0018] Optionally, clustering the potential topography includes:

[0019] Based on the spatial layout of the measurement channels of the EEG data corresponding to the training sample set, a channel adjacency graph is established. Polarity unification and mirror mapping are performed on the potential terrain. The spatial spectral energy ratio of each potential terrain on the graph Laplacian eigenvector basis of the channel adjacency graph is calculated. The correlation coefficient between each potential terrain and the mirror potential terrain is calculated as the mirror symmetry.

[0020] Within a preset set of candidate microstates, constrained clustering is performed on the potential terrain, with the goal of maximizing the spatial correlation within the group, and the constraint that the spatial spectral energy ratio of each cluster center is not greater than a first upper limit and the mirror symmetry is not less than a first lower limit is imposed.

[0021] Optionally, determining the number of microstates through stability evaluation and generating a microstate template set includes:

[0022] For each candidate microstate, hierarchical resampling is performed on a sample-by-sample basis within the training sample set to form multiple sets of potential terrain. After performing the constrained clustering on each set of potential terrain, the obtained cluster centers are back-projected onto the corresponding potential terrain set. The reprojection error and the matching consistency of the cluster centers are calculated, and a weighted combination is performed with the spatial spectral energy ratio and mirror symmetry of the corresponding cluster centers to determine the stability score.

[0023] Based on the stability score sequence in the training sample set, calculate the stability score gain of the number of adjacent candidate microstates, and define the preset score gain threshold as the second threshold.

[0024] The number of microstates that maximizes the stability score and whose score gain for the number of adjacent candidate microstates is less than a second threshold is selected as the target number of microstates, and a set of microstate templates is generated based on the clustering results corresponding to the target number of microstates.

[0025] Optionally, calculating the spatial spectral energy ratio of each potential topography on the graph Laplacian eigenvector basis of the channel adjacency graph includes:

[0026] Based on the normalized graph Laplacian matrix of the channel adjacency graph, the eigenvalue sequence and eigenvector sequence are obtained, and the first spectral threshold is determined according to the maximum spectral gap in the eigenvalue sequence.

[0027] The ratio of the projected energy of the potential terrain after polarity unification processing on the set of basis vectors with eigenvalues ​​not greater than the first spectral threshold to the projected energy on all basis vectors is determined as the spatial spectral energy ratio.

[0028] Optionally, calculating the correlation coefficient between each potential terrain and its mirror potential terrain as the mirror symmetry includes:

[0029] The midline of the skull is determined based on the spatial layout of the measurement channels of the EEG data corresponding to the training sample set, and the channel correspondence is established based on the mirror image of the coordinates of each measurement channel relative to the midline of the skull to determine the mirror mapping matrix.

[0030] The weights of the measurement channels are determined by normalizing the distances from each measurement channel to the midline of the skull.

[0031] The potential terrain vector after polarity unification processing and the potential terrain vector after the mirror mapping matrix transformation are calculated in the space after removing constant components, and the weighted correlation coefficient is used as the mirror symmetry.

[0032] Optionally, the obtained subset of features of the object to be tested includes:

[0033] The time-matrix feature matrix is ​​divided into multiple feature groups according to microstates. Each feature group includes at least the average duration, time coverage, occurrence frequency, and transition probability vector from the microstate to other microstates, resulting in a grouped time-matrix feature matrix.

[0034] Multiple training sub-matrices are generated on the grouped temporal feature matrix. The training sub-matrices are obtained by performing hierarchical bootstrap sampling on the sample rows. For each training sub-matrice, a regularization weight is calculated based on the coefficient of variation of each feature in the corresponding training sub-matrice, and a candidate feature index set is calculated using weighted group sparse selection.

[0035] Stability statistics are performed on multiple candidate feature index sets, the selection frequency of each feature is calculated and compared with a third threshold to determine the target feature index set;

[0036] The target feature index set is applied to the temporal feature vector of the object under test, and the corresponding components are extracted according to the index order to obtain the feature subset of the object under test.

[0037] Optionally, the step of performing temporal reconfiguration of the EEG data of the test subject using the microstate template set includes:

[0038] Based on the time domain standard deviation of the EEG data corresponding to the training sample set and the EEG data of the test subject in each measurement channel, the channel gain normalization matrix is ​​determined, and the channel gain normalization matrix is ​​used to perform channel gain normalization processing on the EEG data of the test subject.

[0039] On the normalized graph Laplacian eigenvector basis of the channel adjacency graph, orthogonal registration is performed on the potential terrain set of the object under test after the channel gain normalization process and the microstate template set to determine a first alignment matrix, and the microstate template set is transformed with the first alignment matrix to obtain an alignment template set;

[0040] Based on the alignment template set, labels are assigned to each time point of the EEG data of the test subject according to the criterion of maximizing spatial correlation, thereby generating the microstate sequence of the test subject.

[0041] Optionally, constructing a classifier with regularization constraints includes:

[0042] The lower and upper limits of the preset risk threshold range are defined as the fourth and fifth thresholds, respectively;

[0043] On the training sample set, for the risk threshold interval defined by the fourth threshold and the fifth threshold, pre-classification models with different class weight ratios are trained respectively and cross-validation predictions are generated. Based on the cross-validation predictions, the average net benefit of decision curve analysis is calculated within the risk threshold interval to obtain the evaluation value of each class weight ratio.

[0044] The class weight ratio that maximizes the average net benefit is selected as the target weight ratio, and the classifier with regularization constraints is trained using the training sample set and the target feature index set under the target weight ratio to obtain the classification model.

[0045] Compared with existing technologies, this application can automatically complete the spatiotemporal feature extraction and classification model training of EEG signals without manual parameter setting. The system adaptively determines the number of microstates through a stability evaluation mechanism, achieving objective division of EEG microstate templates; it constructs a low-redundancy temporal feature set through sparse feature selection, enhancing the model's generalization ability; and it combines an XGBoost classifier with regularization constraints to effectively suppress overfitting under small sample conditions, improving diagnostic robustness. The final output risk assessment results can provide clinicians with objective auxiliary diagnostic evidence, realizing an automated process from EEG signal acquisition to autism risk assessment. Compared with existing technologies, it has higher spatiotemporal feature expression capabilities, better classification accuracy, and higher diagnostic efficiency. Attached Figure Description

[0046] Figure 1 A flowchart of an intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost provided in an embodiment of this application;

[0047] Figure 2 A flowchart of a method for extracting potential topography provided in this application embodiment;

[0048] Figure 3 A flowchart illustrating a method for dividing a first channel group and a second channel group, as provided in an embodiment of this application;

[0049] Figure 4This is a flowchart of a method for clustering potential terrain provided in an embodiment of this application. Detailed Implementation

[0050] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0051] See Figure 1 The diagram shown is a flowchart of an intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost, provided in an embodiment of this application. The system includes:

[0052] The acquisition module is used to acquire EEG information, which includes: EEG data corresponding to the training sample set and EEG data of the test subject;

[0053] The processing module is used to select representative moments from the EEG data corresponding to the training sample set through a representative selection mechanism, and extract potential topography at the representative moments; cluster the potential topography and determine the number of microstates through stability evaluation, and generate a microstate template set.

[0054] The selection module is used to perform temporal reconfiguration on the EEG data of the test subject using the microstate template set to obtain the temporal feature vector of the test subject; based on the training sample set, a temporal feature matrix is ​​obtained; for the temporal feature matrix, sparse feature selection is performed to determine the target feature index set, and the target feature index set is applied to the temporal feature vector of the test subject to obtain a feature subset of the test subject;

[0055] The output module is used to construct a classifier with regularization constraints, train a classification model using the training sample set, receive a subset of features of the test object to generate a discrimination score, and output auxiliary diagnostic information to characterize the risk of autism based on the discrimination score.

[0056] Regarding the aforementioned data acquisition module:

[0057] In one specific embodiment of the present invention, the EEG information includes EEG data corresponding to the training sample set and EEG data of the subject to be tested. The EEG data corresponding to the training sample set may be derived from multiple labeled individual EEG signals for system training. The EEG data of the subject to be tested is derived from the individual to be diagnosed and is used to generate diagnostic results.

[0058] Specifically, the acquisition module can acquire data using an EEG acquisition device. For example, the EEG acquisition device can be a multi-channel EEG recorder, such as a standard international 10-20 system electrode cap with 19 or 32 channels, and a sampling frequency of 250Hz to 1000Hz, preferably 500Hz, to capture sufficient EEG signal details. During acquisition, Ag / AgCl electrodes can be used to ensure skin impedance is below 5kΩ to reduce noise interference.

[0059] For the EEG data corresponding to the training sample set, the acquisition module can import a pre-acquired dataset from a database or external storage medium. For example, this dataset may include at least 50 samples, with half from the ASD group and half from the control group. The duration of EEG data for each sample is 5 to 30 minutes, including resting-state or task-state EEG signals. The data format can be EDF (European Data Format) or MAT file for easy subsequent processing.

[0060] Preferably, the collected EEG data is resting-state EEG data. This is because resting-state brain activity better reflects its intrinsic, spontaneous neural network dynamics, and EEG microstates are the electrophysiological representations of these large-scale brain network activities. Therefore, using resting-state data helps to more stably capture the neurodynamic features related to the core pathology of autism, thereby improving the accuracy of subsequent diagnostic models.

[0061] In addition, the acquisition module can integrate software interfaces, such as using Python's MNE library or MATLAB's EEGLAB toolbox, to perform data reading and preliminary format conversion.

[0062] The acquisition module supports both real-time and offline acquisition of EEG data from the subject. In real-time mode, data can be transmitted directly from the EEG signal acquisition device via USB or wireless connection; in offline mode, pre-recorded EEG data files can be uploaded.

[0063] Before acquisition, basic calibration operations can be performed, including setting the reference electrode position, such as selecting a Cz reference or an average reference. To remove low-frequency drift and high-frequency electromyography artifacts, a bandpass filter can be applied to the acquired signal, typically set to a passband range of 0.5 Hz to 40 Hz.

[0064] It should be noted that the filtering range, reference electrode configuration, and acquisition mode can be adjusted according to specific needs in different experimental or clinical scenarios, and this application does not limit them.

[0065] The training sample set undergoes standard preprocessing before entering the processing module. This includes: bandpass filtering (0.5–40 Hz), power frequency notch filtering (50 Hz or 60 Hz, selected according to the local power grid), downsampling to a uniform frequency to eliminate equipment differences, channel quality detection and bad sector interpolation, and rereference. When necessary, independent component analysis or adaptive subspace restoration methods can be used to weaken eye movement and electromyography components. Triggering conditions and thresholds are recorded according to quality control specifications, and processing records are archived along with the samples. The signal obtained through the above process serves as input for representative moment selection and terrain extraction. Specific thresholds and window lengths are given as ranges in the embodiments but are not uniquely limited, to adapt to different devices and populations.

[0066] Regarding the above processing modules:

[0067] In one specific embodiment of the present invention, the processing module is used to extract features from the EEG data corresponding to the training sample set acquired by the acquisition module to generate a microstate template set. Specifically, the module identifies information-rich time points from the EEG data and constructs templates based on clustering and evaluation mechanisms to support subsequent time-series analysis.

[0068] First, representative moments are selected using a representativeness selection mechanism. For example, signal characteristic scores at each time point in the EEG data can be calculated, such as a combined index based on global field power (GFP) and local correlation. Specifically, after standardizing the multi-channel EEG signals, the GFP value at each time point is calculated, and its average spatial correlation within a preset window (e.g., 1-5 seconds) is evaluated to obtain a representative score. Then, threshold screening or ranking methods are applied to select moments with higher scores as representative moments, preferably 5%-20% of each sample's EEG data, to balance computational efficiency and information coverage.

[0069] Secondly, potential topography is extracted at the representative time points. Potential topography refers to the potential value vector of all channels at that time point, which can be obtained by directly extracting the standardized channel data. For example, for 19-channel EEG data, a 19-dimensional vector is extracted at each representative time point and its polarity is normalized, such as by taking the absolute value or unifying the positive direction, to form a topography sample set. This set can be stored in matrix form for easy subsequent processing.

[0070] Next, the potential terrain is clustered. For example, K-means or hierarchical clustering algorithms are used to group the terrain samples into several clusters, each cluster representing a potential microstate. During the clustering process, spatial correlation coefficients, such as Pearson correlation, are used as similarity measures. The initial number of clusters can be preset to 3-10, and iterative optimization is performed to minimize the within-group variance.

[0071] Then, the number of microstates is determined through stability evaluation. Specifically, resampling methods can be used, such as repeatedly extracting subsets from the training sample set for clustering, and calculating the consistency index of the cluster centers, such as the average correlation coefficient or reprojection error. The number of clusters that makes the stability index meet preset conditions, such as maximizing or exceeding a threshold, is selected as the number of microstates, preferably 4-7, to conform to the typical range of EEG microstates.

[0072] The training sample set can be derived from publicly available research databases or self-built clinical cohorts, or a combination of both. For example, the total sample size can be in the hundreds, including both autism spectrum disorder and control groups; age and gender distribution are determined according to clinical inclusion criteria and stratified balance is maintained during training and validation; diagnostic labels are provided and verified by trained clinicians according to standard diagnostic guidelines. The acquisition device can be a multi-channel EEG system compliant with medical electrical safety standards, with 32 or 64 channels being common; electrode layout follows a 10-20 or 10-10 system; sampling frequency is typically in the range of 250–1000 Hz; resting state conditions can be with eyes open or closed, and single recording duration is typically 3–10 minutes. Reference settings can use Cz reference or average reference; electrode impedance is controlled within commonly used engineering thresholds.

[0073] Furthermore, based on the determined number of microstates and the clustering results, a set of microstate templates is generated. For example, the center vector of each cluster is used as a template, and optional normalization processing is performed to form a set containing multiple templates. This set can be used for subsequent temporal matching of EEG data.

[0074] In this way, the processing module achieves effective modeling of the spatiotemporal characteristics of EEG, ensuring that the generated template set is robust and representative.

[0075] Regarding the above selection module:

[0076] In one specific embodiment of the present invention, the selection module performs feature extraction and feature selection based on the microstate template set generated by the processing module and the EEG data acquired by the acquisition module to generate an optimized feature subset, providing input data for the subsequent classification module. This module achieves efficient representation of the temporal dynamic characteristics of EEG through a combination of temporal matching and sparse optimization.

[0077] Specifically, the microstate template set is used to perform temporal reconfiguration on the EEG data of the test subject to generate a temporal feature vector of the test subject. In this process, the similarity between the potential topography at each time point in the EEG signal under test and each template in the template set can be calculated, for example, using the spatial correlation coefficient as a metric, and the template label with the highest similarity is assigned to that time point, thereby forming a microstate sequence.

[0078] Based on the obtained microstate sequence, temporal characteristic indicators can be further calculated, including average duration, frequency of occurrence, and coverage, and these indicators can be combined into a temporal characteristic vector. The dimension of this vector is usually the product of the number of microstates and the number of feature types, preferably between 20 and 50 dimensions, in order to fully characterize the spatiotemporal dynamics of EEG signals.

[0079] Subsequently, for the training sample set, the same temporal matchback method was used to construct a temporal feature matrix. Specifically, for the EEG data corresponding to the training sample set, temporal feature vectors for each sample were generated based on the same microstate template set, and the feature vectors of all samples were stacked row-wise to form a matrix. The number of rows in this matrix equals the number of samples, and the number of columns equals the feature dimension, which is used for subsequent feature selection and model training.

[0080] Next, sparse feature selection is performed on the temporal feature matrix to determine the target feature index set. For example, an embedded feature selection method can be used, such as an L1-regularized linear model or a recursive feature elimination algorithm, and the importance metrics of each feature, such as coefficient magnitude or variation contribution, are evaluated within a cross-validation framework. Feature indices with importance higher than a preset threshold are selected to form the final target feature index set. This feature selection process can be implemented using the scikit-learn library in a Python programming environment, preferably with 5 to 10 iterations to ensure the stability and robustness of the results.

[0081] Finally, the target feature index set is applied to the temporal feature vector of the object under test, and the corresponding components are extracted according to the index to obtain a feature subset of the object under test. This feature subset has a lower dimension than the original vector, which effectively reduces redundant information and improves computational efficiency. Through the above steps, the selection module achieves automated feature filtering and optimization, ensuring the representativeness and applicability of the input data.

[0082] For example, this application employs weighted sparse group selection at the temporal feature level. This method simultaneously implements two types of constraints within a single model: one at the group level that causes the entire group's parameters to tend to zero, and another within the group that causes some column parameters to tend to zero. These are controlled by two regularization strength parameters, which respectively constrain sparsity at the group level and within the group, thus balancing the structural information of groups based on microstates with the discriminative differences in columns within the group.

[0083] The input consists of a grouped temporal feature matrix and its corresponding class labels. Each column of the matrix belongs to a microstate group, and each group contains at least the duration, coverage, occurrence frequency, and several transition probability columns. To eliminate dimensional differences, the matrix is ​​standardized by zero mean and unit scale along each column, and the scaling factor is recorded for subsequent back-substitution and deployment consistency.

[0084] Group weights are used to balance the differences in group size and stability. The group weight is proportional to the square root of the number of columns in the group and is adjusted in conjunction with the group's stability index during the training phase. The higher the stability, the greater the weight can be. All group weights are eventually normalized to a commonly used engineering range.

[0085] Column weights are used to suppress unstable columns. The coefficient of variation of the column in the corresponding training submatrix is ​​used as an indicator of instability. After quantile pruning and interval normalization, this coefficient is mapped to a regularization weight; the greater the variation, the stronger the weight. Within-group column weights and group weights work together.

[0086] For example, the optimization objective consists of two parts: one part is the training loss, such as log loss or squared loss, either of which can be used to keep the training and evaluation caliber consistent; the other part is the sum of two types of weighted regularization terms, where the group-level regularization term uses the length of each group parameter vector as the statistic, and the intra-group regularization term uses the sum of the absolute values ​​of each column parameter as the statistic, and the two are multiplied by the aforementioned group weight and column weight, respectively.

[0087] Two common methods can be used to solve this problem: proximal coordinate descent or proximal gradient. Each iteration first updates the non-regularized gradient step, then performs proximal mapping on each group: first, group-level truncation is performed, then column-by-column soft thresholding is applied within each group; after the update, cyclic polarity and convergence are determined. Convergence is determined based on the relative magnitude of parameter changes between adjacent iterations.

[0088] Regarding the above output module:

[0089] In one specific embodiment of the present invention, the output module is used to construct a classifier with regularization constraints and to train the classification model based on the training sample set in order to generate auxiliary diagnostic results for autism risk.

[0090] The output module may include a model building unit, a training unit, a prediction unit, and an output unit. The model building unit defines the structure and constraints of the classifier; the training unit uses the temporal feature matrix in the training sample set to learn the model; the prediction unit receives a subset of features of the object to be tested and calculates the discrimination score; the output unit generates risk assessment information based on the discrimination score and visualizes the output.

[0091] In its implementation, the classifier can employ an XGBoost model with regularization constraints. This model introduces L1 and L2 regularization terms within the basic gradient boosting framework to constrain the complexity of the tree structure and the weights of leaf nodes, thereby reducing overfitting and improving the model's generalization ability. During training, the input features are a temporal feature matrix filtered by a selection module, and the labels are the diagnostic categories corresponding to the training samples. The training process can employ 5-fold or 10-fold cross-validation to evaluate model stability, and the optimal hyperparameters, such as learning rate, tree depth, subsampling ratio, and regularization coefficient, are determined by minimizing the validation error.

[0092] During the prediction phase, the output module receives a subset of features from the test subject and inputs it into the pre-trained classification model. The classifier calculates the probability output or decision function value for each category based on the distribution of the feature subset and generates a discrimination score. The discrimination score can be mapped to a probability quantification index in the [0,1] interval, representing the relative probability that the test subject belongs to the autism risk category.

[0093] The output module generates auxiliary diagnostic information based on the discrimination score. Specifically, the discrimination score can be classified according to a preset threshold; for example, a score higher than the threshold is marked as high risk, and otherwise as low risk. A report containing the score value, threshold position, and model confidence level can be generated for clinical reference. The output results can be provided through a graphical interface, data files, or a remote interface, enabling doctors to quickly obtain auxiliary decision-making information.

[0094] Through the above structure, the output module can maintain regularization constraints in both the training and prediction stages, ensuring that the classifier's response to input features is stable and interpretable, and realizing the automatic conversion from feature subsets to diagnostic results, thereby improving the accuracy and clinical usability of EEG-assisted diagnosis of autism.

[0095] Optional, see Figure 2 The flowchart below shows a method for extracting potential topography provided in this application, including steps S101 to S104, wherein:

[0096] S101: Divide the measurement channels of the EEG data corresponding to the training sample set into a first channel group and a second channel group;

[0097] S102: The first artifact coefficient is calculated for the channels of the first channel group based on the absolute deviation of the instantaneous amplitude relative to the median of the channel; the second artifact coefficient is calculated for the channels of the second channel group based on the degree of inconsistency of the spatial correlation between the channels; and the low artifact factor is calculated by combining the first artifact coefficient and the second artifact coefficient.

[0098] S103: Based on the low artifact factor, the global field strength quantile at each time point, and the spatial correlation average of the potential topography within the first window length at each time point, the representative score at each time point is calculated.

[0099] S104: Perform non-maximum suppression based on the representative score to determine several representative moments, and extract potential terrain at the representative moments to form a terrain sample set.

[0100] In one specific embodiment of the present invention, the process of extracting potential terrain at representative moments is used to identify moments with stable signals and significant physiological significance from the EEG data corresponding to the training sample set, and to form a terrain sample set based on the spatial potential distribution at that moment for subsequent microstate clustering and template construction.

[0101] In practice, the measurement channels of the EEG data corresponding to the training sample set are divided into a first channel group and a second channel group. Channel division can be based on the symmetry of the channel layout on the scalp or noise sensitivity. For example, channels easily affected by electromyography (EMG) or eye movement artifacts, such as those in the forehead, periocular region, and jawline, are classified as the second channel group, while the remaining relatively stable channels are classified as the first channel group. In this way, differentiated artifact evaluation strategies can be adopted based on the noise characteristics of different channels.

[0102] Furthermore, artifact indices are calculated for the first and second channel groups respectively. For the first channel group, a first artifact coefficient is calculated based on the absolute deviation of the instantaneous amplitude of each channel relative to the median of that channel. This coefficient reflects the degree of signal amplitude anomaly. For the second channel group, a second artifact coefficient is calculated based on the degree of inconsistency in spatial correlation between different channels at the same time point. This coefficient describes the degree of anomaly in spatial potential distribution. Then, the two types of artifact coefficients are combined, and a low artifact factor is calculated using a weighted average or geometric mean, thereby quantifying the overall signal quality at each time point.

[0103] Next, based on the low artifact factor, the global field strength quantile at each time point, and the spatial correlation average of the potential topography within the first window length, a representative score for each time point is calculated. The global field strength is used to measure the spatial activity of the EEG signal at that moment, and the spatial correlation average reflects the consistency of the potential pattern between that moment and adjacent time points. After standardizing these three indicators, a weighted sum is obtained to obtain a representative score that comprehensively reflects the spatiotemporal saliency and stability of the EEG signal.

[0104] Subsequently, a non-maximum suppression algorithm was used to screen representative scoring sequences. Specifically, only time points where the scores reached local maxima were retained within a continuous time window, and the minimum interval between adjacent selected time points was set to 50–100 milliseconds to avoid temporal redundancy. The final selected representative time points corresponded to the steady-state transition points of the EEG signals.

[0105] Potential values ​​from all channels are extracted at the representative time points to form a potential vector, which represents the potential topography. The potential topography corresponding to all representative time points is then set into a matrix and stored as input for subsequent cluster analysis.

[0106] For example, taking resting-state EEG signals as an example, representative moments and terrain extraction were performed on multi-channel EEG data of an individual to be analyzed. The EEG signal had a sampling rate of 500 Hz, a total duration of approximately 5 minutes, and contained simultaneous recordings of 32 channels.

[0107] First, based on the aforementioned representative score calculation method, a representative score sequence covering all time points is obtained. After normalization, the score sequence exhibits a curve that fluctuates over time. Then, a non-maximum suppression algorithm is executed within a defined sliding window: in each window, only time points with local maxima in the score are retained as candidate representative moments, while points with lower scores within the time intervals before and after them are removed. The minimum time interval between adjacent representative moments can be set according to the sampling rate and the time scale of EEG dynamic changes, typically on the order of tens of milliseconds, to avoid information duplication between nearby time points. This method yields a series of temporally evenly distributed and non-overlapping representative moments.

[0108] Furthermore, at each representative moment, the instantaneous potential values ​​of all channels are extracted to form a potential vector. For example, for a 32-channel signal, a 32-dimensional potential vector can be obtained. The potential vectors of all representative moments are arranged in chronological order to form a matrix, with each row corresponding to the potential distribution at a given moment and each column corresponding to the time series value of a channel. This matrix is ​​the potential terrain sample set, which serves as the input for subsequent cluster analysis.

[0109] In this way, low-noise, highly representative spatiotemporal potential distributions can be automatically extracted from large-scale EEG data, effectively improving the accuracy and computational efficiency of microstate analysis.

[0110] Optional, see Figure 3 The flowchart illustrates a method for dividing a first channel group and a second channel group according to an embodiment of this application, including steps S201 to S203, wherein:

[0111] S201: Calculate the artifact sensitivity index of each measurement channel based on the EEG data corresponding to the training sample set, and divide each measurement channel into a first channel group and a second channel group according to the quantile threshold of the artifact sensitivity index.

[0112] S202: Calculate the first artifact coefficient and the second artifact coefficient for the first channel group and the second channel group respectively; determine the contribution as the weight based on the relative change of the first artifact coefficient and the second artifact coefficient in the potential topography clustering stability evaluation index within the corresponding channel group; perform a weighted geometric average on the first artifact coefficient and the second artifact coefficient according to the weight to generate a low artifact factor.

[0113] S203: Based on the training sample set, determine the quantile threshold used for channel division according to the absolute deviation rules of quantiles and medians within the channel group.

[0114] In practice, subjects often exhibit frequent artifacts due to fluctuations in attention or involuntary movements, particularly in the forehead, temporal region, and periocular channels, which are sensitive to blinking, eye movements, and facial muscle electrodes. This application provides an adaptive processing mechanism to address this high artifact characteristic during channel segmentation and the generation of low-artifact factors.

[0115] For example, in a 32-channel EEG acquisition system, channels can be automatically grouped based on their spatial location and historical signal stability. Channels Fp1, Fp2, F7, and F8 near the eyes, and T3 and T4 in the temporal region, are classified into the second channel group due to their susceptibility to eye-movement artifacts; the remaining channels are classified into the first channel group. After reading the EEG data from the training sample set, the system calculates a long-term artifact sensitivity index for each channel, including the channel signal amplitude fluctuation rate and the average correlation coefficient between adjacent channels. The obtained sensitivity index is normalized, and a quantile threshold is calculated. Channels with sensitivity higher than this threshold are automatically classified as high-artifact groups (second channel group), and the rest are classified as low-artifact groups (first channel group).

[0116] During the real-time analysis phase, the system independently calculates artifact indices for the two channel groups. For the first channel group, the mean absolute deviation of the instantaneous amplitude of the channel from its median is calculated; for the second channel group, the spatial correlation matrix of the channels in this group is analyzed and the correlation variance is obtained to reflect the degree of imbalance in the local potential distribution.

[0117] The two artifact indices were further standardized, and the weighting coefficients were determined by combining the stability indices obtained from the historical clustering stability analysis of each channel group.

[0118] For example, channels with high stability are given slightly higher weights to ensure the reliability of the overall signal evaluation. A low-artifact factor, with a value between 0 and 1, is obtained through a weighted geometric average to quantitatively reflect the signal quality at each time point.

[0119] After statistically analyzing the distribution of low-artifact factors across multiple training samples, a threshold can be adaptively set using the intra-channel quantile and median absolute deviation (MAD) rule. When the low-artifact factor exceeds this threshold, the system marks that time period as an artifact enhancement interval, automatically reducing its weight or removing it during subsequent representative time-time filtering.

[0120] Thus, the channel segmentation and artifact quantification method of this application can dynamically adjust the artifact features of different regions in the autism EEG signal without manual setting. It is suitable for multi-channel EEG acquisition scenarios and can improve the stability of the input signal and the accuracy of subsequent microstate extraction.

[0121] Optional, see Figure 4 The flowchart of a method for clustering potential terrain provided in this application embodiment includes steps S301 to S302, wherein:

[0122] S301: Based on the measurement channel spatial layout of the EEG data corresponding to the training sample set, establish a channel adjacency graph, perform polarity unification and mirror mapping on the potential terrain, calculate the spatial spectral energy ratio of each potential terrain on the graph Laplacian eigenvector basis of the channel adjacency graph, and calculate the correlation coefficient between each potential terrain and the mirror potential terrain as the mirror symmetry.

[0123] S302: Perform constrained clustering on the potential terrain within a preset set of candidate microstates, with the goal of maximizing the spatial correlation within the group, and impose constraints on the spatial spectral energy ratio of each cluster center not being greater than a first upper limit and the mirror symmetry not being less than a first lower limit.

[0124] To ensure that the microstate templates balance physiological rationality and statistical stability, this application performs constrained clustering on the potential topography. This process constructs a channel adjacency graph based on the channel spatial layout, completes polarity unification and mirror mapping, and calculates the spatial spectral energy ratio and mirror symmetry for each potential topography. Subsequently, cluster optimization is performed within the set of candidate microstates, and constraints are imposed on the cluster centers: the spatial spectral energy ratio is not greater than a first upper limit, and the mirror symmetry is not less than a first lower limit.

[0125] In practice, a weighted adjacency matrix is ​​constructed based on the three-dimensional coordinates or standardized planar coordinates of the acquisition cap, such as a 10-20 or 10-10 system, with channels as graph nodes and the distance between channels as the metric.

[0126] For example, channels with an Euclidean distance less than a set radius can be considered adjacent, with edge weights using Gaussian kernels or inverse distance weights, and a normalized graph Laplacian matrix obtained using the degree matrix for subsequent graph spectrum expansion.

[0127] The potential topography vectors are arranged in channel order. To eliminate the influence of equivalent polarity reversal, the system compares the potential topography with a reference direction. The reference direction can be the smoothed shape of the potential topography or the direction corresponding to the initial center of the template. When the comparison result shows a negative correlation, the potential topography is assigned the opposite sign to ensure consistency of signs. Mirror mapping establishes channel pairing relationships based on the cranial midline. The cranial midline is formed by the geometric axis determined by the nasal root point, occipital point, and the two preauricular points. The system constructs channel pairing relationships based on this midline, forming a mirror mapping matrix. For channels that are not strictly paired, the corresponding mirror potential value can be determined by interpolation based on the spatial coordinates of the symmetrical position of the midline.

[0128] The potential topography after polarity unification is expanded on the Graph Laplacian eigenvector basis, and the low eigenvalues ​​correspond to the smooth (low frequency) components.

[0129] For example, a feature subspace less than or equal to a certain spectral threshold can be taken as the low-frequency subspace. The energy proportion of the potential terrain in this subspace is calculated and recorded as the spatial spectral energy ratio, which is used to characterize the spatial smoothness of the terrain. The mirror symmetry can be calculated by setting weights according to the distance from the channel to the midline after removing the mean, and calculating the weighted correlation coefficient between the potential terrain and its mirror potential. The higher the value, the stronger the left-right symmetry.

[0130] Within the set of candidate microstates, clustering is performed using correlation distance or cosine distance, with the optimization objective being to maximize intra-group similarity. To satisfy physiological and physical constraints, two types of constraints are introduced for each cluster center: the spatial-spectral energy ratio must not exceed a first upper limit, and the mirror symmetry must not be lower than a first lower limit.

[0131] For example, constraints can be incorporated into the penalty term of the objective function, or projection corrections can be performed after each center update: when the spatial spectral energy ratio of the center exceeds the upper limit, its projection in the Graph Laplace high-frequency subspace is scaled down proportionally; when the mirror symmetry is below the lower limit, the symmetry is improved by using a weighted average of the center and its mirror.

[0132] For example, the penalty weights can be selected within an empirical range to strike a balance between fit and physical realizability.

[0133] Initial centers can be obtained using k-means++ or a sparse set of representative terrain features. Sample assignment and center updates are performed in each round, with constraint corrections implemented after each update. Convergence occurs when the center change between two adjacent rounds falls below a set threshold or the maximum number of iterations is reached. To reduce the uncertainty of random initialization, the same number of candidates can be solved multiple times, and the solution with the optimal objective function value is selected. The clustered terrain features form a micro-state template set for subsequent temporal reconfiguration and feature extraction.

[0134] For example, on 32-channel resting-state data, the channel adjacency graph can be given by a locally connected structure consisting of three to five neighboring channels; the graph spectrum expansion can select a small number of low eigenvalue vectors as low-frequency basis to ensure spatial smoothness prior; when calculating mirror symmetry, the temporal region and extra-side channels that are far from the midline can be given slightly higher weights to enhance the constraint effect on obviously lateral terrain.

[0135] In this way, the resulting cluster centers usually exhibit a smoother, more symmetrical potential distribution, which facilitates obtaining a stable microstate sequence in subsequent time-series reconfiguration.

[0136] Optionally, determining the number of microstates through stability evaluation and generating a microstate template set includes:

[0137] For each candidate microstate, hierarchical resampling is performed on a sample-by-sample basis within the training sample set to form multiple sets of potential terrain. After performing the constrained clustering on each set of potential terrain, the obtained cluster centers are back-projected onto the corresponding potential terrain set. The reprojection error and the matching consistency of the cluster centers are calculated, and a weighted combination is performed with the spatial spectral energy ratio and mirror symmetry of the corresponding cluster centers to determine the stability score.

[0138] Based on the stability score sequence in the training sample set, calculate the stability score gain of the number of adjacent candidate microstates, and define the preset score gain threshold as the second threshold.

[0139] The number of microstates that maximizes the stability score and whose score gain for the number of adjacent candidate microstates is less than a second threshold is selected as the target number of microstates, and a set of microstate templates is generated based on the clustering results corresponding to the target number of microstates.

[0140] To ensure the objectivity and reproducibility of determining the number of microstates across data batches, this application conducts a stability evaluation within the candidate microstate quantity set, selects the target microstate quantity based on the joint score of resampling consistency and physical consistency, and generates a microstate template set under this quantity.

[0141] In practice, stratified resampling is performed within the training sample set. The stratification criteria may include sample category, such as autism group / control group, collection batch, or conversation source, to ensure that the composition of each resampling is representative.

[0142] For example, the number of candidate microstates can be an integer ranging from 3 to 8, and the number of resampling groups can be set to several tens. Each group of resampling data is subjected to constrained clustering according to the method described above to obtain the cluster center set of that group.

[0143] To quantify the interpretability of the cluster centers, backprojection is performed on the cluster centers within each resampling group: for each potential topography in the group, the spatial correlation coefficient is calculated with each center, and the maximum correlation coefficient is used as the fitting similarity of that sample; the average fitting similarity of all samples in the group is taken as the matching consistency of the group; simultaneously, the reprojection error of the group is statistically analyzed according to the rule that "the lower the fitting similarity, the higher the error," for example, by recording the proportion of samples with fitting similarity below a preset lower limit, or by inverting the similarity and averaging it, to characterize the residual level of the centers on the data. To avoid differences in caliber, the above two quantities are linearly normalized within the number of candidates to unify them to the same comparison scale.

[0144] To incorporate physiological and physical properties into the evaluation, the spatial spectral energy ratio and mirror symmetry of each cluster center were calculated using the method described above, and then averaged after normalization to obtain physical consistency.

[0145] This yields three comparable scoring criteria: matching consistency, reprojection error (converted to the principle that smaller errors are better), and physical consistency.

[0146] For example, the three weights can be set to be equal or unequal based on validation set experience, and then a weighted sum can be performed to obtain the stability score for that number of candidates. In scenarios with small samples or high noise, the weight of physical consistency can be appropriately increased to enhance the suppression of unstable clustering.

[0147] After forming a stable scoring sequence based on the number of candidates, the score increase between adjacent candidate numbers is calculated, i.e., the improvement of the current score relative to the previous candidate number score. A preset "score increase threshold" is defined as the second threshold; this threshold can be adaptively determined by statistically analyzing the quantiles of the increase on the training sample set or according to the rule of median absolute deviation.

[0148] The selection rules are as follows: priority is given to the position where the stability score reaches its peak; when there are multiple local peaks, priority is given to the smaller number of candidates whose score increase is lower than the second threshold and no longer shows significant improvement, so as to avoid template redundancy and subsequent feature expansion caused by excessive subdivision.

[0149] After determining the number of target microstates, constrained clustering is reproduced on the full training sample set to generate the final microstate template set. Initialization can use a set of centers with higher scores in the resampling phase, or average multiple sets of centers after correlation alignment, to reduce initialization randomness. The output template set includes the center potential topography vector of each template and its corresponding summary statistics, such as a summary of intra-group similarity distribution, for subsequent temporal reassembly and chronological feature extraction.

[0150] Optionally, calculating the spatial spectral energy ratio of each potential topography on the graph Laplacian eigenvector basis of the channel adjacency graph includes:

[0151] Based on the normalized graph Laplacian matrix of the channel adjacency graph, the eigenvalue sequence and eigenvector sequence are obtained, and the first spectral threshold is determined according to the maximum spectral gap in the eigenvalue sequence.

[0152] The ratio of the projected energy of the potential terrain after polarity unification processing on the set of basis vectors with eigenvalues ​​not greater than the first spectral threshold to the projected energy on all basis vectors is determined as the spatial spectral energy ratio.

[0153] To measure the spatial smoothness of potential topography, this application expands the potential topography on the graph Laplacian eigenvector basis of the channel adjacency graph and uses the proportion of low-frequency component energy in the total energy as the spatial spectral energy ratio. This index is used to impose smoothness constraints on cluster centers in constrained clustering.

[0154] For example, channel coordinates use standardized two-dimensional coordinates based on the international 10-20 or 10-10 system, or the projection of three-dimensional coordinates obtained from device calibration onto the scalp tangent plane. Each channel is treated as a graph node, and weighted edges are established based on spatial distance. For example, the K-nearest neighbor method is used by default, where K is an integer from 3 to 5. In layouts with significantly uneven channel spacing, radius adjacency can be used instead, with the radius taking a commonly used multiple of the average distance between channels. Edge weights decrease as distance increases, using Gaussian or inverse distance weights, and metric normalization is performed across the entire graph to form a normalized graph Laplacian matrix. If bad channels or missing coordinates exist, neighbor interpolation is used to fill in the gaps first; if this is not possible, the minimum necessary edges are established using the nearest valid channel to ensure connectivity.

[0155] The normalized graph Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain a sequence of eigenvalues ​​arranged in ascending order of eigenvalues ​​and their corresponding eigenvector sequences. These eigenvector sequences constitute the graph Laplacian eigenvector basis of the channel adjacency graph.

[0156] To determine the boundary between low and high frequencies, the feature value sequence is scanned, the position of the largest gap between adjacent feature values ​​is identified, the set of feature vectors before this position is determined as the low frequency basis, and the first spectral threshold is determined accordingly.

[0157] For example, gap retrieval can be performed within the first few dozen feature values ​​to reduce the interference of high-order noise on threshold selection; when no obvious gaps appear, quantile rules or a fixed number of low-order feature vectors can be used as low-frequency bases to ensure the executability of the method on different datasets.

[0158] Potential topography vectors are arranged in channel order and used for spectral expansion after polarity unification. During expansion, the potential topography is represented on both the low-frequency basis and the full basis. The cumulative energy corresponding to the two sets of expansion coefficients is calculated, and the spatial spectral energy ratio is determined according to the ratio of low-frequency energy to total energy. To avoid the influence of constant terms on the statistical caliber, the potential topography can be mean-removed before expansion. When the total energy is too small, which may lead to an unstable ratio, a minimum energy clamping strategy can be set or the ratio calculation at that moment can be skipped to ensure numerical robustness.

[0159] For example, 32-channel data can typically use three to five nearest neighbors to build adjacency relationships, and the size of the low-frequency basis can fall within a small integer range; for 64-channel data, the number of nearest neighbors or the radius of the connection can be appropriately relaxed to maintain the stability of the graph structure. In the case of missing coordinates or individual bad channels, spatial interpolation, nearest neighbor substitution, or removal and normalization can be used to ensure that the eigenvector basis and the energy statistics process are continuous and consistent.

[0160] Optionally, calculating the correlation coefficient between each potential terrain and its mirror potential terrain as the mirror symmetry includes:

[0161] The midline of the skull is determined based on the spatial layout of the measurement channels of the EEG data corresponding to the training sample set, and the channel correspondence is established based on the mirror image of the coordinates of each measurement channel relative to the midline of the skull to determine the mirror mapping matrix.

[0162] The weights of the measurement channels are determined by normalizing the distances from each measurement channel to the midline of the skull.

[0163] The potential terrain vector after polarity unification processing and the potential terrain vector after the mirror mapping matrix transformation are calculated in the space after removing constant components, and the weighted correlation coefficient is used as the mirror symmetry.

[0164] To measure the left-right symmetry of potential topography, this application determines the midline of the skull based on the spatial layout of scalp channels, establishes channel mirror correspondence and mirror mapping matrix, and calculates weighted correlation coefficients in the space after removing constant components as mirror symmetry, which are used to apply symmetry constraints in constrained clustering.

[0165] The midline of the skull can be determined using the spatial coordinates provided by the acquisition cap. The coordinates can be standardized two-dimensional projected coordinates from the international 10-20 or 10-10 system, or three-dimensional coordinates obtained from equipment calibration. The geometric definition of the midline uses the nasal root point, occipital point, and the two preauricular points as references, and fits a symmetry axis running through the anterior-posterior direction using the minimum deviation principle. When only two-dimensional projection is available, the midline is a straight line; when using three-dimensional coordinates, the channel coordinates can be projected onto the scalp tangent plane before fitting the midline. To improve robustness, outlier channels can be removed during the fitting process, or a robust regression caliber can be used.

[0166] The channel mirror correspondence is established after the centerline is determined. For each channel, the corresponding channel is found according to the mirror position of the centerline. When the mirror position does not completely coincide with the actual electrode, several nearest channels can be selected near the mirror position, and interpolation weights are assigned according to the spatial distance from the mirror position to obtain the potential estimate of the channel on the mirror side. This forms a sparse mirror mapping matrix: each row corresponds to the linear combination coefficient of a channel on the mirror side, and unpaired centerline channels can be regarded as self-mapping or given neutral treatment in related calculations.

[0167] Channel weights are determined and normalized based on the distance of the channel from the midline. The distance can be calculated directly within the coordinate system, and the weight increases with the distance, thus giving greater influence to channels that are far from the midline and more sensitive to differences on the opposite side. To avoid extreme weights affecting the results, the distance can be scaled using quantiles, and upper and lower limits can be set for the weights.

[0168] For example, channels located in the high quantile of the sample distribution are given higher weights, while channels located in the low quantile are given lower weights; channels in the middle line may be given the minimum weight or handled separately in the relevant calculations.

[0169] After polarity unification, the potential terrain vector participates in the correlation metric. Polarity unification is achieved by comparing it with a reference direction, which can be a smoothed version of the terrain or the initial template direction; when the correlation is negative, the sign is uniformly reversed. Subsequently, the original potential terrain vector undergoes mean removal processing, and a corresponding mirror potential vector is constructed (obtained by applying a mirror mapping matrix to the original terrain), which also undergoes mean removal processing. The calculation of the weighted correlation coefficient follows a collaborative statistical process involving channel weights: first, the weighted covariance metric is obtained by aggregation along the channel dimension according to weights, and simultaneously, the weighted energy metrics of the two vectors are obtained separately, which are then normalized to the correlation coefficient. To avoid numerical instability, a minimum energy clamping strategy or skipping the calculation at a certain moment can be used when the energy metric is too small.

[0170] For example, in a 32-channel layout of a 10-20 system, F3 and F4, C3 and C4, P3 and P4, O1 and O2, F7 and F8, T3 and T4 are common symmetrical pairs; the midline electrodes Fz, Cz, and Pz can be treated neutrally as described above. When the three-dimensional coordinates are incomplete or bad channels exist, the signal estimation of the mirror position can be completed by neighborhood interpolation, or the nearest channel can be used to replace the signal and the weights can be re-normalized when constructing the mapping matrix. To ensure consistency in the data across different batches, the channel weight normalization rule and the midline fitting rule are kept consistent during the training and application phases.

[0171] In the international 10-20 or 10-10 electrode system, the letters in the electrode markings represent anatomical regions: F for the frontal region, C for the central region, P for the parietal region, O for the occipital region, and T for the temporal region. Numbers distinguish the left and right hemispheres, with odd numbers on the left and even numbers on the right. The suffix "z" indicates a position near the midline. Therefore, F3 is located in the left frontal region, F4 in the right frontal region, C3 and C4 in the left and right central regions respectively, P3 and P4 in the left and right parietal regions respectively, O1 and O2 in the left and right occipital regions respectively, and T3 and T4 in the left and right temporal regions respectively; Fz, Cz, and Pz are located near the midline. Based on this naming convention, F3-F4, C3-C4, P3-P4, O1-O2, F7-F8, and T3-T4 can be considered common left-right mirror pairings. Midline electrodes such as Fz, Cz, and Pz are treated neutrally in mirror pairings and related calculations.

[0172] The mirror mapping matrix is ​​used to generate a corresponding potential vector on the mirror side for any potential topography. Each row of the matrix corresponds to the linear combination coefficient of a channel on the mirror side: when there is a definite one-to-one mirror channel, the coefficients are concentrated on that mirror channel; when the mirror position does not completely coincide with the actual electrode, several nearest channels can be selected near the mirror position, and interpolation weights are assigned according to the spatial distance from the mirror position, and the weights of that row are normalized so that the sum of the weights of all rows is one. The mirror of the centerline electrode can be set as self-mapping or assigned a minimum influence weight in related calculations to maintain the stability of the symmetry evaluation.

[0173] Channel weights are used to emphasize channels that are more sensitive to differences on the opposite side in the weighted correlation coefficient. The weights are calculated and normalized from the geometric distance from the channel to the midline, and can be quantile scaling and upper / lower bound truncation can be used to avoid extreme weights dominating the results while maintaining the distinctiveness of channels far from the midline. Consistent distance metrics and normalization rules should be used in both the training and application phases to ensure the comparability of different batches of data.

[0174] Bad channels refer to channels marked as abnormal during data acquisition or quality control, such as those with prolonged high impedance, saturation, persistent missing channels, or significant drift. Neighborhood interpolation can be used to complete bad channels, constructing a linear estimate based on the potential and spatial distance of several surrounding channels. If stable completion is not possible, the nearest valid channel can be used to replace the bad channel when constructing the mirror mapping matrix, and the weights of that row can be renormalized to ensure the sparsity and numerical stability of the mapping matrix.

[0175] When performing mirror symmetry calculations, the potential topography needs to undergo polarity unification and mean removal. Polarity unification is achieved by comparing with a reference direction, which can be the smoothed shape of the topography or the initial template direction; if the correlation is negative, the opposite sign is used to eliminate the equivalent polarity reversal. Mean removal is used to eliminate the influence of constant terms on the correlation metric. The weighted correlation coefficient participates in statistical aggregation with channel weights, and the resulting value is within a bounded range; the higher the value, the stronger the left-right symmetry. To avoid numerical instability, when the weighted energy is extremely small or the weights are concentrated in a few channels, minimum energy clamping or skipping the calculation at that moment can be used, and the report should indicate that the moment is invalid.

[0176] The above-mentioned standards ensure that electrode naming, mirror relationship, channel weight, and bad sector handling have clear meanings and consistent usage in the instruction manual, making it easy for readers to directly reproduce the mirror symmetry calculation process and connect it with the symmetry constraints of constrained clustering.

[0177] Thus, the resulting weighted correlation coefficient varies within the range, with higher values ​​indicating stronger left-right symmetry. This mirror symmetry serves as a constraint for constrained clustering, limiting the geometrical laterality of cluster centers and thereby improving the physiological rationality and cross-sample stability of the micro-state template.

[0178] Optionally, the subset of features of the object to be tested includes:

[0179] The time-matrix feature matrix is ​​divided into several feature groups according to microstates. Each feature group includes at least the average duration, time coverage, occurrence frequency, and transition probability vector from the microstate to other microstates, thus obtaining the grouped time-matrix feature matrix.

[0180] Multiple training sub-matrices are generated on the grouped temporal feature matrix. The training sub-matrices are obtained by performing hierarchical bootstrap sampling on the sample rows. For each training sub-matrice, a regularization weight is calculated based on the coefficient of variation of each feature in the corresponding training sub-matrice, and a candidate feature index set is calculated using weighted group sparse selection.

[0181] Stability statistics are performed on multiple candidate feature index sets, the selection frequency of each feature is calculated and compared with a third threshold to determine the target feature index set;

[0182] The target feature index set is applied to the temporal feature vector of the object under test, and the corresponding components are extracted according to the index order to obtain the feature subset of the object under test.

[0183] To enhance the stability and reproducibility of classification without increasing model complexity, this application employs grouped representation, hierarchical bootstrapping, and weighted group sparse selection at the temporal feature level, and uses the selection frequency as a stability criterion to determine the target feature index set, ultimately generating a feature subset of the object to be tested.

[0184] The temporal features are grouped on a microstate basis. Each microstate corresponds to a set of features, including at least the average duration, temporal coverage, occurrence frequency, and transition probability vector from that microstate to other microstates. The temporal feature vectors of the training samples are stacked row-wise according to the above criteria to obtain a temporal feature matrix, and the grouped temporal feature matrix is ​​constructed accordingly. The column grouping information remains fixed in the subsequent selection process to facilitate the implementation of group sparsity constraints.

[0185] Multiple training sub-matrices are generated on the grouped temporal feature matrix. The sub-matrices are obtained by stratified bootstrapping of sample rows. Stratification criteria can include sample category, collection batch, or session source, ensuring that the category distribution of each sub-matrix is ​​consistent with the overall distribution. Each sub-matrix reuses only the rows of the training samples, without changing the definition and grouping of the feature columns, thus guaranteeing comparability between different sub-matrices.

[0186] For each training submatrix, calculate the weights used for regularization. Calculate the coefficient of variation for each feature within that submatrix, column by column, and map the features to regularization weights according to their degree of variation. The weight mapping follows a monotonicity principle; features with greater variation are assigned stronger regularization to suppress unstable features. To avoid the influence of extreme values, quantile pruning and interval normalization can be applied to the coefficients of variation.

[0187] After the weights are calculated, weighted group sparse selection is performed: sparse constraints are applied at the group level, with each microstate as a group, while preserving the selection freedom of columns within each group. This allows the model to retain a small number of the most discriminative columns within the same microstate group. This process outputs a set of candidate feature indices, corresponding to the set of selected columns for the current submatrix.

[0188] Stability statistics are performed on multiple candidate feature index sets obtained from all training sub-matrices. The selection frequency is calculated column by column and compared with a third threshold. Columns with a selection frequency not lower than the third threshold are included in the target feature index set. The third threshold can be adaptively set according to the stability distribution of the training sample set, for example, by determining the specific value range based on the quantile of the selection frequency or the performance of the reference validation set. When there are highly correlated adjacent columns, only one column can be retained according to the correlation merging strategy or retained with a fixed priority.

[0189] The target feature index set is applied to the temporal feature vector of the object under test, and the corresponding components are extracted according to the index order to obtain the feature subset of the object under test. To ensure consistency in the deployment phase, the index set and its column order are kept consistent with those in the training phase and are fixed in the model file along with the classifier.

[0190] For example, when the number of microstates is within a common range, the number of columns in each microstate group is within a controllable range; the number of hierarchical bootstrapping rounds can be set to several tens; and the third threshold can be selected to be in the middle to high quantile range. This process can suppress unstable columns while preserving key temporal features, reduce input dimensionality, and improve the robustness of the classifier under small sample and cross-batch conditions.

[0191] Optionally, the step of performing temporal reconfiguration of the EEG data of the test subject using the microstate template set includes:

[0192] Based on the time domain standard deviation of the EEG data corresponding to the training sample set and the EEG data of the test subject in each measurement channel, the channel gain normalization matrix is ​​determined, and the channel gain normalization matrix is ​​used to perform channel gain normalization processing on the EEG data of the test subject.

[0193] On the normalized graph Laplacian eigenvector basis of the channel adjacency graph, orthogonal registration is performed on the potential terrain set of the object under test after the channel gain normalization process and the microstate template set to determine a first alignment matrix, and the microstate template set is transformed with the first alignment matrix to obtain an alignment template set;

[0194] Based on the alignment template set, labels are assigned to each time point of the EEG data of the test subject according to the criterion of maximizing spatial correlation, thereby generating the microstate sequence of the test subject.

[0195] To obtain stable microstate assignments under different individual and device conditions, this application performs channel gain normalization before time-series reconfiguration, and performs orthogonal registration of the terrain set and microstate template set of the test object within the graph Laplacian eigenvector basis. Subsequently, label assignment is performed according to the criterion of maximizing spatial correlation to generate the microstate sequence of the test object.

[0196] The purpose of channel gain normalization is to eliminate channel amplitude calibration differences across individuals, sessions, and devices. The normalization coefficients are derived from the comparison of the time-domain statistics of the training sample set and the test object on each measurement channel.

[0197] Specifically, for each channel, its time-domain standard deviation or equivalent robust statistic is calculated. The gain coefficient of that channel is determined by the ratio of the training-side statistic to the test-side statistic, forming a channel gain normalization matrix. The normalization matrix is ​​then scaled channel by channel on the EEG data of the test subject so that the processed terrain set is consistent with the training-side template in terms of dimensions and amplitude.

[0198] For example, extreme coefficients can be clamped and their gain coefficients can be supplemented by neighborhood interpolation when bad sectors exist, thus ensuring the numerical stability of the normalization process.

[0199] Orthogonal registration is performed within the normalized graph Laplacian eigenvector basis of the channel adjacency graph to align the spatial coordinate systems of the terrain set and the microstate template set. The system projects the terrain set and template set, after channel gain normalization, onto the same subspace of the graph Laplacian eigenvector basis, preferably using a subspace spanned by low-order eigenvectors to enhance noise suppression.

[0200] Within this subspace, an orthogonal transformation preserving length and angle is obtained based on the cooperative statistics of the two sets, serving as the first alignment matrix. This can be implemented using an orthogonal registration method based on singular value decomposition to obtain a uniquely determined rotation or rotation plus reflection transformation. After obtaining the first alignment matrix, a linear transformation is performed on the microstate template set to form an alignment template set for subsequent assignment.

[0201] Label assignment is based on maximizing spatial correlation. The system calculates the spatial correlation coefficient between the potential topography of the target object at each time point and each template in the alignment template set. The template label with the highest correlation coefficient is selected as the microstate label for that time point, and the corresponding correlation coefficient is recorded as the matching strength. To reduce high-speed jitter, a short-term smoothing strategy can be introduced, such as setting a minimum dwell time or maintaining the previous label when the correlation between adjacent time points is close. For time periods marked as low quality by prior quality control, no label can be assigned or a missing label can be used as a placeholder, and this can be addressed during statistical temporal characteristics analysis.

[0202] For example, for 32-channel resting-state or task-state data, a channel adjacency graph can be constructed using 3 to 5 nearest neighbors. The dimension of the low-order feature vector subspace can fall within a small integer range to balance expressiveness and robustness. Channel gain statistics can be updated within a few-second windows to accommodate the slow drift of long records. When encountering missing coordinates or bad channels, the system ensures the continuity of the projection and registration process through spatial interpolation and weight normalization. The alignment matrix and alignment template set are bound and saved with the current session to ensure that a consistent spatial caliber is used in subsequent feature statistics and classification stages.

[0203] In this way, temporal alignment is achieved in both amplitude scale and spatial coordinates, label assignment is completed based on a unified similarity metric, and the output microstate sequence of the test object can be directly used for statistical analysis of temporal features and subsequent classification. This process reduces the impact of cross-device and cross-session differences on microstate recognition and improves the stability of microstate sequences.

[0204] Optionally, constructing a classifier with regularization constraints includes:

[0205] The lower and upper limits of the preset risk threshold range are defined as the fourth and fifth thresholds, respectively;

[0206] On the training sample set, for the risk threshold interval defined by the fourth threshold and the fifth threshold, pre-classification models with different class weight ratios are trained respectively and cross-validation predictions are generated. Based on the cross-validation predictions, the average net benefit of decision curve analysis is calculated within the risk threshold interval to obtain the evaluation value of each class weight ratio.

[0207] The class weight ratio that maximizes the average net benefit is selected as the target weight ratio, and the classifier with regularization constraints is trained using the training sample set and the target feature index set under the target weight ratio to obtain the classification model.

[0208] To ensure that the classifier maintains stable clinical decision-making value under different risk assessment criteria, this application introduces the average net benefit of risk threshold interval and decision curve analysis as the basis for weight selection when constructing a classifier with regularization constraints, and completes model solidification and deployment under the target weight ratio.

[0209] The risk threshold range is defined by the fourth and fifth thresholds, describing the permissible probability range for a clinically determined high risk. The upper and lower limits of the range can be set based on clinical experience, historical interpretation records, or the risk distribution of the training sample set to ensure consistency with the decision-making criteria in actual application. Once determined, the range remains unchanged during the training and validation phases.

[0210] A set of class weight ratios is constructed on the training sample set, ranging from equal weights to multiple weights applied to minority classes. The weighting is set using several discrete levels, and the actual weight sequence is recorded in the metadata. Hierarchical cross-validation is performed for each weight ratio to obtain the predicted probabilities at the sample level paired with the true labels. The number of folds, random seed, and stratification of the cross-validation are fixed to prevent information leakage. To improve the reliability of threshold determination, probability calibration can be performed on the predicted probabilities within each fold. Calibration can use order-preserving or binning methods. The calibration mapping is fitted only in the training fold and applied in the corresponding validation fold. The calibration method and parameters are recorded in the metadata.

[0211] A threshold grid with equal spacing is constructed between the fourth and fifth thresholds, with the number of grid points ranging from 20 to 100, including the endpoints. The grid size and endpoint inclusion details are recorded in the metadata. For each threshold point, the predicted probability is bisected by the threshold, resulting in a positive judgment set and a negative judgment set for the model. The number of true positives judged as positive is counted, and the number of true negatives judged as positive is counted, with the overall sample size used as the normalization base. The cost of false positives is reduced according to the odds corresponding to the threshold, yielding the net benefit value for that threshold point. Simultaneously, the net benefits of both the all-positive and all-negative reference strategies at this threshold point are calculated and used as reference curves. When the model's net benefit is not higher than that of the reference strategy, the threshold point can be marked as invalid or truncated to the highest reference value to avoid invalid segments affecting subsequent averaging. To reduce the impact of random disturbances, the threshold-net benefit sequence can be lightly smoothed; the smoothing window and boundary handling methods are recorded in the metadata.

[0212] For small sample sizes, repeated cross-validation or bootstrap resampling can be used to repeatedly estimate the net benefit, and the average can be taken among the repeated estimates. The number of repeated estimates is written into the metadata.

[0213] The class weight ratio that maximizes the average net benefit is selected as the target weight ratio. Under the target weight ratio, a classifier with regularization constraints is trained using the training sample set and the determined target feature index set, and overfitting is prevented by early stopping and validation set monitoring. For example, early stopping can be triggered after several rounds without improvement; the random seed and data partitioning strategy can be fixed to improve the reproducibility of results. After training, the model parameters, target weight ratio, feature index order, and risk threshold range are fixed as a unified standard for the inference phase.

[0214] During the inference phase, the output module receives a subset of features of the test object and generates a discrimination score. The discrimination score is used to present a continuous risk probability, or to provide risk indications consistent with clinical standards within an interval defined by a fourth and fifth threshold. To facilitate integration into clinical workflows, the report can simultaneously provide the score value, risk interval, and explanation of the currently used category weight ratio, ensuring comparability and traceability between different batches and different departments.

[0215] In this way, the selection of classifier weights and regularization training are aligned based on the average net benefit, thereby improving the practical decision-making value while keeping the model complexity under control.

[0216] For example, the classifier in this application employs a gradient boosting tree model. The model consists of several decision trees stacked sequentially in an iterative order. In each round, a new tree is trained based on the residuals of the previous round and combined additively with the outputs of the existing trees. The weak learner is a binary decision tree, where internal nodes divide samples using a single feature threshold, and leaf nodes store the output increment for that round. The regularization constraints include two parts: tree structure constraints and parameter constraints. The structure constraints limit the tree depth, the number of leaf nodes, and the minimum sample size, while the parameter constraints limit the amplitude of the leaf node outputs and the learning rate. Together, they suppress overfitting.

[0217] The input features are vectors of temporal features filtered through the target feature index set. The vector dimension is determined by the number of microstates and the types of features, and remains consistent throughout the training and inference phases. The output is a risk probability or an equivalent continuous score between 0 and 1, representing the relative likelihood that the test subject belongs to a high-risk category. Labels are derived from clinical annotations or audited scale conclusions of the training sample set, and the category definitions are consistent with the risk caliber of the training task.

[0218] The training process includes data partitioning, probability calibration, weight search, and model solidification. Data partitioning employs cross-validation, with the number of folds set within a commonly used range. To improve reproducibility, a fixed random seed and consistent stratified sampling caliber can be maintained. For each class weight ratio, a preliminary model is trained with the same range of hyperparameters. These hyperparameters include the number of trees, tree depth, learning rate, leaf node output penalty coefficient, sample subsampling ratio, and feature subsampling ratio. These values ​​can be set according to empirical ranges and monitored using a validation set; early stopping is triggered when the validation metrics no longer improve after several rounds of updates.

[0219] The risk threshold interval is defined by a fourth and a fifth threshold, which can be derived from historical clinical interpretation suggestions, statistical distribution of training samples, or pre-defined safety strategies. Within this interval, the risk thresholds are scanned with a uniform step size. The net benefit for each threshold is calculated, and the average of these values ​​over the interval is taken as the average net benefit for that group of class weight ratios. To reduce the impact of random noise, the threshold scan curve can be smoothed, or the results of multiple cross-validation repetitions can be averaged. The combination with the largest average net benefit among multiple group class weight ratios is selected as the target weight ratio. Under this weight ratio, the classifier is trained using the full training sample set and the target feature index set to obtain the final model.

[0220] Probability calibration can employ methods such as equidistant binning or order-preserving calibration as needed to better align model scores with empirical probabilities. The calibration samples should be consistent with those used in training and validation, and the calibrated mappings should be consistently used during the inference phase.

[0221] During the inference phase, the system receives a subset of features of the target object and outputs a discrimination score. When discrete risk warnings are required, a threshold consistent with the clinical procedure can be selected within the range defined by the fourth and fifth thresholds for judgment. To ensure deployment consistency, model parameters, target weight ratios, target feature index sets, risk threshold ranges, and calibration mappings are uniformly fixed and saved with version numbers after training. The inference and reporting modules load from the same version snapshot.

[0222] The search range for class weight ratios covers an area from equal weighting to multiple weighting for minority classes, with the step size adjustable to several levels. The number of trees can be set to hundreds, with tree depth and learning rate adjusted in tandem within a commonly used range. The sampling ratio is selected from a suitable range of 0 to 1 to balance bias and variance. The above numerical ranges are not limited to specific values ​​and can be determined according to the training sample size, feature dimensions, and target performance using conventional parameter tuning strategies.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost, characterized in that, include: The acquisition module is used to acquire EEG information, which includes: EEG data corresponding to the training sample set and EEG data of the test subject; The processing module is configured to select representative moments from the EEG data corresponding to the training sample set through a representativeness selection mechanism, and extract potential topography at the representative moments; cluster the potential topography and determine the number of microstates through stability evaluation, and generate a microstate template set; the step of determining the number of microstates through stability evaluation and generating the microstate template set includes: For each number of candidate microstates, hierarchical resampling is performed on a sample-by-sample basis within the training sample set to form multiple sets of potential terrain. After performing constrained clustering on each set of potential terrain, the obtained cluster centers are back-projected onto the corresponding potential terrain set. The reprojection error and the matching consistency of the cluster centers are calculated, and a weighted combination is performed with the spatial spectral energy ratio and mirror symmetry of the corresponding cluster centers to determine the stability score. Based on the stability score sequence in the training sample set, calculate the stability score gain of the number of adjacent candidate microstates, and define the preset score gain threshold as the second threshold. The number of microstates that maximizes the stability score and whose score gain is less than a second threshold is selected as the target number of microstates, and a set of microstate templates is generated based on the clustering results corresponding to the target number of microstates. The selection module is used to perform temporal reconfiguration on the EEG data of the test subject using the microstate template set to obtain the temporal feature vector of the test subject; based on the training sample set, a temporal feature matrix is ​​obtained; for the temporal feature matrix, sparse feature selection is performed to determine the target feature index set, and the target feature index set is applied to the temporal feature vector of the test subject to obtain a feature subset of the test subject; The output module is used to construct a classifier with regularization constraints, train a classification model using the training sample set, receive a subset of features of the test object to generate a discrimination score, and output auxiliary diagnostic information to characterize the risk of autism based on the discrimination score.

2. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 1, characterized in that, The extraction of potential topography at the representative time includes: The measurement channels of the EEG data corresponding to the training sample set are divided into a first channel group and a second channel group. The first artifact coefficient is calculated for the channels of the first channel group based on the absolute deviation of the instantaneous amplitude relative to the median of the channel. The second artifact coefficient is calculated for the channels of the second channel group based on the degree of inconsistency of the spatial correlation between the channels. The low artifact factor is calculated by combining the first artifact coefficient and the second artifact coefficient. Based on the low artifact factor, the global field strength quantile at each time point, and the spatial correlation average of the potential topography within the first window length at each time point, the representative score for each time point is calculated. Non-maximum suppression is performed based on the representative scores to determine multiple representative moments, and potential terrain is extracted at the representative moments to form a terrain sample set.

3. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 2, characterized in that, The step of dividing the measurement channels of the EEG data corresponding to the training sample set into a first channel group and a second channel group includes: Based on the EEG data corresponding to the training sample set, the artifact sensitivity index of each measurement channel is calculated, and each measurement channel is divided into a first channel group and a second channel group according to the quantile threshold of the artifact sensitivity index. The first artifact coefficient and the second artifact coefficient are calculated for the first channel group and the second channel group respectively; the contribution of the first artifact coefficient and the second artifact coefficient in the potential topography clustering stability evaluation index within the corresponding channel group is determined as the weight; and a weighted geometric mean is performed on the first artifact coefficient and the second artifact coefficient according to the weight to generate a low artifact factor. Based on the training sample set, the quantile thresholds used for channel segmentation are determined according to the absolute deviation rules of quantiles and medians within the channel group.

4. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 2, characterized in that, The clustering of the potential terrain includes: Based on the spatial layout of the measurement channels of the EEG data corresponding to the training sample set, a channel adjacency graph is established. Polarity unification and mirror mapping are performed on the potential terrain. The spatial spectral energy ratio of each potential terrain on the graph Laplacian eigenvector basis of the channel adjacency graph is calculated. The correlation coefficient between each potential terrain and the mirror potential terrain is calculated as the mirror symmetry. Within a preset set of candidate microstates, constrained clustering is performed on the potential terrain, with the goal of maximizing the spatial correlation within the group, and the constraint that the spatial spectral energy ratio of each cluster center is not greater than a first upper limit and the mirror symmetry is not less than a first lower limit is imposed.

5. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 4, characterized in that, The calculation of the spatial spectral energy ratio of each potential topography on the graph Laplacian eigenvector basis of the channel adjacency graph includes: Based on the normalized graph Laplacian matrix of the channel adjacency graph, the eigenvalue sequence and eigenvector sequence are obtained, and the first spectral threshold is determined according to the maximum spectral gap in the eigenvalue sequence. The ratio of the projected energy of the potential terrain after polarity unification processing on the set of basis vectors with eigenvalues ​​not greater than the first spectral threshold to the projected energy on all basis vectors is determined as the spatial spectral energy ratio.

6. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 4, characterized in that, The calculation of the correlation coefficient between each potential terrain and its mirror potential terrain as the mirror symmetry includes: The midline of the skull is determined based on the spatial layout of the measurement channels of the EEG data corresponding to the training sample set, and the channel correspondence is established based on the mirror image of the coordinates of each measurement channel relative to the midline of the skull to determine the mirror mapping matrix. The weights of the measurement channels are determined by normalizing the distances from each measurement channel to the midline of the skull. The potential terrain vector after polarity unification processing and the potential terrain vector after the mirror mapping matrix transformation are calculated in the space after removing constant components, and the weighted correlation coefficient is used as the mirror symmetry.

7. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 1, characterized in that, The obtained feature subset of the object to be tested includes: The time-matrix feature matrix is ​​divided into multiple feature groups according to microstates. Each feature group includes at least the average duration, time coverage, occurrence frequency, and transition probability vector from the microstate to other microstates, resulting in a grouped time-matrix feature matrix. Multiple training sub-matrices are generated on the grouped temporal feature matrix. The training sub-matrices are obtained by performing hierarchical bootstrap sampling on the sample rows. For each training sub-matrice, a regularization weight is calculated based on the coefficient of variation of each feature in the corresponding training sub-matrice, and a candidate feature index set is calculated using weighted group sparse selection. Stability statistics are performed on multiple candidate feature index sets, the selection frequency of each feature is calculated and compared with a third threshold to determine the target feature index set; The target feature index set is applied to the temporal feature vector of the object under test, and the corresponding components are extracted according to the index order to obtain the feature subset of the object under test.

8. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 4, characterized in that, The step of performing temporal reconfiguration of the EEG data of the test subject using the microstate template set includes: Based on the time domain standard deviation of the EEG data corresponding to the training sample set and the EEG data of the test subject in each measurement channel, the channel gain normalization matrix is ​​determined, and the channel gain normalization matrix is ​​used to perform channel gain normalization processing on the EEG data of the test subject. On the normalized graph Laplacian eigenvector basis of the channel adjacency graph, orthogonal registration is performed on the potential terrain set of the object under test after the channel gain normalization process and the microstate template set to determine a first alignment matrix, and the microstate template set is transformed with the first alignment matrix to obtain an alignment template set; Based on the alignment template set, labels are assigned to each time point of the EEG data of the test subject according to the criterion of maximizing spatial correlation, thereby generating the microstate sequence of the test subject.

9. The intelligent auxiliary diagnostic system for autism based on EEG microstates and XGBoost according to claim 8, characterized in that, The construction of a classifier with regularization constraints includes: The lower and upper limits of the preset risk threshold range are defined as the fourth and fifth thresholds, respectively; On the training sample set, for the risk threshold interval defined by the fourth threshold and the fifth threshold, pre-classification models with different class weight ratios are trained respectively and cross-validation predictions are generated. Based on the cross-validation predictions, the average net benefit of decision curve analysis is calculated within the risk threshold interval to obtain the evaluation value of each class weight ratio. The class weight ratio that maximizes the average net benefit is selected as the target weight ratio, and the classifier with regularization constraints is trained using the training sample set and the target feature index set under the target weight ratio to obtain the classification model.

Images