A brain state continuous quantitative evaluation method and device based on brain network characteristics, equipment and medium

Abstract

The present application relates to medical instrument technology, equipment and medical special software, and provides a brain state continuous quantitative evaluation method, device, equipment and medium based on brain network characteristics. In view of the problem that the existing brain state evaluation depends on discrete labels and has poor generalization ability, the present application extracts the electroencephalogram time-frequency characteristics and brain network topological characteristics of the subject, and filters out the inter-group characteristics and individual common characteristics through statistical analysis; the baseline state data is used to train and freeze the autoencoder; the two types of characteristics of the target state data are respectively input into the autoencoder, the cross-state reconstruction error is compared, the branch with smaller error is locked as the target evaluation path, and the reconstruction error value is output as a continuous quantitative index. The present application filters out the physiological noise interference, directly quantifies the topological deviation distance by the reconstruction error, and provides a highly interpretable continuous evaluation scale for clinical use.

Description

Technical Field

[0001] This invention relates to medical instrument technology, equipment and medical software, and in particular to a method, device, equipment and medium for continuous quantitative assessment of brain state based on brain network characteristics. Background Technology

[0002] Stroke, especially cerebral infarction, is a cerebrovascular disease with high incidence, high disability rate, and high mortality rate, and has become a major public health problem threatening human health. Before patients are discharged from the hospital and enter the rehabilitation stage, it is usually necessary to conduct an objective assessment of their neurological function recovery to guide the development of subsequent rehabilitation training programs and predict long-term prognosis. Traditional neurological function assessment mainly relies on clinical scales. Such scales are often subjective and cannot directly reflect the degree of recovery of brain network function from a neurophysiological perspective. Therefore, using objective physiological signals to quantitatively assess the state of neurological recovery has gradually become a research direction in related fields.

[0003] Electroencephalography (EEG) is a non-invasive neurophysiological technique that records the electrical activity of brain neurons using scalp electrodes. It features high temporal resolution, low equipment cost, and the ability to perform long-term dynamic monitoring. Studies have shown that the brain activity patterns of stroke patients undergo significant changes during recovery. EEG signal analysis typically extracts two types of features: time-frequency features reflecting signal energy distribution, such as power in each frequency band, δ-α ratio, θ-β ratio, and spectral asymmetry index; and topological features characterizing the functional coupling relationships between brain regions, such as functional connectivity networks constructed based on weighted phase lag indices and graph theory metrics extracted from them, such as local efficiency, global efficiency, and clustering coefficients. With the development of machine learning and deep learning technologies, researchers have begun to use automatic feature learning models such as autoencoders to model EEG data, using the error between input and reconstruction to reflect differences in data structure.

[0004] In some existing brain state assessment schemes, various brain features extracted from patients at different times are often directly spliced ​​together indiscriminately to construct and train supervised deep classifiers, such as support vector machines or graph convolutional neural networks, for end-to-end binary classification to determine whether a patient is in the onset or recovery phase. However, such direct classification methods can only establish classification boundaries and output discrete labels in a high-dimensional feature space, failing to provide continuous quantitative indicators reflecting the evolution of brain state. They are also difficult to characterize the specific degree to which the brain's topological structure deviates from the original pathological state, and their assessment process is black-box and has poor interpretability. In clinical practice, doctors not only need to determine whether a patient has recovered but also need to understand the degree of recovery. Existing methods are unable to meet this need due to the lack of continuous and interpretable quantitative output. At the same time, due to significant individual differences in EEG signals, the above methods are prone to overfitting to individual background physiological noise when directly mixing a large number of original or primary fusion features. This causes the model to overlearn the differences in individual basic brain rhythms while ignoring the brain network reorganization patterns related to disease evolution, resulting in a significant decrease in generalization ability when facing new patients.

[0005] Therefore, it is evident that how to overcome the limitations of discrete classification labels in the assessment of brain state in cerebral infarction and provide continuous and interpretable quantitative indicators, while simultaneously suppressing physiological noise introduced by individual differences in EEG to improve cross-individual generalization ability, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] In view of the technical problems that existing brain state assessment technologies rely heavily on subjective scales, and that existing deep learning direct classification methods can only output discrete labels, cannot continuously quantify the degree of brain state recovery and have poor interpretability, and are prone to overfitting to individual background physiological noise, resulting in a significant decrease in generalization ability, the present invention aims to provide a method, device, equipment and medium for continuous quantitative assessment of brain state based on brain network features, so as to solve the above problems.

[0007] This invention provides a method for continuous quantitative assessment of brain states based on brain network features, comprising: acquiring continuous EEG time-series signals of a subject and preprocessing them, dividing them into continuous EEG signal segments of fixed time length; within each continuous EEG signal segment, extracting the subject's time-frequency features and brain network topology features in parallel; performing inter-group statistical analysis and individual paired statistical analysis on the extracted time-frequency features and brain network topology features, screening out inter-group features that have significant group differences between preset first different brain states, and individual common features that have significant differences in each individual; and training a self-test using pre-determined baseline state data. The encoder minimizes the same-state reconstruction error and freezes the network weights of the autoencoder after training convergence. The inter-group features and individual common features corresponding to the target state data are fed as two independent inputs into the autoencoder with frozen weights, and the cross-state reconstruction error of each input is calculated independently. The magnitudes of the cross-state reconstruction errors of the two inputs are compared, and the feature branch corresponding to the path with the smaller cross-state reconstruction error is locked as the target evaluation path. The cross-state reconstruction error value of the target state data on the target evaluation path is output as a continuous quantitative indicator representing the deviation distance of the brain network topology. Using the above technical solution, this invention establishes the most manifold-stable inter-group brain network topology features as the core input through a system-level self-verification mechanism. This effectively filters out time-frequency features sensitive to individual background physiological noise and transforms the brain state discrimination problem, which originally relied on discrete classification labels, into a continuous distance measurement problem characterized by the magnitude of reconstruction error. This allows doctors to further grasp the true recovery trajectory and recovery amount from onset to discharge, overcoming the shortcomings of poor interpretability and weak cross-patient generalization ability of direct classification methods.

[0008] As a preferred embodiment of the present invention, the acquisition and preprocessing of the subject's continuous EEG temporal signal, dividing it into continuous EEG signal segments of fixed time length, includes: synchronously recording the subject's continuous EEG temporal signal through multiple electrode channels; bandpass filtering the continuous EEG temporal signal within a frequency range of 1 to 30 Hz; detecting and removing bad segments from the filtered EEG signal, and using independent component analysis to identify and remove artifacts from electrooculography (EOG) and electromyography (EMG); and windowing and segmenting the artifact-free EEG signal according to a preset fixed time length to obtain the continuous EEG signal segments. Using the above preferred embodiment, low-frequency drift and high-frequency noise can be suppressed to the greatest extent, eliminating strong interference physiological signals such as body movement, EOG, and EMG, providing a clean and standardized input unit on a uniform scale for downstream multi-feature calculation and extraction, thus solidifying the data foundation for quantitative assessment.

[0009] As a preferred embodiment of the present invention, the step of extracting the time-frequency features and brain network topology features of the subject in parallel within each segment of the continuous EEG signal includes: calculating the power spectral density of the signal from each electrode channel in a preset frequency band using a spectral estimation method, and accordingly calculating the band power, relative power, power ratio between specific frequency bands, and peak value of each preset frequency band. Frequency and the left-right brain region asymmetry index are used as the time-frequency features. Based on the imaginary part of the cross spectrum between the frequency domain representations of different electrode channels, the weighted phase lag index between each pair of electrode channels is calculated to construct a functional connectivity matrix. The functional connectivity matrix is ​​then subjected to a proportional threshold based on a preset quantile and binarized to generate a sparse brain functional connectivity graph structure corresponding to a binary adjacency matrix. On the sparse brain functional connectivity graph structure, the local efficiency of nodes and the regional aggregation value of their corresponding brain regions of interest are calculated as the topological features of the brain network. Using the above preferred technical solution, not only are time-frequency features reflecting energy distribution extracted from the perspective of signal fluctuation, but also a complex network characterizing real functional coupling is constructed from the perspective of brain-inter-brain interactions, and graph theory topological features are extracted. This provides comprehensive high-dimensional feature matrix support for multi-perspective feature comparison and verification in the downstream evaluation stage.

[0010] As a preferred technical solution of the present invention, the peak value The frequency calculation process is as follows: in the range of 8 to 13 Hz... Within the frequency band, the centroid method is used to calculate the first... Peak value of each electrode channel frequency:

[0011] In the formula, Indicates the first Peak value of each electrode channel frequency, Indicates the first Each channel at frequency Power spectral density at that point and They are respectively The lower and upper limits of the frequency band. For frequency variables; and / or, the formula for calculating the weighted phase lag index is:

[0012] In the formula, Indicates the first The electrode channel and the first Weighted phase lag index between electrode channels Indicates the first The electrode channel and the first The imaginary part of the cross spectrum between individual electrode channels This indicates the expected or time-averaged calculation. The sign function is defined as follows. Utilizing the aforementioned optimized technical scheme, the energy-weighted extraction of the spectral centroid improves noise resistance compared to directly extracting spectral peaks. Furthermore, by eliminating zero-phase coupling through the imaginary part of the cross-spectrum, the interference from pseudo-connections caused by volume conduction and common reference is significantly suppressed, ensuring the objective authenticity of the extracted brain network topological features.

[0013] As a preferred technical solution of the present invention, the step of performing inter-group statistical analysis and individual paired statistical analysis on the extracted time-frequency features and brain network topology features to screen out inter-group features and individual common features includes: obtaining two sets of sample data corresponding to the subjects in the first different brain states; comparing the distribution differences of each feature between the two sets of samples through independent sample tests or non-parametric tests; and aggregating features whose significance level meets a preset threshold as the inter-group features; performing paired statistical tests on the corresponding features collected from the same subject in the first different brain states to obtain features that show significant changes in a single subject; and taking the intersection of features that show significant changes in all subjects participating in the screening, and extracting features that show significant changes in all individuals as the individual common features. Using the above preferred technical solution, by separating the two macro- and micro-statistical levels of the group and the individual, the inter-group features that are sensitive to brain neural network reorganization at the group scale and the individual common features that are extremely sensitive to individual background physiology are accurately identified, providing feature comparison targets for subsequent elimination of individual differences.

[0014] As a preferred embodiment of the present invention, the first different brain states include the baseline state of the stroke patient and the evolving recovery state; the combination of the baseline state data and the target state data includes any of the following: using the baseline state data as the baseline state data to train the autoencoder, and using the evolving recovery state data as the target state data, wherein the value of the continuous quantification index positively represents the degree of deviation of the brain network topology from the baseline state; using the evolving recovery state data as the baseline state data to train the autoencoder, and using the baseline state data as the target state data; using the physiological state data of a healthy control individual as the baseline state data to train the autoencoder, and using the stroke patient data as the target state data, wherein the value of the continuous quantification index negatively represents the degree of convergence of the brain network topology relative to the healthy state. Utilizing the above preferred embodiment, the clinical applicability of the assessment method is greatly expanded, enabling flexible support for multi-perspective cross-validation, and achieving continuous quantitative assessment of the subject's brain function from different spatiotemporal dimensions such as the degree of deviation from the initial state and the degree of convergence to the final recovery state.

[0015] As a preferred embodiment of the present invention, the step of training an autoencoder using predetermined baseline state data to minimize the same-state reconstruction error includes: using the autoencoder, which is composed of an encoder and a decoder cascaded, to map the input feature vector in the baseline state to a low-dimensional latent representation and perform reconstruction output; based on the formula

[0016] The reconstruction error is calculated, and the weights of the autoencoder are updated with the goal of minimizing the reconstruction error; wherein... For reconstruction error, For the first One input feature vector, For the corresponding reconstruction output, The sample size is defined as follows: After comparison and locking, the feature branch corresponding to the path with the smaller cross-state reconstruction error is the branch containing the inter-group feature. The inter-group feature is then statistically mapped and established as the brain network topological feature. Utilizing the above-mentioned preferred technical solution, the manifold learning capability of the autoencoder under unsupervised conditions is fully leveraged. Without relying on human experience intervention, the feature least prone to manifold collapse can be identified as the optimal evaluation index through purely data-driven cross-state reconstruction calculation, thus improving the overall robustness of the system.

[0017] This invention provides a continuous quantitative assessment device for brain states based on brain network features, comprising: a preprocessing unit for acquiring and preprocessing continuous EEG time-series signals of a subject, dividing them into continuous EEG signal segments of fixed time lengths; a feature extraction unit for extracting, in parallel, the subject's time-frequency features and brain network topology features within each of the continuous EEG signal segments; a statistical screening unit for performing inter-group statistical analysis and individual paired statistical analysis on the extracted time-frequency features and brain network topology features, screening out inter-group features that exhibit significant group differences between preset first different brain states, and individual common features that exhibit significant differences in each individual; and a reconstruction assessment unit for utilizing preprocessing data to assess the brain state. An autoencoder is trained using pre-determined baseline state data to minimize in-state reconstruction error, and the network weights of the autoencoder are frozen after training convergence. The inter-group features and individual common features corresponding to the target state data are fed into the autoencoder with frozen weights as two independent inputs, and the cross-state reconstruction error of each input is calculated independently. The magnitudes of the cross-state reconstruction errors of the two inputs are compared, and the feature branch corresponding to the one with the smaller cross-state reconstruction error is locked as the target evaluation path. A quantization output unit is used to output the cross-state reconstruction error value of the target state data on the target evaluation path as a continuous quantization index characterizing the topological deviation distance of the brain network.

[0018] The present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for continuous quantitative assessment of brain states based on brain network features.

[0019] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for continuous quantitative assessment of brain states based on brain network features.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention establishes the most manifold-stable inter-group brain network topology features as the core evaluation input through a system-level self-verification mechanism. This effectively filters out time-frequency features sensitive to individual background physiological noise and transforms the brain state discrimination problem, which originally relied on discrete classification labels, into a continuous distance metric problem characterized by the magnitude of reconstruction error. This allows doctors to further grasp the true recovery trajectory and recovery amount from onset to discharge, beyond simply judging whether a patient has recovered. This overcomes the shortcomings of direct classification methods, such as poor interpretability and weak cross-patient generalization. By separating the macro- and micro-statistical levels of the population and individuals, it accurately identifies inter-group features sensitive to brain neural network reorganization at the population scale and individual common features extremely sensitive to individual background physiology, providing feature comparison targets for subsequent elimination of individual differences. The autoencoder is trained only on baseline state data and has its weights frozen. By comparing the cross-state reconstruction errors of the two feature branches, it automatically locks the optimal evaluation path without human intervention, improving the overall robustness of the system. Furthermore, the combination of baseline and target states proposed in this invention has good flexibility and can support multiple cross-validation scenarios such as disease baseline, recovery end state, and healthy control. It can achieve continuous quantitative assessment of the degree of brain function recovery of subjects from different spatiotemporal dimensions, and has strong clinical applicability and promotion value. Attached Figure Description

[0021] Figure 1 The overall flowchart of the continuous quantitative assessment method for brain state based on brain network features provided in the embodiments of the present invention; Figure 2 A schematic diagram of the structure of a brain state continuous quantitative assessment device based on brain network features provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0022] Explanation of reference numerals in the attached figures: 100, Data acquisition unit; 200, Preprocessing unit; 300, Feature extraction unit; 310, Time-frequency feature extraction module; 320, Brain network feature extraction module; 400, Statistical screening unit; 410, Inter-group statistics module; 420, Individual pairing statistics module; 500, Reconstruction evaluation unit; 510, Encoder training module; 520, Cross-reconstruction module; 530, Path locking module; 600, Quantization output unit; 700, Electronic device; 701, Processor; 702, Memory; 703, Communication bus; 704, Communication interface. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention. Unless otherwise defined, the following embodiments and technical features can be combined with each other. Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the invention.

[0024] This invention provides a method for continuous quantitative assessment of brain states based on brain network features, the overall processing flow of which is as follows: Figure 1 As shown, the process generally includes several interconnected steps: acquisition and preprocessing of EEG data, parallel extraction of multiple EEG features, feature selection based on both population and individual statistical levels, autoencoder training for specific baseline states, and assessment of brain state deviation using cross-state reconstruction error as a continuous quantitative indicator. It should be noted that the brain states described in this invention include, but are not limited to, the onset or baseline state of stroke patients before discharge, their evolution or recovery state after discharge, and the physiological state of healthy controls. The core of this invention lies in leveraging the manifold stability of brain network features at the population level to transform the brain state discrimination problem, which originally relied on discrete labels, into a continuous distance metric problem characterized by the magnitude of reconstruction error, thereby providing interpretable and quantifiable assessment data for brain function evolution in clinical practice.

[0025] During the data acquisition phase, the temporal signals of EEG from multiple electrode channels of the subject were simultaneously recorded using a scalp EEG acquisition device. Considering that the amplitude of the EEG signal itself is weak and easily affected by factors such as power frequency interference, eye movement, electromyography, and poor electrode contact, the raw EEG signal needs to be preprocessed before entering feature calculation. Specifically, all EEG data are first bandpass filtered in the range of 1 to 30 Hz. This frequency range covers the main rhythmic components closely related to brain function, including delta waves, theta waves, alpha waves, and beta waves, thereby suppressing low-frequency drift and high-frequency noise while retaining effective neurophysiological information. Subsequently, the filtered signal is subjected to bad segment detection and removal, deleting abnormal segments caused by subject movement, electrode detachment, or transient strong interference. Furthermore, independent component analysis is used to perform blind source separation of the signal, identifying and removing artifacts such as electrooculography and electromyography, to obtain a relatively pure EEG signal. After completing the above purification process, the continuous EEG signal is windowed and segmented according to a pre-set fixed time length to ensure that each segment is consistent in duration, thereby providing a standardized input unit for subsequent extraction of frequency domain features and connectivity features on a uniform scale.

[0026] In the feature extraction stage, this invention extracts two types of features with different properties in parallel from the EEG signals within each time window: one type is time-frequency features reflecting the signal energy distribution, and the other type is brain network topological features characterizing the functional coupling relationships between brain regions. Regarding time-frequency features, frequency domain analysis is first performed on the EEG signals at each electrode channel, and the power spectral density function of each channel is obtained using the Welch method or other spectral estimation methods. For the first... The power of each electrode channel within a specified frequency band can be expressed as: (1) In the formula, Indicates the first The power of each channel within a specified frequency band Indicates the first Each channel at frequency Power spectral density at that point and These represent the lower and upper frequency limits of the frequency band under investigation, respectively. The frequency variable is used. In this embodiment, the boundaries of each frequency band are defined as follows: δ band (0.5 to 4 Hz), θ band (4 to 8 Hz), α band (8 to 13 Hz), and β band (13 to 30 Hz). After calculating the power of each frequency band separately, they are added together to obtain the total power. The relative power is obtained by the ratio of the power of each frequency band to the total power, for example, the relative power of the δ frequency band. The relative powers of the θ, α, and β frequency bands can be deduced similarly. Based on this, ratio indicators characterizing brain injury features are further calculated, including the δ-α ratio. and the θ-β ratio This reflects the trend of relatively increased low-frequency activity or relatively decreased high-frequency activity.

[0027] To characterize the frequency centroid of brainwave rhythms, this invention further calculates the peak α frequency in the α band. Using only spectral data from 8 to 13 Hz, the centroid method is employed to calculate the peak α frequency: (2) In the formula, Indicates the first Peak α frequency of each electrode channel Indicates the first Each channel at frequency Power spectral density at that point and These are the lower and upper limits of the α band, respectively. The frequency variable is used. This index provides the center frequency of the alpha rhythm using a spectral energy weighted method, which has better noise resistance than simply taking the spectral peak position. Considering that cerebral infarction often involves one hemisphere, this invention also calculates an asymmetry index based on the corresponding characteristics of the left and right symmetrical brain regions to quantify the degree of imbalance between the two hemispheres: (3) In the formula, and These represent the values ​​of the left and right hemispheres for the same feature, respectively. The asymmetry index can be applied to frequency band power or indicators such as peak alpha frequency, thereby obtaining a multi-dimensional description of hemispheric imbalance.

[0028] Regarding brain network topological features, this invention constructs a brain functional connectivity network from the perspective of functional coupling between brain regions within a time window and extracts its topological indices. Assume that the electroencephalogram (EEG) signal contains... Each electrode channel has a time-series signal denoted as: (4) In the formula, Indicates the first The EEG signals of each electrode channel are analyzed. To measure the phase synchronization relationship between different channels, this invention employs a weighted phase lag index. First, the phase lag index is calculated based on the frequency domain representation of each channel signal. The first channel and the first Cross spectrum between channels: (5) In the formula, For the first Frequency domain representation of the signal in each channel For the first The conjugate complex number represented in the frequency domain of each channel. Since the zero-phase coupling introduced by volumetric conduction is mainly reflected in the real part of the cross spectrum, while the true phase lag information is contained in the imaginary part of the cross spectrum, this invention calculates the weighted phase lag exponent based on the imaginary part of the cross spectrum: (6) In the formula, Represents the imaginary part of the cross spectrum. This indicates the expected or time-averaged calculation. The sign function is used. The weighted phase lag exponent, weighted by the imaginary part amplitude of the cross spectrum, can largely suppress spurious connections caused by volumetric conduction and common reference, thus more reliably characterizing the true functional coupling between brain regions. By calculating the above index pairwise for all electrode channels, a [value] can be obtained. Functional connection matrix: (7) In the formula, Indicates the first The electrode channel and the first The functional connectivity strength between individual electrode channels is represented by zeros on the diagonal of the matrix. This functional connectivity matrix is ​​used to characterize the functional coupling relationships between different brain regions as a whole, and serves as the basis for further construction of sparse brain functional networks and extraction of topological features.

[0029] Considering that fully connected brain networks contain a large number of weak connections, which are often highly susceptible to noise and lack stable physiological significance, this invention performs proportional thresholding on the functional connectivity matrix, retaining only the connections with the highest strength. The function connects edges. Specifically, first, all connection weights in the connection matrix are extracted, and their i-th weights are calculated. quantiles as thresholds Then, the connections are binarized based on this threshold to obtain a binary adjacency matrix: (8) In the formula, These are elements of a binary adjacency matrix. The threshold value is used as the scaling factor. This yields the sparsified brain functional connectivity graph structure. Based on this graph structure, this invention uses node local efficiency as the fundamental topological metric. Node The local connectivity efficiency of a node is defined as the global efficiency of the subgraph formed by its neighboring nodes. Let node... The set of neighbor nodes is When the number of neighboring nodes is less than 2, the local efficiency of that node is specified. When the number of neighboring nodes is not less than 2, the global efficiency is calculated on the subgraph formed by its neighboring nodes: (9) In the formula, Representing neighboring nodes and The shortest path length between them. This represents the number of neighboring nodes. Local connectivity efficiency reflects the fault tolerance and integration capability of information transmission within a local area of ​​a node. To obtain a topological description at the brain region level, this invention further performs region aggregation on the local efficiency of nodes based on predefined brain regions of interest. Let the set of electrode channels contained in a certain brain region be denoted as . Then the local connectivity efficiency of this brain region is: (10) In the formula, This refers to the number of electrode channels contained in the brain region. In addition to the local efficiency of nodes and their regional aggregation values, those skilled in the art can also extract other graph theory topology indicators such as global efficiency, clustering coefficient, and feature path length from the above-mentioned brain functional connectivity map without departing from the concept of the present invention, as a supplement to the characteristics of the brain network, so as to more comprehensively characterize the information transmission efficiency between brain regions and the neural network reorganization process.

[0030] After extracting the aforementioned time-frequency features and brain network features, this invention performs two levels of statistical analysis and screening on the extracted features to separate brain network features with stable differences at the group level and time-frequency features that are generally significant at the individual level. The first level is the inter-group statistical analysis before and after discharge. Let the sample set of a certain feature in the pre-discharge group be... The sample set in the post-discharge group is Through independent samples If the normality assumption is not met, an appropriate non-parametric test method can be used to compare whether there is a significant difference between the means or distributions of the two groups of features, thereby obtaining the significance level of each feature. When satisfied At that time, among them A pre-defined significance level threshold is used, meaning that the feature is considered to have a significant difference at the group level before and after discharge. All features that meet this condition are grouped together and defined as between-group features. Since functional connectivity and topological indices between brain regions are highly sensitive to brain network reorganization at the group scale, most between-group features statistically map to brain network topological features.

[0031] The second level is paired individual statistical analysis, used to assess trends in characteristic changes at the individual level. Characteristics collected from the same subject before and after discharge are denoted as follows: and Paired statistical tests were used to analyze the differences between the two groups, thereby obtaining the features that showed significant changes at the individual level for each subject. After performing the above analysis on all subjects, the intersection of the significant feature results for each subject was taken, that is, the features that showed significant changes in every individual were selected and defined as individual common features. Since there are significant individual differences in EEG among different subjects, features that are significant in all individuals often concentrate on time-frequency features that are more sensitive to the individual's background rhythm. Therefore, individual common features are statistically mostly mapped to time-frequency features. The inter-group features and individual common features selected at the above two levels constitute two candidate inputs for the subsequent brain state coding and reconstruction assessment stages.

[0032] In the encoder training phase, this invention utilizes the characteristic of autoencoders to learn data manifolds under unsupervised conditions to construct a benchmark for measuring the deviation of brain states. The autoencoder consists of a cascaded encoder and decoder. The encoder maps the input feature vector to a low-dimensional latent representation, and the decoder then reconstructs an output with the same dimension as the input from this latent representation. Its training objective is to minimize the mean squared error between the input and the reconstructed output. For samples in the training set, the reconstruction error can be expressed as: (11) In the formula, For the first One input feature vector, For the corresponding reconstruction output, The number of samples is denoted by . A fully trained autoencoder can only achieve low-error reconstruction of data distributed in the same way as its training set, while generating significantly increased reconstruction errors for data distributed differently from the training set. This mathematical characteristic is the basis for this invention to measure the deviation of brain states. Based on this, this invention selects specific baseline state data as the training set according to the combined needs of clinical assessment, such as using only pre-discharge data from stroke patients or only data from healthy control individuals. The goal is to minimize the reconstruction error within the same state, and after training convergence, the network weights are frozen, thus solidifying the autoencoder into a scale representing the manifold of specific baseline state data.

[0033] During the testing phase, the target state data is input into an autoencoder with frozen weights via two input paths: between-group features and individual common features. The reconstruction error of each input path is calculated independently. By comparing the magnitudes of the two reconstruction errors, the feature corresponding to the path with the smaller reconstruction error is identified as the more stable feature in cross-state mapping. Based on this, the branch containing this feature is identified as the target evaluation path and serves as the core input variable for subsequent distance calculation. Through this system-level self-verification mechanism, this invention can confirm at the algorithmic level that brain network features have lower cross-state reconstruction errors than time-frequency features without relying on manual prior knowledge. This establishes the most manifold-stable brain network features as the core input for continuous quantitative assessment of brain states. Simultaneously, time-frequency features that are sensitive to individual background physiological noise and unstable in cross-state mapping are filtered out to avoid introducing high-frequency noise and manifold collapse, which would weaken the physical accuracy of the reconstructed distance.

[0034] After identifying the target assessment path, this invention directly uses the cross-state reconstruction error value of the target state data along that path as an objective quantitative indicator characterizing the deviation distance of the brain network topology. Unlike traditional supervised classifiers that can only output discrete labels such as the onset period or recovery period, this reconstruction error is a continuous value, and its magnitude dynamically reflects the degree of deviation or convergence of the subject's brain network topology relative to a given baseline state. When the autoencoder is trained using pre-discharge data as a baseline, the larger the reconstruction error of the post-discharge data, the farther the brain network topology deviates from the onset baseline, and the higher the degree of recovery. When the autoencoder is trained using healthy control data as a baseline, the smaller the reconstruction error of the patient data, the closer the brain network topology is to a healthy state. Thus, doctors can not only obtain a judgment on whether the patient has recovered, but also intuitively grasp the patient's true recovery trajectory and recovery amount from onset to discharge based on the continuously changing reconstruction error value, thereby obtaining a highly interpretable and continuously quantifiable benchmark for brain function rehabilitation assessment.

[0035] It should be noted that the setting of the baseline and target states described in this invention is flexible and can be modified in various equivalent ways according to specific clinical needs. In addition to the typical combination of training with pre-discharge data and testing with post-discharge data, a reverse combination of training with post-discharge data and testing with pre-discharge data can also be used, or a combination of training with healthy control data and testing with patient data can be used to continuously quantify brain state evolution from different perspectives such as the degree of convergence to the final recovery state and the deviation from the disease baseline. The above-mentioned baseline combinations can be used individually or in combination to achieve multi-perspective cross-validation, but their core remains the same: using the continuous error value obtained by reconstructing the brain network characteristics between groups through cross-state autoencoding as a distance metric for brain state quantification. Therefore, all of these are natural extensions of the technical concept of this invention and fall within the protection scope of this invention.

[0036] To verify the feasibility and effectiveness of this invention, the inventors collected multi-channel EEG time-series data of 24 stroke patients, including their baseline state before discharge and their post-discharge recovery status, as well as EEG data from healthy controls. Following the aforementioned preprocessing and feature extraction process, a multimodal feature matrix containing both spatial topological connectivity and time-frequency energy fluctuations of the brain network was obtained. Furthermore, the two-level feature selection scheme of this invention was strictly followed to obtain inter-group statistical features and individual common significant features. Based on this, unsupervised training and weight freezing of the autoencoder were completed using the pre-discharge data as the training set. During the testing phase, the two features from the post-discharge data were input into the frozen model to calculate the cross-state reconstruction error. Simultaneously, reverse cross-validation was performed using various baseline combinations, such as training with post-discharge data and testing with pre-discharge data, and training with healthy control data and testing with patient data. Experimental results show that when the input is a group statistical feature that, after system analysis, is mostly mapped to brain network features, the model's cross-reconstruction error in cross-state testing is extremely small, stabilizing at the order of 0.1. However, when the input is a group common significant feature that, after system analysis, is mostly mapped to time-frequency features, the model exhibits significant reconstruction difficulties, with errors reaching the order of 0.3. This objective data difference of 0.1 (less than 0.3) verifies the core judgment of this invention from the algorithm's underlying perspective: the time-frequency features extracted based on the intersection of individual differences contain a large amount of physiological high-frequency background noise that is difficult to map across states, leading to a significant increase in reconstruction error. In contrast, the brain network features selected through group inter-group statistics retain a highly stable underlying physical topology of the brain and have higher manifold stability, resulting in extremely small reconstruction errors. This demonstrates that using the continuous reconstruction error value of the preferred features of this invention as the topological deviation distance can continuously and objectively quantify the brain state evolution of stroke patients before and after discharge, possessing high clinical assessment benchmark value and engineering feasibility.

[0037] Based on the above-described methodological concept, embodiments of the present invention also provide a continuous quantitative assessment device for brain states based on brain network features, the structure of which is as follows: Figure 2 As shown, the system includes interconnected data acquisition unit 100, preprocessing unit 200, feature extraction unit 300, statistical screening unit 400, reconstruction and evaluation unit 500, and quantization output unit 600. The data acquisition unit 100 is used to simultaneously acquire the subject's EEG time-series signals through multiple electrode channels, serving as the data source for the entire evaluation process. The preprocessing unit 200, connected to the data acquisition unit 100, performs bandpass filtering on the acquired EEG signals within the range of 1 to 30 Hz, performs bad segment removal and EEG artifact removal based on independent component analysis, and then windows and segments the signals according to a fixed time length, outputting standardized EEG signal segments.

[0038] The feature extraction unit 300 is connected to the preprocessing unit 200 and is used to extract time-frequency features and brain network features in parallel within each time window. Accordingly, the feature extraction unit 300 includes a time-frequency feature extraction module 310 and a brain network feature extraction module 320. The time-frequency feature extraction module 310 obtains the power spectral density of each channel through spectral estimation and calculates time-frequency features such as band power, relative power, δ-α ratio, θ-β ratio, peak α frequency, and left-right brain region asymmetry index. The brain network feature extraction module 320 calculates a weighted phase lag index based on inter-channel cross-spectrum to construct a functional connectivity matrix, then obtains a binary adjacency matrix through proportional thresholding, and calculates the local efficiency of nodes and brain region aggregation values, etc., on the obtained brain functional connectivity map. The statistical screening unit 400 is connected to the feature extraction unit 300 and is used to perform two-level statistical analysis on the extracted features. Accordingly, the statistical screening unit 400 includes an inter-group statistical module 410 and an individual pairing statistical module 420. The intergroup statistics module 410 is used to compare the characteristics of the two groups before and after discharge through independent samples test or nonparametric test, and to screen out the intergroup characteristics based on the significance level; the individual paired statistics module 420 is used to perform paired tests on the characteristics of the same subject before and after discharge, and to take the intersection of the significant characteristics of all subjects to screen out the common characteristics of individuals.

[0039] The reconstruction evaluation unit 500 is connected to the statistical screening unit 400 and is used to complete the training of the autoencoder and the calculation of cross-state reconstruction errors. Accordingly, the reconstruction evaluation unit 500 includes an encoder training module 510, a cross-reconstruction module 520, and a path locking module 530. The encoder training module 510 trains the autoencoder using selected baseline state data as the training set, with the goal of minimizing the same-state reconstruction error, and freezes its weights. The cross-reconstruction module 520 inputs the inter-group features and individual common features of the target state data into the autoencoder with frozen weights, and independently calculates the reconstruction errors of the two inputs. The path locking module 530 compares the two reconstruction errors, locks the branch containing the feature of the path with the smaller reconstruction error as the target evaluation path, and establishes the inter-group brain network features as the core input. The quantization output unit 600 is connected to the reconstruction evaluation unit 500 and outputs the continuous reconstruction error values ​​of the target state data on the target evaluation path as a quantitative indicator representing the topological deviation distance of the brain network, thereby providing continuous brain state evolution evaluation results for clinical use. The above-mentioned units and their internal modules can be implemented by dedicated hardware circuits or by executing corresponding program modules by a general-purpose processor. Those skilled in the art can divide and combine them according to actual needs. Any implementation that achieves the same function does not depart from the protection scope of this invention.

[0040] This invention also provides an electronic device, the structure of which is as follows: Figure 3As shown, the electronic device 700 includes a processor 701, a memory 702, and a computer program stored in the memory 702 and executable on the processor 701. The processor 701 and the memory 702 are connected via a communication bus 703. The electronic device 700 may also include a communication interface 704 for data interaction with external EEG acquisition devices or other devices. The memory 702 may include volatile or non-volatile storage media such as random access memory, read-only memory, or flash memory, used to store the computer program and intermediate data during the evaluation process. The processor 701 may be a central processing unit, digital signal processor, application-specific integrated circuit, or field-programmable gate array, etc., with computational and control capabilities, used to read and execute the computer program stored in the memory 702. When the processor 701 executes the computer program, it implements each step of the aforementioned continuous quantitative evaluation method for brain states based on brain network features, namely, the complete processing process from preprocessing of EEG signals, extraction of time-frequency features and brain network features, two-level feature screening, baseline state autoencoder training, to the calculation and quantization output of cross-state reconstruction errors.

[0041] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various steps of the aforementioned method for continuous quantitative assessment of brain states based on brain network characteristics. The computer-readable storage medium includes, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks. It should be noted that the execution order of the steps in this invention can be appropriately adjusted without violating logical causal relationships, and some steps can be executed in parallel. The embodiments of the aforementioned methods, devices, electronic devices, and computer-readable storage media correspond to and corroborate each other, and the explanation of one embodiment also applies to the other embodiments. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for continuous quantitative assessment of brain states based on brain network features, characterized in that, include: The continuous EEG time-series signals of the subjects were acquired and preprocessed, and divided into continuous EEG signal segments of fixed time length; Within each segment of the continuous EEG signal, the subject's time-frequency features and brain network topology features are extracted in parallel. The extracted time-frequency features and brain network topology features are subjected to inter-group statistical analysis and individual pairing statistical analysis to screen out inter-group features that have significant group differences between preset first different brain states, as well as individual common features that have significant differences in each individual. An autoencoder is trained using predetermined baseline state data to minimize the same-state reconstruction error, and the network weights of the autoencoder are frozen after training convergence. The inter-group features and the common features of individuals corresponding to the target state data are respectively sent as two independent inputs to the autoencoder with frozen weights, and the cross-state reconstruction error of each of the two inputs is calculated independently. By comparing the magnitudes of the cross-state reconstruction errors of the two input paths, the feature branch corresponding to the path with the smaller cross-state reconstruction error is locked as the target evaluation path. The cross-state reconstruction error value of the target state data on the target evaluation path is output as a continuous quantitative index characterizing the deviation distance of the brain network topology.

2. The method according to claim 1, characterized in that, The process of acquiring and preprocessing the subject's continuous EEG temporal signals, dividing them into segments of fixed-length continuous EEG signals, includes: The subject's continuous EEG temporal signals were recorded synchronously through multiple electrode channels; The continuous EEG timing signal is bandpass filtered in the frequency range of 1 to 30 Hz. Bad segments were detected and removed from the filtered EEG signals, and independent component analysis was used to identify and remove artifacts from electrooculography and electromyography. The artifact-free EEG signal is windowed and segmented according to a preset fixed time length to obtain the continuous EEG signal segment.

3. The method according to claim 1, characterized in that, Within each of the continuous EEG signal segments, the time-frequency features and brain network topology features of the subject are extracted in parallel, including: The power spectral density of each electrode channel signal within a preset frequency band is calculated using a spectral estimation method. Based on this, the band power, relative power, power ratio between specific frequency bands, and peak value are calculated for each preset frequency band. Frequency and left-right brain region asymmetry index are used as the time-frequency features; Based on the imaginary part of the cross spectrum between the frequency domain representations of different electrode channels, the weighted phase lag exponent between each pair of electrode channels is calculated to construct the functional connectivity matrix. Extract a proportional threshold based on a preset quantile from the functional connectivity matrix and perform binarization to generate a sparse brain functional connectivity graph structure corresponding to the binary adjacency matrix. On the sparse brain functional connectivity graph structure, the local efficiency of nodes and the regional aggregation value of their corresponding brain regions of interest are calculated as the topological features of the brain network.

4. The method according to claim 3, characterized in that, The peak value The frequency calculation process is as follows: Between 8 and 13 Hz... Within the frequency band, the centroid method is used to calculate the first... Peak value of each electrode channel frequency: ; In the formula, Indicates the first Peak value of each electrode channel frequency, Indicates the first Each channel at frequency Power spectral density at that point and They are respectively The lower and upper limits of the frequency band. For frequency variables; And / or, the formula for calculating the weighted phase lag index is: ; In the formula, Indicates the first The electrode channel and the first Weighted phase lag index between electrode channels Indicates the first The electrode channel and the first The imaginary part of the cross spectrum between individual electrode channels This indicates the expected or time-averaged calculation. It is a symbolic function.

5. The method according to claim 1, characterized in that, The extracted time-frequency features and brain network topological features are subjected to inter-group statistical analysis and individual paired statistical analysis to screen out inter-group features and common individual features, including: Two sets of sample data corresponding to the subject in the first different brain states are obtained. The distribution differences of each feature between the two sets of samples are compared by independent sample test or nonparametric test. Features whose significance level meets the preset threshold are aggregated into the inter-group features. Paired statistical tests were performed on the corresponding features collected from the same subject under the first different brain states to obtain the features that caused significant changes in a single subject; The intersection of features that show significant changes in all participating subjects is taken, and the features that show significant changes in all individuals are extracted as the common features of the individuals.

6. The method according to claim 1, characterized in that, The first distinct brain state includes the baseline state of onset and the evolving recovery state of patients with cerebral infarction. The selected combination of the baseline state data and the target state data includes any of the following: The baseline state data of the onset of disease is used as the baseline state data to train the autoencoder, and the evolutionary recovery state data is used as the target state data. The value of the continuous quantization index positively represents the degree of deviation of the brain network topology from the recovery baseline of the onset of disease. The evolutionary rehabilitation state data is used as the baseline state data to train the autoencoder, and the onset baseline state data is used as the target state data. The physiological state data of healthy control individuals are used as the baseline state data to train the autoencoder, and the data of stroke patients are used as the target state data. The value of the continuous quantification index inversely represents the degree to which the brain network topology approaches the healthy state.

7. The method according to claim 1, characterized in that, The step of training an autoencoder using pre-determined baseline state data to minimize the same-state reconstruction error includes: The autoencoder, which consists of a cascaded encoder and decoder, maps the input feature vector in the baseline state to a low-dimensional latent representation and reconstructs the output. Based on formula Calculate the reconstruction error and update the weights of the autoencoder with minimizing the reconstruction error as the training objective; in, For reconstruction error, For the first One input feature vector, For the corresponding reconstruction output, The number of samples; After comparison and locking, the feature branch corresponding to the path with smaller cross-state reconstruction error is the branch where the inter-group feature is located, and the inter-group feature is established as the brain network topological feature after statistical mapping.

8. A device for continuous quantitative assessment of brain states based on brain network characteristics, characterized in that, include: The preprocessing unit is used to acquire and preprocess the continuous EEG time-series signals of the subject, dividing them into continuous EEG signal segments of fixed time length. The feature extraction unit is used to extract the time-frequency features and brain network topology features of the subject in parallel within each segment of the continuous EEG signal. The statistical screening unit is used to perform inter-group statistical analysis and individual pairing statistical analysis on the extracted time-frequency features and brain network topology features, and to screen out inter-group features that have significant group differences between preset first different brain states, as well as individual common features that have significant differences in each individual. A reconstruction evaluation unit is used to train an autoencoder using predetermined baseline state data to minimize the same-state reconstruction error, and to freeze the network weights of the autoencoder after training convergence. The inter-group features and the common features of individuals corresponding to the target state data are respectively sent as two independent inputs to the autoencoder with frozen weights, and the cross-state reconstruction error of each of the two inputs is calculated independently. Then, by comparing the magnitudes of the cross-state reconstruction errors of the two inputs, the feature branch corresponding to the one with the smaller cross-state reconstruction error is locked as the target evaluation path; The quantization output unit is used to output the cross-state reconstruction error value of the target state data on the target evaluation path as a continuous quantization index characterizing the deviation distance of the brain network topology.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method for continuous quantitative assessment of brain states based on brain network features as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for continuous quantitative assessment of brain states based on brain network features as described in any one of claims 1 to 7.

Images