A method and system for assessing the quality of resting-state norm EEG

Abstract

This invention discloses a method for assessing the quality of resting-state normative EEG signals, comprising the following steps: extracting EEG signal features related to signal quality from three complementary perspectives—time domain, frequency domain, and spatial domain—from the obtained resting-state normative EEG signals; dividing the extracted normative EEG signal feature dataset into training and test sets using K-fold cross-validation; selecting random forest base classifiers and inputting the training sets of normative EEG signal features corresponding to the three perspectives into the three random forest base classifiers for individual training; dynamically allocating the weights of each perspective based on the classification accuracy of the three perspectives in each fold of the K-fold cross-validation; and integrating the classification results based on the predicted labels of each perspective on the test set and the allocated weights to obtain the final resting-state normative EEG signal quality assessment result. A resting-state normative EEG quality assessment system is also disclosed. This invention is the first to employ a multi-perspective integration method for assessing the quality of resting-state normative EEG signals.

Description

Technical Field

[0001] This invention relates to the field of normative EEG signal processing technology, and in particular to a method and system for assessing the quality of resting-state normative EEG. Background Technology

[0002] Resting-mode norm EEG refers to EEG signals acquired from healthy individuals under natural, awake, and closed-eye conditions without a specific task, using a standardized acquisition process. This data forms a population reference baseline based on large-sample statistical modeling. Its core value lies in eliminating non-physiological interference through rigorous quality control, establishing a statistical distribution range of EEG parameters reflecting a specific population, and providing a quantitative assessment basis for individual EEG data deviations. Conventional resting-mode EEG only records the instantaneous state of spontaneous neural activity, lacking population statistical significance; general clinical EEG focuses on pathological waveform detection and does not provide a quantitative reference range for healthy individuals; induced EEG relies on external task paradigms to analyze event-related potentials; brain-computer interface EEG focuses on decoding the characteristics of motor imagery or cognitive tasks, ignoring the physiological frequency domain distribution of the resting state. In contrast, resting-mode norm EEG emphasizes standardized acquisition and large-sample representativeness, aiming to eliminate individual variability and equipment differences, and constructing a reusable neural oscillation baseline. For example, a deviation of the peak frequency of alpha waves in children from the age-matched norm may indicate cognitive developmental delay, while delta power exceeding the age-matched norm in the elderly may predict neurodegenerative diseases. This population-based statistical reference makes resting norm EEG irreplaceable in neuroscience research and clinical applications.

[0003] Most existing automated assessment methods for normative EEG signal quality are based on single-perspective feature analysis. For example, they might divide a normative EEG signal into time windows, calculate statistical features such as peak value, mean, and variance, and then classify the signal quality based on a fixed threshold. Alternatively, they might extract only frequency domain features such as power spectral density and power ratio and assess quality based on a threshold. These methods are not comprehensive enough, potentially leading to increased misclassification rates for complex noise such as EMG artifacts. Furthermore, fixed thresholds result in poor adaptability and reduced reliability of the assessment results.

[0004] It is worth noting that while specialized pipelines for specific age groups have been developed in the international EEG preprocessing technology field—for example, the APICE and HAPPE pipelines are adapted to the high noise and irregular artifacts in infants and young children, the RELAX pipeline optimizes data segmentation strategies for children, and the elderly pipeline combined with PREP / wICA suppresses EMG interference—these segmented solutions fundamentally conflict with the core requirement of constructing unified and comparable EEG norms across the entire life cycle. The high artifact tolerance of the infant pipeline and the strict EMG suppression standards of the elderly pipeline create a disconnect in cross-age data cleaning. Age-locked parameters in children's and infant pipelines cannot be adaptively transferred to other age groups; for instance, filtering parameters that suppress infant motion artifacts weaken physiological low-frequency oscillations in the elderly. Relying on segmented pipelines to process narrow-age data introduces artificial breaks in the norm curves, disrupting the continuous evolution trajectory of neurophysiological indicators. Furthermore, the high interpolation tolerance strategy of the infant pipeline may lead to excessive interpolation, potentially distorting the signal topology and causing non-random quality bias. The existing age-isolated pipeline system cannot guarantee the uniformity and comparability of cross-age data processing, severely restricting the construction of high-quality neurodevelopmental and aging norms. Some studies have used deep learning methods to assess the quality of normative EEG signals, but deep models require a large amount of labeled raw normative EEG data for training, and the risk of overfitting is significant in small sample scenarios. Furthermore, the black-box nature of deep learning models makes it impossible for them to correspond to clear neurophysiological interpretations.

[0005] Therefore, there is an urgent need for a novel norm EEG quality assessment method that can adaptively integrate multidimensional features to fundamentally solve the quality assessment bottleneck caused by individual differences, noise complexity, and insufficient age adaptability, and to provide reliable technical support for the construction of resting norm EEG. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a method and system for evaluating the quality of resting norm EEG based on multi-view ensemble learning when constructing norm tasks for large-scale EEG acquisition across the entire population and throughout the entire life cycle. This method can more comprehensively capture the dynamic characteristics, rhythm distribution, and spatial topological features of norm EEG signals, and effectively solve the problem of misjudgment of EEG data quality in complex noise scenarios.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a method for assessing the quality of resting-state norm EEG, comprising the following steps:

[0008] S1: Extract EEG signal features related to signal quality from the obtained resting norm EEG signals from three complementary perspectives: time domain, frequency domain, and spatial domain.

[0009] S2: K-fold cross-validation was used to divide the extracted norm EEG signal feature dataset into training and test sets. Random forest base classifiers were selected, and the training sets of norm EEG signal features corresponding to the three perspectives were input into the three random forest base classifiers for separate training.

[0010] S3: In each fold of the K-fold cross-validation, the weights of each perspective are dynamically allocated based on the classification accuracy of the three perspectives;

[0011] S4: Based on the predicted labels and assigned weights of the test set from each perspective, the classification results are integrated to obtain the final resting norm EEG signal quality assessment results.

[0012] In a preferred embodiment of the present invention, in step S1, the normative EEG signal features extracted from the time domain perspective include mobility and complexity:

[0013]

[0014] in, It is the original signal The variance; It is the first-order difference of the signal. variance ;

[0015]

[0016] in, It is a second-order difference. .

[0017] In a preferred embodiment of the present invention, in step S1, the normative EEG signal features extracted from the frequency domain perspective include Band power spectral density ratio and parallel logarithmic spectral index :

[0018] The frequency band power spectral density ratio is defined as The ratio of the power spectral density of a frequency band to the power spectral density of the entire frequency band;

[0019]

[0020] in, Represents frequency The cross spectrum matrix at the location, Represents frequency Cross spectrum matrix The eigenvalue diagonal matrix obtained by eigenvalue decomposition has diagonal elements that are eigenvalues ​​arranged in descending order. i.e., frequency The largest eigenvalue at that frequency represents the principal component energy. For all frequencies The trace, i.e., the total frequency domain energy.

[0021] In a preferred embodiment of the present invention, in step S1, the normative EEG signal features extracted from the spatial domain perspective include the bad passage rate (BCR) and the proportion of brain signals after classification by the Independent Component Labeling (ICLabel) algorithm. and Band-directed phase lag index dPLI:

[0022]

[0023] in, It is the number of bad sectors. It is the total number of all electrodes;

[0024]

[0025] in, It is the number of independent components labeled as brain activity. It is the total number of components;

[0026]

[0027] in It is the instantaneous phase difference of each channel of the EEG. , , This represents the total number of sampling points within the calculation period. For time sampling point index, and This includes all channels of the norm EEG signal.

[0028] In a preferred embodiment of the present invention, in each fold of the K-fold cross-validation, the base classifier calculates the independent classification accuracy of each viewpoint on the current test set and dynamically allocates weights according to the formula. The calculation process is as follows:

[0029] make Let represent the test set classification accuracy of the i-th viewpoint base classifier in the k-th fold cross-validation, then its weight is . ,in, satisfy and , Indicates the total number of viewpoints.

[0030] In a preferred embodiment of the present invention, the specific steps of step S4 are as follows:

[0031] For test sample X, define the candidate category set as follows: ,in Indicates the category of low-quality signal. Indicates the category of high-quality signal;

[0032] For the i-th viewpoint base classifier, the predicted label output in the k-th fold cross-validation is: The value can be 0 or 1, and a weight is assigned to it. ,in, The weights represent the prediction results of the i-th viewpoint base classifier in the k-th fold cross-validation.

[0033] Introducing indicator functions Indicator value This indicates that if the i-th base classifier predicts the label... For category If the value is 1, then cast one vote for that category; otherwise, cast 0 votes and the value is 0.

[0034] For each category The weighted voting score is calculated according to the formula. The products of the weights of all view base classifiers and their corresponding indicator values ​​are summed. The weighted voting scores of category 0 and category 1 are compared. The category with the highest score is selected as the ensemble prediction result for sample X in the k-th fold cross-validation. The ensemble prediction result in each fold is then calculated. according to Perform calculations, where Indicates from candidate categories The class with the highest score is selected from the list.

[0035] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is: to provide a resting-state norm EEG quality assessment system, comprising:

[0036] The multi-view norm EEG signal feature extraction module is used to extract EEG signal features related to signal quality from the obtained resting norm EEG signals from several complementary perspectives.

[0037] The multi-view base classifier training module is used to divide the extracted norm EEG signal feature dataset into training and test sets using K-fold cross-validation, select random forest base classifiers, and input the training sets of norm EEG signal features corresponding to the three views into the three random forest base classifiers for separate training.

[0038] The multi-view weight allocation module is used to allocate the weights of each viewpoint based on the classification accuracy of the three views in each fold of the K-fold cross-validation.

[0039] The resting norm EEG quality assessment module is used to obtain the final resting norm EEG signal quality assessment result based on the prediction results of the test set from each perspective and the weight integration classification results assigned by the multi-perspective weight allocation module.

[0040] In a preferred embodiment of the present invention, the multi-view norm EEG signal feature extraction module includes a time-domain view norm EEG signal feature extraction unit, a frequency-domain view norm EEG signal feature extraction unit, and a spatial-domain view norm EEG signal feature extraction unit.

[0041] Furthermore, the EEG signal features extracted by the time-domain perspective norm EEG signal feature extraction unit include mobility and complexity; the norm EEG signal features extracted by the frequency-domain perspective norm EEG signal feature extraction unit include... The frequency band power spectral density ratio and parallel logarithmic spectral index PaLOSi; the normative EEG signal features extracted by the spatial normative perspective EEG signal feature extraction unit include the bad passage rate (BCR) and the proportion of brain signals after classification by the independent component auto-labeling algorithm (ICLabel). and The frequency band directed phase lag index dPLI.

[0042] In a preferred embodiment of the present invention, the method by which the multi-view weight allocation module allocates the weights of each viewpoint is as follows:

[0043] make Let represent the test set classification accuracy of the i-th viewpoint base classifier in the k-th fold cross-validation, then its weight is . ,in, satisfy and , Indicates the total number of viewpoints.

[0044] The beneficial effects of this invention are:

[0045] (1) This invention is the first to use a multi-view integration method to evaluate the quality of resting norm EEG signals and select features from each viewpoint. By integrating features from the time domain, frequency domain and spatial domain, and using a random forest base classifier and dynamic weight allocation strategy, a comprehensive evaluation of the quality of resting norm EEG signals is achieved, which determines whether the resting norm EEG signals can be used for subsequent analysis.

[0046] (2) This invention is specifically designed for quality assessment of resting norm EEG signals. By binding a dynamic weight allocation mechanism with neurophysiological characteristics, it solves the problem of insufficient adaptability of traditional multi-view methods in resting-state scenarios. Compared with traditional single-view assessment methods, this invention can more comprehensively capture the dynamic characteristics, rhythm distribution, and spatial topological features of norm EEG signals, providing an objective and non-destructive technical solution for norm EEG quality assessment, and effectively solving the problem of subjective misjudgment of norm EEG quality in complex noise scenarios;

[0047] (3) This invention demonstrates good generalization ability under small sample conditions, avoids the dependence of deep learning models on large-scale labeled data, and enhances the interpretability of the model. The model using traditional machine learning methods is more lightweight and does not require higher environment configuration. Attached Figure Description

[0048] Figure 1 This is a flowchart of the resting-state norm EEG quality assessment method of the present invention;

[0049] Figure 2 This is a schematic diagram of the weight allocation results from three perspectives;

[0050] Figure 3 This is a structural block diagram of the resting-state norm EEG quality assessment system. Detailed Implementation

[0051] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0052] Please see Figure 1 The embodiments of the present invention include:

[0053] A method for assessing the quality of resting-state norm EEG includes the following steps:

[0054] S1: Extract norm EEG signal features related to signal quality from the obtained resting norm EEG signals from three complementary perspectives: time domain, frequency domain, and spatial domain.

[0055] In this example, the resting norm EEG signal integrates four publicly available resting norm EEG datasets: ① normative EEG, pupillary measurement, electrocardiogram, and photoelectric blood pressure monitoring, as well as behavioral data from digital span tasks and resting periods; ② resting and cognitive state normative EEG test-retest datasets; ③ mind-body-brain-body datasets of MRI, EEG, cognition, emotion, and peripheral physiology in young and older adults; and ④ the Healthy Brain Network open / closed eye dataset. All data were obtained from healthy subjects and underwent standardized preprocessing using Automagic (including 1-35Hz notch filtering and ICA artifact removal), and were labeled with quality tags using Automagic. Data with good and poor quality ratings were selected and defined as high-quality and low-quality data, respectively. Due to sample balance requirements, only low-quality data was selected from the Healthy Brain Network open / closed eye dataset. The study also incorporated high-quality data from 273 expert visual examinations, and used 200 high-quality expert visual examination data as the original data. Data simulation was performed to supplement the missing brain signal data samples through ICA decomposition and removal of low-contribution components based on the component's contribution to the total variance of the signal (PVAF). A total of 1732 normative EEG data points (866 "high-quality" and 866 "low-quality" cases) were ultimately used in the examples.

[0056] The multi-view features extracted in this step are as follows:

[0057] (A) Temporal characteristics: Mobility and Complexity;

[0058] Mobility is a dynamic characteristic index in the Hjorth parameter, used to quantify the rate of change of norm EEG signals and reflect the average frequency characteristics of the signal waveform. Excessively high mobility indicates drastic signal changes, possibly caused by high-frequency noise (such as EMG artifacts, power line interference) or rapid neural oscillations (such as gamma wave activity). Excessively low mobility suggests gradual signal changes, possibly dominated by low-frequency drift (such as baseline fluctuations due to sweating) or resting-state slow-wave activity. Its calculation formula is:

[0059]

[0060] in, It is the original signal The variance; It is the first-order difference of the signal. The variance of the signal characterizes the intensity of its instantaneous changes.

[0061] Complexity estimates the bandwidth of an EEG signal by calculating the mobility of the first derivative of the EEG relative to the EEG itself. High complexity reflects the presence of multi-scale dynamic patterns (such as coordinated activation of multiple brain regions during task-oriented processes) or high-frequency noise (such as EMG artifacts). Low complexity suggests highly regularized signals (such as hypersynchronous discharges during epileptic seizures) or low-frequency drift interference (such as baseline drift caused by sweating). The formula is:

[0062]

[0063] in, First-order difference , Second-order difference .

[0064] (B) Frequency domain characteristics: alpha band power spectral density ratio and parallel logarithmic spectral index PaLOSi;

[0065] High-quality resting norm EEG signals The frequency band should be dominant and concentrated in the occipital lobe region. The power spectral density ratio of the frequency band (8-13Hz) is defined as The ratio of the power spectral density of a frequency band to the power spectral density of the entire frequency band.

[0066] The parallel logarithmic spectral index PaLOSi is used to quantify the impact of preprocessing conditions on subsequent frequency domain analysis. The specific formula is as follows:

[0067]

[0068] in, Represents frequency Cross spectrum matrix at position, Represents frequency Cross spectrum matrix The eigenvalue diagonal matrix obtained by eigenvalue decomposition has diagonal elements that are eigenvalues ​​arranged in descending order. i.e., frequency The largest eigenvalue at that frequency represents the principal component energy. For all frequencies The trace (sum of diagonal elements) is the total frequency domain energy. PaLOSi ranges from [0,1]. A larger value indicates higher homogeneity of the cross-frequency spectral structure, which may correspond to over-preprocessing; a lower value reflects enhanced frequency domain heterogeneity, indicating residual noise.

[0069] (C) Spatial characteristics: Bad passage rate (BCR), percentage of brain signal after classification by independent component autolabeling algorithm (ICLabel). and The frequency band directed phase lag index dPLI;

[0070] Poor-quality data often exhibits a high proportion of bad channels (such as impedance imbalance or hardware failure), requiring interpolation algorithms for restoration. However, interpolation essentially estimates missing signals based on the spatial correlation of adjacent channels. When the bad channel rate is too high, interpolation significantly reduces spatial resolution and introduces artifacts, leading to the loss of activity features in key brain regions and consequently affecting the reliability of functional connectivity or source localization analysis. Therefore, the bad channel rate is introduced as a spatial index, calculated using the following formula:

[0071]

[0072] in, It is the number of bad sectors. It represents the total number of all electrodes.

[0073] The proportion of brain signals is crucial to whether norm EEG signals can be further analyzed; high-quality norm EEG signals should have a high proportion of brain signals.

[0074]

[0075] in, It is the number of independent components labeled as brain activity. It represents the total number of components.

[0076] The band-directed phase lag index (dPLI) indirectly reflects the "purity" of the resting norm EEG signal by quantifying the stability of neural-derived phase synchronization. Abnormal values ​​(too high / too low / spatial disorder) are directly related to quality problems such as noise pollution and over-processing. Its calculation formula is as follows:

[0077]

[0078] in It is the EEG that is preserved after bandpass filtering. After the frequency band components are extracted, the instantaneous phase difference of each channel is obtained by Hilbert transform. , N represents the total number of sampling points within the calculation period. For time sampling point index, and This includes all channels of the norm EEG signal.

[0079] The above features can evaluate the quality of resting norm EEG signals from different perspectives, satisfying the following: 1) Spatial, frequency, and temporal features respectively reflect the spatial distribution, frequency components, and temporal dynamics of resting norm EEG signals, achieving information complementarity; 2) Each feature is directly related to the quality of resting norm EEG; 3) The statistical correlation between features from different perspectives is <0.3, avoiding redundant interference. These features all correspond to a 7-dimensional feature vector for each resting norm EEG signal, ensuring that the multi-perspective ensemble model can integrate information from multiple perspectives and improve the model's performance.

[0080] S2: K-fold cross-validation was used to divide the extracted normative EEG signal feature dataset into training and test sets to ensure the reliability of model evaluation. Random forest base classifiers were selected, and the training sets of normative EEG signal features corresponding to the three perspectives were input into three random forest base classifiers for separate training.

[0081] The parameters of the random forest base classifier described in this example were determined through grid search optimization. The grid search range was set to 50-100 decision trees, with a step size of 10, selecting the parameter combination that yielded the highest average accuracy in 10-fold cross-validation. Specifically, this included:

[0082] The spatial domain model input consists of three features of the spatial domain (bad passage rate, percentage of brain signals after classification by the Independent Component Labeling (ICLabel) algorithm, and... The frequency band has a directed phase lag exponent, and there are 90 decision trees. The training subset is generated through Bootstrap sampling. The node splitting adopts the interaction curvature criterion. The minimum number of leaf nodes is 15, and the maximum depth of the tree is limited. Each tree is randomly selected. Each feature is used to split into nodes (d is the dimension of the current view feature).

[0083] The frequency domain model takes two features from the frequency domain as input. The frequency band power spectral density ratio and parallel logarithmic spectral exponent are used to determine the number of decision trees. Training subsets are generated through bootstrap sampling. Node splitting uses the interaction curvature criterion. The minimum number of leaf nodes is 10, and the maximum depth of the trees is limited. Each tree is randomly selected. Each feature is used to split into nodes (d is the dimension of the current view feature).

[0084] The temporal model takes two temporal features (mobility and complexity) as input, and consists of 80 decision trees. A training subset is generated through Bootstrap sampling. Node splitting uses the interaction curvature criterion, with a minimum of 10 leaf nodes and a limit on the maximum tree depth. Each tree is randomly selected. Each feature is used to split into nodes (d is the dimension of the current view feature).

[0085] S3: In each fold of the K-fold cross-validation, the weights of each perspective are dynamically allocated based on the classification accuracy of the three perspectives;

[0086] In this example, K is set to 10, and the current training set is divided into a training subset (90%) and a validation subset (10%). Three random forest base classifiers are trained using time-domain, frequency-domain, and spatial-domain features, respectively.

[0087] make Let represent the test set classification accuracy of the i-th viewpoint base classifier in the k-th fold cross-validation, then its weight is . ,in, satisfy and , Indicates the total number of viewpoints.

[0088] like Figure 2 As shown, this step can further analyze the contribution weight of each perspective. The frequency domain perspective shows the highest weight, with an average weight of 0.3437, which is significantly higher than that of the time domain and spatial domain. The spatial domain has the lowest weight, indicating that the PaLOSi index has core discriminative value in quality assessment. This result also verifies that the time domain and frequency domain features directly quantify the dynamic anomalies of signal energy distribution and have higher discriminative sensitivity for common noise types.

[0089] S4: Based on the prediction results of the test set from each perspective and the weighted classification results, the final resting norm EEG signal quality assessment result is obtained.

[0090] For test sample X, define the candidate category set as follows: ,in Indicates the category of low-quality signal. Indicates the category of high-quality signal;

[0091] For the i-th viewpoint base classifier, the predicted label output in the k-th fold cross-validation is: The value can be 0 or 1, and a weight is assigned to it. ,in, The weights represent the prediction results of the i-th viewpoint base classifier in the k-th fold cross-validation.

[0092] Introducing indicator functions Indicator value This indicates that if the i-th base classifier predicts the label... For category If the value is 1, then cast one vote for that category; otherwise, cast 0 votes and the value is 0.

[0093] For each category The weighted voting score is calculated according to the formula. The products of the weights of all view base classifiers and their corresponding indicator values ​​are summed. The weighted voting scores of category 0 and category 1 are compared. The category with the highest score is selected as the ensemble prediction result for sample X in the k-th fold cross-validation. The ensemble prediction result in each fold is then calculated. according to Perform calculations, where Indicates from candidate categories The class with the highest score is selected from the list.

[0094] The method used in this invention is compared with the quality classification results of single-view feature analysis and the classification results of directly inputting all features into a random forest classifier without viewpoint distinction. The results are shown in Table 1:

[0095] Table 1. Performance Comparison of Single-View Classification, Full-Feature Direct Classification, and Multi-View Dynamic Weighted Integration Framework

[0096]

[0097] Compared with using only single-viewpoint features for quality classification and directly inputting all features into a random forest classifier for classification without distinguishing between viewpoints, the quality classification method described in this invention has significantly improved performance. In particular, compared with extracting features from a single spatial viewpoint and inputting them into a random forest classifier for quality classification, the classification accuracy is improved by nearly 10%.

[0098] To verify the necessity of employing the synergistic use of spatial, frequency, and time domain perspectives in this invention, the technical effects of combining any two perspectives (spatial-frequency, spatial-temporal, and time-frequency) with the integration of the three perspectives in this invention were compared. The specific results are shown in Table 2.

[0099] Table 2 Comparison of classification performance of normative EEG signal quality under different perspective combinations

[0100]

[0101] As shown in Table 2, the average accuracy of any two-view combination is less than 95.5%, while the present invention improves the average accuracy to 97.69% by weighted integration of three-view features, which is up to 4.5 percentage points higher than the two-view combination (compared to 93.19% for the spatiotemporal view combination). The positive class F1 value and area under the curve are also better. The core reasons for the differences in technical performance are as follows: spatial domain features provide information on neural synchronization stability, frequency domain features provide information on rhythm integrity, and time domain features provide information on waveform dynamics. These three correspond to the three irreplaceable dimensions of norm EEG signal quality: spatial distribution, frequency components, and temporal dynamics. Dual-view combinations suffer from information gaps: for example, while time-frequency combinations can capture frequency and time features, they lack spatial synchronization assessment; frequency-space combinations lack transient noise detection in the time domain waveform (such as electromyographic spikes), making it difficult to distinguish between real neural fluctuations and artifact interference; and time-space combinations lack rhythmic information, failing to detect frequency domain structural heterogeneity caused by excessive preprocessing. Three-view integration achieves feature complementarity through weighted voting. When one view fails due to noise, other views can still provide effective quality criteria. By fusing the discriminative information of the three dimensions through a weighted majority voting mechanism, a significant leap in accuracy is ultimately achieved.

[0102] Figure 2 This paper visually demonstrates the dynamic weight allocation results of the time-domain, frequency-domain, and spatial-domain base classifiers during 10-fold cross-validation. The weights are not fixed but dynamically adjusted based on the classification accuracy of each perspective in each fold of validation. The figure shows that the average weight of the frequency-domain features reaches 0.3437, significantly higher than that of the time-domain and spatial-domain features. This verifies that the frequency-domain features have the highest sensitivity to discriminating complex noise such as electromyographic artifacts, consistent with the strong correlation between rhythm abnormalities and noise in neurophysiology. Furthermore, the PaLOSi index plays a significant role in assessing the absence of brain signals. Although the weights of the time-domain and spatial-domain features are relatively low, dynamic integration compensates for the limitations of a single perspective. Traditional single-perspective or fixed-weight methods lead to misclassification due to neglecting the correlation between the frequency and spatial domains. This invention improves overall performance through dynamic weighted integration, as shown in Table 1, achieving an accuracy of 97.69% after integration.

[0103] The method used in this invention is compared with the classification results of various lightweight deep learning models that directly input all features without distinguishing perspectives. The results are shown in Table 3:

[0104] Table 3 Performance Comparison of Lightweight Deep Learning Model and Multi-View Dynamically Weighted Ensemble Framework

[0105]

[0106] Table 3 compares the performance metrics of the multi-view dynamic weighted ensemble framework with three lightweight deep learning models (DNN, ResNet-1D-Slim, and Transformer-Tiny). The multi-view ensemble outperforms all other deep learning models in average accuracy, positive class F1 score, and area under the curve, showing a significant advantage in speed, exceeding the fastest Transformer-Tiny (0.9 minutes) by more than 5 times. This efficiency advantage stems from the parallelization of the random forest base classifier and the computationally lightweight design of the dynamic weighting mechanism. All compared deep learning models employ a feature engineering-driven lightweight architecture: DNN uses a two-layer fully connected network (64→32 nodes) with dropout rate control to prevent overfitting; ResNet-1D-Slim innovatively transforms image residual blocks into 3×1 one-dimensional convolutions (32 filters), introducing skip connections while preserving temporal local features; Transformer-Tiny achieves parameter simplification by compressing the number of attention heads (4 heads) and the feedforward network dimension (64→32).

[0107] See Figure 3 This invention also provides a resting-state norm EEG quality assessment system, including a multi-view norm EEG signal feature extraction module, a multi-view base classifier training module, a multi-view weight allocation module, and a resting-state norm EEG quality assessment module.

[0108] The multi-view norm EEG signal feature extraction module is used to extract norm EEG signal features related to signal quality from the obtained resting norm EEG signals from several complementary perspectives.

[0109] The multi-view base classifier training module is used to divide the extracted norm EEG signal feature dataset into training and test sets using K-fold cross-validation, select random forest base classifiers, and input the norm EEG signal feature training sets corresponding to the three views into the three random forest base classifiers for separate training.

[0110] The multi-view weight allocation module is used to dynamically allocate the weights of each viewpoint based on the classification accuracy of the three views in each fold of the K-fold cross-validation.

[0111] The resting norm EEG quality assessment module is used to obtain the final resting norm EEG signal quality assessment result based on the prediction results of the test set from each perspective and the weight integration classification result assigned by the multi-perspective weight allocation module.

[0112] The multi-view norm EEG signal feature extraction module includes a time-domain view norm EEG signal feature extraction unit, a frequency-domain view norm EEG signal feature extraction unit, and a spatial-domain view norm EEG signal feature extraction unit.

[0113] Furthermore, the normative EEG signal features extracted by the time-domain perspective normative EEG signal feature extraction unit include mobility and complexity; the normative EEG signal features extracted by the frequency-domain perspective normative EEG signal feature extraction unit include... Band power spectral density ratio and parallel logarithmic spectral index The normative EEG signal features extracted by the spatial perspective normative EEG signal feature extraction unit include the bad passage rate and the proportion of brain signals after BCR independent component automatic labeling algorithm ICLabel classification. and The frequency band directed phase lag index dPLI.

[0114] This example provides a resting-state norm-modal EEG quality assessment system that can execute the resting-state norm-modal EEG quality assessment method provided by this invention. It can perform any combination of the steps in the method example and has the corresponding functions and beneficial effects of the method.

[0115] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for assessing the quality of resting-state norm EEG, characterized in that, Includes the following steps: S1: Extract norm EEG signal features related to signal quality from the obtained resting norm EEG signals from three complementary perspectives: time domain, frequency domain, and spatial domain. S2: K-fold cross-validation was used to divide the extracted norm EEG signal feature dataset into training and test sets. Random forest base classifiers were selected, and the training sets of norm EEG signal features corresponding to the three perspectives were input into the three random forest base classifiers for separate training. S3: In each fold of the K-fold cross-validation, the weights of each perspective are dynamically allocated based on the classification accuracy of the three perspectives; S4: Based on the integrated classification results of the predicted labels and assigned weights of the test set from various perspectives, the final resting norm EEG signal quality assessment results are obtained.

2. The method for assessing the quality of resting-state norm EEG according to claim 1, characterized in that, In step S1, the normative EEG signal features extracted from the time domain perspective include mobility and complexity: ， in, It is the original signal The variance; It is the first-order difference of the signal. variance ; ， in, It is a second-order difference. .

3. The method for assessing the quality of resting-state norm EEG according to claim 1, characterized in that, In step S1, the norm EEG signal features extracted from the frequency domain perspective include Band power spectral density ratio and parallel logarithmic spectral index : The frequency band power spectral density ratio is defined as The ratio of the power spectral density of a frequency band to the power spectral density of the entire frequency band; ， in, Represents frequency The cross spectrum matrix at the location, Represents frequency Cross spectrum matrix The eigenvalue diagonal matrix obtained by eigenvalue decomposition has diagonal elements that are eigenvalues ​​arranged in descending order. i.e., frequency The largest eigenvalue at that frequency represents the principal component energy. For all frequencies The trace, i.e., the total frequency domain energy.

4. The method for assessing the quality of resting-state norm EEG according to claim 1, characterized in that, In step S1, the normative EEG signal features extracted from the spatial domain perspective include the bad passage rate (BCR) and the proportion of brain signals after classification by the Independent Component Labeling (ICLabel) algorithm. and Band-directed phase lag index dPLI: ， in, It is the number of bad sectors. It is the total number of all electrodes; ， in, It is the number of independent components labeled as brain activity. It is the total number of components; ， in It is the instantaneous phase difference of each channel of the EEG. , , This represents the total number of sampling points within the calculation period. For time sampling point index, and This includes all channels of the norm EEG signal.

5. The method for assessing the quality of resting-state norm EEG according to claim 1, characterized in that, In step S3, in each fold of the K-fold cross-validation, the base classifier calculates the independent classification accuracy for each viewpoint on the current test set and dynamically assigns weights according to the formula. The calculation process is as follows: make Let represent the test set classification accuracy of the i-th viewpoint base classifier in the k-th fold cross-validation, then its weight is . ,in, satisfy and , Indicates the total number of viewpoints.

6. The method for assessing the quality of resting-state norm EEG according to claim 1, characterized in that, The specific steps of step S4 are as follows: For test sample X, define the candidate category set as follows: ,in Indicates the category of low-quality signals. Indicates the category of high-quality signal; For the i-th viewpoint base classifier, the predicted label output in the k-th fold cross-validation is: The value can be 0 or 1, and a weight is assigned to it. ,in, The weights represent the prediction results of the i-th viewpoint base classifier in the k-th fold cross-validation. Introducing indicator functions Indicator value This indicates that if the i-th base classifier predicts the label... For category If the value is 1, then cast one vote for that category; otherwise, cast 0 votes and the value is 0. For each category The weighted voting score is calculated according to the formula. The weights of all view base classifiers are multiplied by their corresponding indicator values ​​and summed. The weighted voting scores of category 0 and category 1 are compared. The category with the highest score is selected as the ensemble prediction result of sample X in the k-th fold cross-validation.

7. A resting-state norm EEG quality assessment system, characterized in that, include: A multi-view normative EEG signal feature extraction module is used to extract EEG signal features related to signal quality from the obtained resting normative EEG signals from three complementary perspectives: time domain, frequency domain, and spatial domain. The multi-view normative EEG signal feature extraction module includes a time domain perspective normative EEG signal feature extraction unit, a frequency domain perspective normative EEG signal feature extraction unit, and a spatial domain perspective normative EEG signal feature extraction unit. The multi-view base classifier training module is used to divide the extracted norm EEG signal feature dataset into training and test sets using K-fold cross-validation, select random forest base classifiers, and input the training sets of norm EEG signal features corresponding to the three views into the three random forest base classifiers for separate training. The multi-view weight allocation module is used to allocate the weights of each viewpoint based on the classification accuracy of the three views in each fold of the K-fold cross-validation. The resting norm EEG quality assessment module is used to obtain the final resting norm EEG signal quality assessment result based on the prediction results of the test set from each perspective and the weight integration classification results assigned by the multi-perspective weight allocation module.

8. The resting-state norm EEG quality assessment system according to claim 7, characterized in that, The EEG signal features extracted by the time-domain perspective normative EEG signal feature extraction unit include mobility and complexity; the normative EEG signal features extracted by the frequency-domain perspective normative EEG signal feature extraction unit include... Band power spectral density ratio and parallel logarithmic spectral index The normative EEG signal features extracted by the spatial normative perspective EEG signal feature extraction unit include the bad passage rate (BCR) and the proportion of brain signals after classification by the independent component labeling algorithm (ICLabel). and The frequency band directed phase lag index dPLI.

9. The resting-state norm EEG quality assessment system according to claim 7, characterized in that, The method by which the multi-view weight allocation module allocates weights for each viewpoint is as follows: make Let represent the test set classification accuracy of the i-th viewpoint base classifier in the k-th fold cross-validation, then its weight is . ,in, satisfy and , Indicates the total number of viewpoints.

Images