Web-based disturbance of consciousness electroencephalogram data analysis system

Abstract

The invention discloses a Web-based disturbance of consciousness electroencephalogram data analysis system, which relates to the technical field of medical data processing, and comprises a data acquisition and storage layer, which comprises four levels of front-end acquisition, data buffering, original data storage and a system database, and is used for realizing real-time acquisition, efficient caching, safe storage and rapid retrieval of original electroencephalogram data; extracting obstacle features including spectrum features, complexity indexes, continuity indexes and ERP evoked potentials; the obstacle feature analysis model is used for screening obstacle features, eliminating redundant or low-distinction-degree data, determining key feature data, establishing a key feature data relation network, and combining real-time input data changes and the key feature data relation network to determine the consciousness obstacle state. According to the method, the key feature data relation network is established by obtaining the key feature data, accurate recognition and classification of the consciousness disorder state are achieved, and the consciousness disorder state of the patient is determined by combining with the real-time data.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a Web-based EEG data analysis system for consciousness disorders. Background Art

[0002] Disorders of consciousness (DoC) refer to varying degrees of disturbance or loss of awareness of the external environment and one's own state. These conditions can range from transient (e.g., delirium) to chronic (e.g., vegetative state / unresponsive wakeful state, minimally conscious state) and include various clinical manifestations. DoC is categorized as vegetative state / unresponsive wakeful state (VS / UWS), minimally conscious state (MCS), and completely locked-in state (CLIS). In clinical research, behavioral rating scales (e.g., CRS-R) are often used to grade patients' levels of consciousness. Electroencephalogram (EEG) data are collected during hospital admission and follow-up for rehabilitation prognosis assessment and treatment decision support.

[0003] Announcement No. CN112927793A discloses a Web-based EEG data analysis system for disorders of consciousness. The platform mainly consists of two parts, namely the EEG acquisition experiment and the data analysis platform. First, a resting-state experimental paradigm is selected to collect EEG data using an EEG cap, and then the data is imported into the platform's back-end database for data analysis. The front end selects the algorithm for EEG data analysis and presents the result interface. The back end receives the front-end request and calls the EEG data in the data analysis algorithm database to process and return the results to the front end. Through the disorder of consciousness (DOC) EEG data analysis system, a consciousness level assessment and scientific research platform that integrates the needs of medical staff and scientific researchers can be realized, which will help promote the rapid transformation of computer-assisted analysis technology based on neuroimaging data in DOC clinical evaluation.

[0004] The current mainstream process is for researchers to obtain raw clinical EEG data, perform offline analysis, and then provide feedback to clinicians. This process is lengthy and difficult for interdisciplinary collaboration.

[0005] Lengthy process: Multiple rounds of feedback from scientific research to clinical practice to scientific research take days to weeks, affecting the efficiency of diagnosis and treatment.

[0006] Talent scarcity: There is a lack of compound talents who are proficient in both clinical and signal processing, resulting in high collaboration costs and poor results.

[0007] Huge amount of data: A single resting-state or stimulated EEG recording can reach several GB, and traditional full-data analysis requires enormous computational complexity. Summary of the Invention

[0008] One of the purposes of the present invention is to provide a Web-based EEG data analysis system for disorders of consciousness, which analyzes the disorder characteristics of different disorders of consciousness, screens the EEG data, identifies key segments, eliminates redundant or low-discrimination data, and improves the speed of EEG data analysis.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: A Web-based EEG data analysis system for impaired consciousness, comprising:

[0010] The user and device management layer allows for adding, deleting, modifying, and checking user profiles, viewing details, batch importing and exporting, and registering, status monitoring, maintenance records, and activity logs for EEG and auxiliary devices.

[0011] The data acquisition and storage layer includes four levels: front-end acquisition, data buffering, raw data storage, and system database, which enable real-time acquisition, efficient caching, secure storage, and rapid retrieval of raw EEG data.

[0012] The data preprocessing and feature extraction layer performs denoising, filtering, and segmentation preprocessing on the raw EEG data to extract the obstacle features including spectral features, complexity index, continuity index, and ERP evoked potentials;

[0013] Obstacle feature analysis model, which screens obstacle features, eliminates redundant or low-discrimination data, identifies key feature data, establishes a key feature data relationship network, and combines real-time input data changes with the key feature data relationship network to determine the state of consciousness impairment;

[0014] The visualization module generates clinical reports through visualization, including patient information, DoC determination, and key indicators.

[0015] In one or more embodiments of the present invention, the user and device management layer includes an operation log, which records events and their corresponding time periods, and associates events with patient and doctor information. When EEG and auxiliary equipment generate new data, the events and their corresponding patients and doctors are determined based on the time when the new data is generated, and the new data is systematically configured.

[0016] In one or more embodiments of the present invention, the front-end acquisition is responsible for reading analog electrical signals from EEG and auxiliary devices, performing A / D conversion, converting the analog signals into digital signals, generating raw EEG data, and adding EEG and auxiliary device analog electrical signal timestamps;

[0017] Data buffering is achieved by periodically sending heartbeat frames from the front-end acquisition layer and measuring the round-trip delay to calculate the offset ΔT between the buffer module and the device. During data buffering, the EEG and auxiliary devices are corrected to a unified clock domain based on ΔT.

[0018] In one or more embodiments of the present invention, the data preprocessing steps are as follows:

[0019] Read raw EEG data, check data integrity, plot time domain waveforms and power spectral density, and assess noise levels and artifact types;

[0020] Eliminate irrelevant channels, retain valid signals, and perform filtering;

[0021] Detect artifacts and mark abnormal segments, apply independent component analysis to separate interference signals, and extract pure EEG features;

[0022] Linear detrending was performed on each segment to eliminate electrode offset, and the ERP data were baseline corrected using the baseline segment;

[0023] Window at a fixed time step, overlap and smooth features, and save timestamps, lead numbers, and operation logs;

[0024] Normalize within the window, align based on timestamp or synchronization pulse interpolation, and resample to a uniform sampling rate.

[0025] In one or more embodiments of the present invention, the obstacle feature extraction steps are as follows:

[0026] Load the signal matrix and event labels of each window after preprocessing, and set the global parameters including sampling rate, window length, and δ / θ / α / β / γ frequency bands;

[0027] The power spectral density was estimated using the Welch method, and the absolute power, relative power, spectral entropy, and band ratio were extracted in the δ / θ / α / β / γ frequency bands.

[0028] Calculate sample entropy, approximate entropy, and Lempel–Ziv complexity to quantify the nonlinear dynamic complexity of raw EEG data;

[0029] Based on Hilbert transform and vector autoregression model, the phase locking value, weighted PLI, coherence and Granger causal connectivity index of the channel pair are calculated respectively;

[0030] The stimulation events were divided into epochs of -200–800ms and baseline corrected, and the peak amplitude and latency of ERP in P300, N100 / N200, MMN, and SSEP were extracted;

[0031] All time-frequency, complexity, connectivity, and ERP features are spliced ​​into a unified structured table at the window or epoch level.

[0032] In one or more embodiments of the present invention, redundant or low-discrimination data are eliminated to determine key feature data:

[0033] Eliminate noise and invalid features through preprocessing;

[0034] Use filtering methods to delete highly redundant or weakly correlated features;

[0035] Applying a wrapper approach to iteratively find the most discriminative feature subset in classifier training;

[0036] Automatically compress or remove unimportant features during model training using embedded methods;

[0037] Evaluate the stability of features under different resampling or validation trade-offs through multi-method fusion and nested cross-validation;

[0038] Combine classification performance metrics with clinical interpretability validation to identify and deploy key features with high discrimination and actionable value.

[0039] In one or more embodiments of the present invention, a key feature data relationship network is constructed:

[0040] Summarize the sample value matrix of key feature data and standardize each key feature uniformly;

[0041] Calculate the correlation coefficient matrix of key feature data and generate the fully connected weight matrix W;

[0042] Set the significance level or correlation threshold, retain high-confidence edges, where nodes represent features and edge weights correspond to correlation strength, and construct a key feature data relationship network;

[0043] Take the intersection of the network after multiple resampling, remove low-confidence edges, apply community detection to identify feature modules, and draw local relationship subnetworks for important modules or core nodes;

[0044] Identify hub features, construct feature paths based on hub features, analyze the synergy between features on the path, and when there are sequential data, compare network evolution in different time windows to reveal intervention effects and pathologies.

[0045] In one or more embodiments of the present invention, the state of impaired consciousness is determined by combining real-time input data changes with a network of key feature data relationships:

[0046] Continuously receive the spectrum characteristics, complexity index, connectivity index and ERP evoked potential feature vector of the new data window or epoch, compare the key feature data relationship network, retain the index corresponding to the node as the graph signal input, and load the real-time features as node attributes to the relationship network. The node attribute matrix X∈R (N×F)， Where N is the number of nodes, and F is the single feature dimension of the node;

[0047] Use graph convolution or graph signal processing operators to allow nodes to diffuse information in the relationship network;

[0048] The diffused node activation graph is obtained and aggregated into a single vector g through graph-level embedding. The single vector g is similar to the standard template of the state of consciousness disorder, the score is normalized, and the probability distribution of each state of consciousness disorder is output.

[0049] In one or more embodiments of the present invention, the standard template tk is a standard image-level embedding vector of the k-th type of consciousness disorder state. The calculation process of the standard template is:

[0050] Collect and label training set samples, and define the consciousness disorder classification label yi = k for each sample;

[0051] Build a relationship network for each sample, load real-time features, and calculate graph embedding through graph neural network:

[0052] g i =Readout(GNN(X i ,A));

[0053] Among them, g i is the graph embedding vector of the i-th sample, Readout is the graph-level aggregation function, GNN is the graph neural network operator, X i is the node attribute matrix of the i-th sample, and A is the adjacency matrix of the relationship network;

[0054] Average all graph embeddings of the same type to obtain the template vectors of k classes:

[0055]

[0056] Among them, H k is the total number of samples of the kth class, To sum all sample indexes i with label k;

[0057] Calculate the embedding g of the image to be tested and perform cosine similarity calculation on various standard templates, and select the template with the highest similarity as the judgment type.

[0058] In one or more embodiments of the present invention, visualization results are generated based on the state of consciousness disorder, and the DoC determinations and key indicator results during the analysis process are intuitively displayed in the form of charts.

[0059] Through the above technical solution, the present invention has the following beneficial effects:

[0060] 1. This application obtains key feature data to establish a key feature data relationship network, realizes the accurate identification and classification of the state of consciousness disorder, combines it with real-time data, determines the patient's state of consciousness disorder, improves the accuracy of diagnosis, optimizes the treatment plan, provides strong support for clinical decision-making, and promotes the advancement of medical technology.

[0061] 2. Obtain real-time data for preprocessing, remove noise interference, extract feature parameters, construct feature vectors, ensure data quality, and accurately map the patient's state of consciousness by combining real-time data feature vectors with the relationship network to improve diagnostic efficiency.

[0062] 3. By constructing standard templates corresponding to each state of consciousness disorder, the representativeness of the templates is ensured. Through multiple iterative optimizations, the recognition accuracy is improved, and ultimately accurate classification is achieved, providing a reliable basis for clinical practice.

[0063] Other features and advantages of the present invention will be described in the following description, and part of them will become obvious from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0065] The following drawings illustrate various embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are optional. Furthermore, features from different embodiments may be interchangeably applicable, where practically possible.

[0066] Unless otherwise defined, all words used herein (including technical and scientific terms) have their ordinary meanings as understood by those skilled in the art. Furthermore, the definitions of the above-mentioned words in commonly used dictionaries should be interpreted in the context of this specification as having the same meanings as those in the relevant field of the present invention. Unless otherwise explicitly defined, these words should not be interpreted as having idealized or overly formal meanings.

[0067] The following is an explanation of the terms used in this application:

[0068] EEG (electroencephalogram) records the potential fluctuations of the cerebral cortex in real time by placing electrodes on the surface of the scalp, which is used to evaluate brain function and status. EEG signals mainly come from the synchronous discharge of cortical pyramidal cells. The potential changes generated by these cells after dendritic synapses are conducted through the brain tissue, skull and skin and converge on the scalp, where they can be detected by electrodes.

[0069] VS / UWS: Maintains sleep-wake rhythm, with spontaneous or stimulus-induced eyelid movements, but without any manifestation of conscious content.

[0070] MCS: Discontinuous, occasional conscious behaviors (such as object tracking, localization of painful stimuli, limited communication intentions) occur, but complex tasks cannot be completed continuously.

[0071] EMCS: "Disengagement" from the MCS state and restoration of repeatable purposeful communication (language or gestures) or object use.

[0072] CLIS: The patient is fully conscious but cannot move or communicate due to paralysis of skeletal muscles. Only minimal movement of the eyeballs or palpebral fissures is retained.

[0073] Spectral characteristics refer to the energy distribution of brain electrical activity in different frequency bands. They reflect the unique signs of the state of consciousness, the energy distribution and rhythmic changes of brain waves in different frequency bands, and are associated with wakefulness and cognitive activities.

[0074] The complexity index measures the nonlinear dynamic characteristics of EEG signals, reflects the diversity of neural activities and information processing capabilities, measures the nonlinear dynamic complexity of signals, and is related to the information processing capabilities of neural networks and the level of consciousness.

[0075] Continuity indicators, also known as functional connectivity or network integration, are used to characterize the synchronization or causal relationship between brain regions, describe the synchronization and causal coupling between brain regions, and evaluate network integration and information flow.

[0076] ERP (event-related potential) refers to the time-phase-locked EEG response to specific sensory, cognitive or motor stimuli. The event-related potential reflects the cognition of the stimulus or the integrity of the conduction pathway, revealing residual or latent conscious responses.

[0077] DoCDoC refers to different degrees of consciousness dysfunction after brain injury or disease. The main types include: coma; vegetative state / unresponsive wakeful state (VS / UWS); minimally conscious state (MCS); minimally conscious state (EMCS).

[0078] The above explanations do not fully include the definitions of terms given in this application, but only represent part of the definitions in this application.

[0079] The present invention provides a Web-based EEG data analysis system for disorders of consciousness, which is used to analyze EEG data and improve the processing efficiency of EEG data.

[0080] The analysis system includes:

[0081] The user and device management layer allows for adding, deleting, modifying, and checking user profiles, viewing details, batch importing and exporting, and registering, status monitoring, maintenance records, and activity logs for EEG and auxiliary devices.

[0082] The data acquisition and storage layer includes four levels: front-end acquisition, data buffering, raw data storage, and system database, which enable real-time acquisition, efficient caching, secure storage, and rapid retrieval of raw EEG data.

[0083] The data preprocessing and feature extraction layer performs denoising, filtering, and segmentation preprocessing on the raw EEG data to extract the obstacle features including spectral features, complexity index, continuity index, and ERP evoked potentials;

[0084] Obstacle feature analysis model, which screens obstacle features, eliminates redundant or low-discrimination data, identifies key feature data, establishes a key feature data relationship network, and combines real-time input data changes with the key feature data relationship network to determine the state of consciousness impairment;

[0085] The visualization module generates clinical reports through visualization, including patient information, DoC determination, and key indicators.

[0086] In one feasible solution, the analysis system sets up a web interface to receive EEG data, automatically performs preprocessing and feature extraction, inputs the extracted feature data into the analysis model, evaluates the degree of consciousness disorder in real time, and presents intuitive results through a visualization module to facilitate medical staff to make quick decisions.

[0087] Connect the analysis system to EEG and auxiliary equipment, directly obtain the original EEG data through the analysis system, extract the obstacle features in the original EEG data, establish a relationship network between the obstacle features, and through real-time input of EEG data, determine the obstacle features in the relationship network associated with the input EEG data, and quickly realize the analysis of the input EEG data status.

[0088] In one embodiment, the user and device management layer includes an operation log, which records events and their corresponding time periods, and associates events with patient and doctor information. When EEG and auxiliary devices generate new data, the events and their corresponding patients and doctors are determined based on the time when the new data is generated, and the new data is systematically configured.

[0089] In one feasible solution, the time, patient, and doctor information of the event are recorded through the operation log. When the EEG and auxiliary equipment are in working status at different times, the corresponding events are automatically matched, the data is accurately associated, and the system configuration, data labeling and distribution are automatically triggered.

[0090] For example, the automation task triggering process is as follows:

[0091] New data is stored → patients / doctors are automatically labeled → preprocessing and feature extraction are called;

[0092] Device online / status changed → Update device management table → Refresh status monitoring;

[0093] User actions (such as permission changes) → Dynamically redraw the front-end menu and button visibility.

[0094] By replacing manual configuration with operation logs, system efficiency is improved, ensuring the accuracy and timeliness of data processing, and automatically triggering processes to facilitate subsequent compliance and accountability requirements. This empowers the analysis system with real-time, adaptive, and highly available capabilities, ensuring that clinical data and analysis results are always synchronized and up-to-date.

[0095] In one embodiment, the front-end acquisition is responsible for reading analog electrical signals from EEG and auxiliary devices, completing A / D conversion, converting the analog signals into digital signals, generating raw EEG data, and adding EEG and auxiliary device analog electrical signal timestamps;

[0096] Data buffering is achieved by periodically sending heartbeat frames from the front-end acquisition layer and measuring the round-trip delay to calculate the offset ΔT between the buffer module and the device. During data buffering, the EEG and auxiliary devices are corrected to a unified clock domain based on ΔT.

[0097] In one feasible solution, the timestamp of the front-end acquisition layer is implemented by combining hardware marking with software recording. The EEG amplifier is accompanied by a hardware sampling time for each sample group, and the auxiliary device also outputs the corresponding time on its own sampling interface. When receiving the raw EEG data frame, the host reception time is written into the frame header.

[0098] Dual hardware and software timestamps ensure microsecond accuracy and data synchronization accuracy, and data buffer layer delay correction eliminates network and system jitter to ensure that EEG and auxiliary device data correspond in the clock domain.

[0099] In one embodiment, the data preprocessing steps are as follows:

[0100] Read raw EEG data, check data integrity, plot time domain waveforms and power spectral density, and assess noise levels and artifact types;

[0101] Eliminate irrelevant channels, retain valid signals, and perform filtering;

[0102] Detect artifacts and mark abnormal segments, apply independent component analysis to separate interference signals, and extract pure EEG features;

[0103] Linear detrending was performed on each segment to eliminate electrode offset, and the ERP data were baseline corrected using the baseline segment;

[0104] Window at a fixed time step, overlap and smooth features, and save timestamps, lead numbers, and operation logs;

[0105] Normalize within the window, align based on timestamp or synchronization pulse interpolation, and resample to a uniform sampling rate.

[0106] Exemplarily, the filtering parameters are as follows:

[0107] High-pass filter (0.5 Hz): removes low-frequency drift and electrode drift.

[0108] Low-pass filtering (45–70 Hz): suppresses high-frequency myoelectricity and environmental noise.

[0109] Power frequency notch (50 / 60Hz): Use narrow-band notch filter to eliminate power grid interference.

[0110] ERP baseline interval: -200–0 ms before stimulation. Correction method: Subtract the mean value of this interval from each channel.

[0111] By preprocessing the data, cleaning and standardizing the EEG data and auxiliary equipment data, it is possible to extract obstacle features more accurately, ensuring the reliability and consistency of subsequent analysis results.

[0112] In one embodiment, the obstacle feature extraction steps are as follows:

[0113] Load the signal matrix and event labels of each window after preprocessing, and set the global parameters including sampling rate, window length, and δ / θ / α / β / γ frequency bands;

[0114] The power spectral density was estimated using the Welch method, and the absolute power, relative power, spectral entropy, and band ratio were extracted in the δ / θ / α / β / γ frequency bands.

[0115] Calculate sample entropy, approximate entropy, and Lempel–Ziv complexity to quantify the nonlinear dynamic complexity of raw EEG data;

[0116] Based on Hilbert transform and vector autoregression model, the phase locking value, weighted PLI, coherence and Granger causal connectivity index of the channel pair are calculated respectively;

[0117] The stimulation events were divided into epochs of -200–800ms and baseline corrected, and the peak amplitude and latency of ERP in P300, N100 / N200, MMN, and SSEP were extracted;

[0118] All time-frequency, complexity, connectivity, and ERP features are spliced ​​into a unified structured table at the window or epoch level.

[0119] In an implementable solution, subjective behavior is converted into objective numerical indicators through spectral characteristics, complexity indicators, continuity indicators and ERP-evoked potential disorder characteristics, accurately reflecting the neurodynamic differences of different types of consciousness disorders. The standardized extraction process achieves feature updates in seconds, providing real-time and dynamic brain function assessment basis for clinical diagnosis, assisting in precise medical decision-making and improving treatment effects.

[0120] In one embodiment, redundant or low-discrimination data is eliminated to determine key feature data:

[0121] Eliminate noise and invalid features through preprocessing;

[0122] Use filtering methods to delete highly redundant or weakly correlated features;

[0123] Applying a wrapper approach to iteratively find the most discriminative feature subset in classifier training;

[0124] Automatically compress or remove unimportant features during model training using embedded methods;

[0125] Evaluate the stability of features under different resampling or validation trade-offs through multi-method fusion and nested cross-validation;

[0126] Combine classification performance metrics with clinical interpretability validation to identify and deploy key features with high discrimination and actionable value.

[0127] In an implementable solution, by eliminating redundant and low-discrimination data and outputting them as key feature data, the accuracy and effectiveness of the key feature data can be improved, avoiding the interference of redundant information when establishing the relationship network, which leads to a decrease in the effectiveness of the relationship network, further improving the analysis and diagnosis accuracy of consciousness disorders, and ensuring the accuracy and efficiency of clinical decision-making.

[0128] For example, for the processing of missing values ​​and outliers in preprocessing, a missing threshold is set. When the missing rate reaches the missing threshold, the data are directly eliminated. For data that does not reach the missing threshold, interpolation or nearest neighbor filling is performed to ensure data integrity. For outliers, Z-score standardization is used to identify and eliminate outliers to retain valid information.

[0129] Filtering methods include variance threshold filtering based on variance size, correlation filtering based on correlation coefficient, and mutual information filtering based on the information gain of measurement features and labels;

[0130] The wrapper method uses recursive feature elimination (RFE) based on iterative training of the classifier to eliminate the weakest features and combines cross-validation to determine the optimal feature subset size;

[0131] The embedded method automatically compresses unimportant coefficients to zero during model training through Lasso regression.

[0132] Multi-method intersection and weighted fusion are used to compare the results of the Filter, Wrapper, and Embedded methods, and the intersection or weighted ranking is taken.

[0133] In one embodiment, a key feature data relationship network is constructed:

[0134] Summarize the sample value matrix of key feature data and standardize each key feature uniformly;

[0135] Calculate the correlation coefficient matrix of key feature data and generate the fully connected weight matrix W;

[0136] Set the significance level or correlation threshold, retain high-confidence edges, where nodes represent features and edge weights correspond to correlation strength, and construct a key feature data relationship network;

[0137] Take the intersection of the network after multiple resampling, remove low-confidence edges, apply community detection to identify feature modules, and draw local relationship subnetworks for important modules or core nodes;

[0138] Identify hub features, construct feature paths based on hub features, analyze the synergy between features on the path, and when there are sequential data, compare network evolution in different time windows to reveal intervention effects and pathologies.

[0139] In an implementable solution, through key feature path analysis, the intensity of interaction between features is quantified, combined with time series evolution, the intervention effect is dynamically evaluated, providing accurate basis for clinical decision-making and improving the personalization and effectiveness of diagnosis and treatment plans.

[0140] After generating the relationship network, the node centrality and edge weight are visualized through visualization tools.

[0141] By constructing a relationship network, the strength of the correlation between features can be intuitively displayed. When a patient's EEG data changes, key feature nodes can be quickly located, their dynamic relationship with other features can be analyzed, and the progression of the disease can be monitored in real time. This provides doctors with precise intervention timing and treatment plan adjustment suggestions, significantly improving treatment outcomes.

[0142] In one embodiment, the state of impaired consciousness is determined by combining real-time input data changes with a network of key feature data relationships:

[0143] Continuously receive the spectrum characteristics, complexity index, connectivity index and ERP evoked potential feature vector of the new data window or epoch, compare the key feature data relationship network, retain the index corresponding to the node as the graph signal input, and load the real-time features as node attributes to the relationship network. The node attribute matrix X∈R(N×F) , where N is the number of nodes and F is the single feature dimension of the node;

[0144] Use graph convolution or graph signal processing operators to allow nodes to diffuse information in the relationship network;

[0145] The diffused node activation graph is obtained and aggregated into a single vector g through graph-level embedding. The single vector g is similar to the standard template of the state of consciousness disorder, the score is normalized, and the probability distribution of each state of consciousness disorder is output.

[0146] In one feasible solution, building a relationship network provides the skeleton and collaborative rules between features. Real-time data input is loaded into the network nodes as graph signals. Graph convolution allows collaborative features to amplify each other and suppress isolated features. The aggregated graph-level embedding is matched with a predefined template to provide an interpretable state output.

[0147] By building a relationship network and combining it with real-time data input, we consider the collaborative patterns in brain network dynamics to achieve more accurate, robust and explainable recognition of states of consciousness disorders.

[0148] In one embodiment, the standard template tk is a standard image-level embedding vector of the k-th type of consciousness disorder state. The calculation process of the standard template is:

[0149] Collect and label training set samples, and define the consciousness disorder classification label yi = k for each sample;

[0150] Build a relationship network for each sample, load real-time features, and calculate graph embedding through graph neural network:

[0151] g i =Readout(GNN(X i ,A));

[0152] Among them, g i is the graph embedding vector of the i-th sample, Readout is the graph-level aggregation function, GNN is the graph neural network operator, X i is the node attribute matrix of the i-th sample, and A is the adjacency matrix of the relationship network;

[0153] Average all graph embeddings of the same type to obtain the template vectors of k classes:

[0154]

[0155] Among them, H k is the total number of samples of the kth class, To sum all sample indexes i with label k;

[0156] Calculate the embedding g of the image to be tested and perform cosine similarity calculation on various standard templates, and select the template with the highest similarity as the judgment type.

[0157] For example, Table 1 shows the template values ​​based on typical features:

[0158] Classification α / δ ratio SampEn PLIC3–C4 P300 amplitude Coma <0.2 <0.3 <0.1 <2μV Vegetative state (VS) 0.2–0.4 0.3–0.5 0.1–0.2 1–3 μV Minimum Consciousness (MCS) 0.4–0.7 0.4–0.6 0.2–0.4 3–6 μV Health (HC) >1.0 >0.7 >0.5 8–15 μV

[0159] Table 1 Average range of four types of consciousness states in commonly used EEG feature space

[0160] In one feasible solution, the standard template represents the average manifestation of the characteristic coordination pattern of this type of consciousness disorder state in the relationship network.

[0161] In one embodiment, visualization results are generated based on the state of consciousness disorder, and the DoC determinations and key indicator results during the analysis process are intuitively displayed in the form of charts.

[0162] Exemplarily, the DoC determination result is:

[0163] Vegetative state (VS / UWS): 0.12;

[0164] Minimally conscious state (MCS): 0.78;

[0165] Elimination Minimum Consciousness (EMCS): 0.07;

[0166] Conscious healthy controls (HC): 0.03;

[0167] Final determination: Minimally Conscious State (MCS).

[0168] The key indicators are summarized in Table 2:

[0169] Feature Category Indicator name Patient Values Normal / Abnormal Controls* Spectral characteristics α / δ ratio 0.45 MCS typically ranges from 0.4 to 0.7 Complexity Index SampEn 0.52 MCS typically ranges from 0.4 to 0.6 Connectivity indicators PLI (C3–C4) 0.28 MCS typically ranges from 0.2 to 0.4 ERP Features P300 amplitude 4.2μV MCS typically ranges from 3–6 μV

[0170] Table 2 Comparison of key patient indicators and classification templates

[0171] Comparative analysis confirmed that all patient indicators met the characteristics of a minimally conscious state (MCS), validating the accuracy of the model's assessment. Visual charts further demonstrated subtle changes in the patient's level of consciousness, providing strong support for clinical decision-making.

[0172] Although the present invention is disclosed in conjunction with the above embodiments, it is not intended to limit the present invention. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the definition of the attached claims.

Claims

1. A web-based EEG data analysis system for impaired consciousness, characterized by: include: The user and device management layer allows for adding, deleting, modifying, and checking user profiles, viewing details, batch importing and exporting, and registering, status monitoring, maintenance records, and activity logs for EEG and auxiliary devices. The data acquisition and storage layer includes four levels: front-end acquisition, data buffering, raw data storage, and system database, which enable real-time acquisition, efficient caching, secure storage, and rapid retrieval of raw EEG data. The data preprocessing and feature extraction layer performs denoising, filtering, and segmentation preprocessing on the raw EEG data to extract the obstacle features including spectral features, complexity index, continuity index, and ERP evoked potentials; Obstacle feature analysis model, which screens obstacle features, eliminates redundant or low-discrimination data, identifies key feature data, establishes a key feature data relationship network, and combines real-time input data changes with the key feature data relationship network to determine the state of consciousness impairment; The visualization module generates clinical reports through visualization, including patient information, DoC determination, and key indicators.

2. A Web-based EEG data analysis system for impaired consciousness according to claim 1, characterized in that: The user and device management layer includes operation logs, which record events and their corresponding time periods, and associate events with patient and doctor information. When EEG and auxiliary devices generate new data, the events and their corresponding patients and doctors are determined based on the time when the new data is generated, and the system configures the new data.

3. A Web-based EEG data analysis system for impaired consciousness according to claim 2, characterized in that: The front-end acquisition is responsible for reading analog electrical signals from EEG and auxiliary equipment, completing A / D conversion, converting analog signals into digital signals, generating raw EEG data, and adding EEG and auxiliary equipment analog electrical signal timestamps; Data buffering is achieved by periodically sending heartbeat frames from the front-end acquisition layer and measuring the round-trip delay to calculate the offset ΔT between the buffer module and the device. During data buffering, the EEG and auxiliary devices are corrected to a unified clock domain based on ΔT.

4. A Web-based EEG data analysis system for impaired consciousness according to claim 3, characterized in that: The data preprocessing steps are as follows: Read raw EEG data, check data integrity, plot time domain waveforms and power spectral density, and assess noise levels and artifact types; Eliminate irrelevant channels, retain valid signals, and perform filtering; Detect artifacts and mark abnormal segments, apply independent component analysis to separate interference signals, and extract pure EEG features; Linear detrending was performed on each segment to eliminate electrode offset, and the ERP data were baseline corrected using the baseline segment; Window at a fixed time step, overlap and smooth features, and save timestamps, lead numbers, and operation logs; Normalize within the window, align based on timestamp or synchronization pulse interpolation, and resample to a uniform sampling rate.

5. A Web-based EEG data analysis system for impaired consciousness according to claim 4, characterized in that: The steps for obstacle feature extraction are as follows: Load the signal matrix and event labels of each window after preprocessing, and set the global parameters including sampling rate, window length, and δ / θ / α / β / γ frequency bands; The power spectral density was estimated using the Welch method, and the absolute power, relative power, spectral entropy, and band ratio were extracted in the δ / θ / α / β / γ frequency bands. Calculate sample entropy, approximate entropy, and Lempel–Ziv complexity to quantify the nonlinear dynamic complexity of raw EEG data; Based on Hilbert transform and vector autoregression model, the phase locking value, weighted PLI, coherence and Granger causal connectivity index of the channel pair are calculated respectively; The stimulation events were divided into epochs of -200–800ms and baseline corrected, and the peak amplitude and latency of ERP in P300, N100 / N200, MMN, and SSEP were extracted; All time-frequency, complexity, connectivity, and ERP features are spliced ​​into a unified structured table at the window or epoch level.

6. A Web-based EEG data analysis system for impaired consciousness according to claim 5, characterized in that: Eliminate redundant or low-discrimination data and identify key feature data: Eliminate noise and invalid features through preprocessing; Use filtering methods to delete highly redundant or weakly correlated features; Applying a wrapper approach to iteratively find the most discriminative feature subset in classifier training; Automatically compress or remove unimportant features during model training using embedded methods; Evaluate the stability of features under different resampling or validation trade-offs through multi-method fusion and nested cross-validation; Combine classification performance metrics with clinical interpretability validation to identify and deploy key features with high discrimination and actionable value.

7. A Web-based EEG data analysis system for impaired consciousness according to claim 6, characterized in that: Constructing a key feature data relationship network: Summarize the sample value matrix of key feature data and standardize each key feature uniformly; Calculate the correlation coefficient matrix of key feature data and generate the fully connected weight matrix W; Set the significance level or correlation threshold, retain high-confidence edges, where nodes represent features and edge weights correspond to correlation strength, and construct a key feature data relationship network; Take the intersection of the network after multiple resampling, remove low-confidence edges, apply community detection to identify feature modules, and draw local relationship subnetworks for important modules or core nodes; Identify hub features, construct feature paths based on hub features, analyze the synergy between features on the path, and when there are sequential data, compare network evolution in different time windows to reveal intervention effects and pathologies.

8. The web-based EEG data analysis system for impaired consciousness according to claim 7, characterized in that: Combining real-time input data changes with the key feature data relationship network to determine the state of consciousness disorder: Continuously receive the spectrum characteristics, complexity index, connectivity index and ERP evoked potential feature vector of the new data window or epoch, compare the key feature data relationship network, retain the index corresponding to the node as the graph signal input, and load the real-time features as node attributes to the relationship network. The node attribute matrix X∈R (N×F) , where N is the number of nodes and F is the single feature dimension of the node; Use graph convolution or graph signal processing operators to allow nodes to diffuse information in the relationship network; The diffused node activation graph is obtained and aggregated into a single vector g through graph-level embedding. The single vector g is similar to the standard template of the state of consciousness disorder, the score is normalized, and the probability distribution of each state of consciousness disorder is output.

9. The web-based EEG data analysis system for impaired consciousness according to claim 8, characterized in that: The standard template tk is the standard graph-level embedding vector of the k-th type of consciousness disorder state. The calculation process of this standard template is as follows: Collect and label training set samples, and define the consciousness disorder classification label yi = k for each sample; Build a relationship network for each sample, load real-time features, and calculate graph embedding through graph neural network: g i =Readout(GNN(X i ,A)); Among them, g i is the graph embedding vector of the i-th sample, Readout is the graph-level aggregation function, GNN is the graph neural network operator, X i is the node attribute matrix of the i-th sample, and A is the adjacency matrix of the relationship network; Average all graph embeddings of the same type to obtain the template vectors of k classes: Among them, H k is the total number of samples of the kth class, To sum all sample indexes i with label k; Calculate the embedding g of the image to be tested and perform cosine similarity calculation on various standard templates, and select the template with the highest similarity as the judgment type.

10. A Web-based EEG data analysis system for disorders of consciousness according to claim 9, characterized in that: Generate visual results based on the state of impaired consciousness, and intuitively display the DoC judgment and key indicator results during the analysis process in the form of charts.

Images