Method, system and storage medium for analyzing eeg of a conscious disorder

Abstract

The application discloses a kind of consciousness disorder electroencephalogram analysis method, system and storage medium.The method obtains the multi-channel electroencephalogram data of the object to be evaluated, completes filtering, re-reference, artifact removal and hierarchical anomaly processing;According to the electrode and the mapping relationship of multiple brain function partitions established by electroencephalogram lead information, the partition weight is calculated by combining electrode space mapping coefficient and lead quality factor, and the representative electroencephalogram signal of each partition is generated;Extract the partition multimodal electroencephalogram feature and form the partition feature matrix, calculate the undirected functional connection, the directed directional connection and the undirected topological feature with brain function partition as node;The above-mentioned features are spliced into fixed-dimension input vector, and after feature screening, input hierarchical classification model, output electroencephalogram typing probability for consciousness disorder auxiliary evaluation, and calculate the function retention degree score of each brain function partition based on reference parameter library, generate visualization result and standardized report.

Description

Technical Field

[0001] This invention relates to the field of computer-aided assessment technology, specifically to a method, system, and storage medium for analyzing electroencephalograms (EEGs) of consciousness disorders. Background Technology

[0002] The assessment of brain function in patients with disorders of consciousness typically requires a comprehensive evaluation combining behavioral scales, imaging studies, and neurophysiological signals. Electroencephalography (EEG) is often used as an adjunct to the analysis of brain function in these patients due to its advantages such as convenient acquisition, high temporal resolution, and bedside administration. Current EEG analysis methods usually reflect the patient's brain function through spectral power, complexity, coherence, phase synchronization, or brain network topology indicators. However, most methods still primarily analyze whole-brain average characteristics, single-lead characteristics, or coarse-grained brain network indicators. While these methods can reflect overall brain activity levels to some extent, they tend to average differences between different brain regions, making it difficult to distinguish whether the patient's brain is specifically damaged in the prefrontal cortex (executive function related areas), the temporal lobe (language and auditory related areas), or the posterior parietal lobe (multimodal integration related areas), and also failing to indicate the extent of preservation of local brain function.

[0003] On the other hand, EEG data from patients with altered consciousness are often affected by poor electrode contact, electromyography artifacts, electrooculography artifacts, bad channels, short-term abnormal bands, and differences in acquisition equipment. In existing technologies, bad channel identification, repair, removal of bad segments, feature extraction, functional connectivity calculation, and classification model input are often processed separately, lacking unified engineering constraints. This easily leads to inconsistent feature matrix dimensions, unclear connection network topology sources, unstable model input, or inconsistent functional scoring benchmarks. Especially in scenarios requiring simultaneous output of subtype probabilities and functional retention scores, if the data flow relationships between different processing steps, abnormal data processing rules, fixed feature dimensions, and reference parameter library construction methods are not clearly defined, it is difficult to guarantee the reproducibility and clinical interpretability of EEG analysis results. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problem that existing EEG analysis methods for disorders of consciousness are unable to accurately reflect the preservation of local brain function, local brain function damage, and the network connectivity status between sub-sub ...

[0005] The first aspect of this invention provides a method for analyzing electroencephalograms (EEGs) of consciousness disorders, comprising:

[0006] Acquire multi-channel EEG data of the subject to be evaluated, and perform filtering, rereference, bad channel identification, artifact removal, bad channel repair, bad segment removal and data quality assessment on the multi-channel EEG data; A mapping relationship between EEG electrodes and multiple brain functional regions is established based on EEG lead information. These multiple brain functional regions are partitions obtained according to the functional differences of brain regions. For each brain functional region, the final weight of the corresponding electrode in the brain functional region is calculated based on the electrode spatial mapping coefficient and the lead quality factor. The preprocessed multi-channel EEG signals are then weighted and fused according to the final weight to generate representative EEG signals for each brain functional region. Multimodal EEG features were extracted from representative EEG signals of each brain functional region to form a region feature matrix; Using brain functional regions as nodes, calculate the undirected functional connectivity features and directed directional connectivity features between regions. Construct an undirected connectivity matrix based on the undirected functional connectivity features between regions, and extract network topology features from the undirected connectivity matrix. The partition feature matrix, the undirected functional connectivity features of the partition, the directed directional connectivity features and the network topology features are concatenated into a fixed-dimensional input vector. After feature normalization and fixed-dimensional feature filtering, the vector is input into a hierarchical classification model, and the output is the EEG typing probability result used for auxiliary assessment of consciousness disorders. The functional retention scores for each brain functional region are calculated based on a reference parameter library, and visual analysis results and standardized reports are generated.

[0007] In a second aspect, the present invention provides an electroencephalogram (EEG) analysis system for disorders of consciousness, comprising: The EEG data acquisition / import module is used to acquire multi-channel EEG data of the subject to be evaluated. The signal preprocessing module is used to filter, rereference, identify bad channels, remove artifacts, repair bad channels, remove bad segments, and evaluate data quality for the multi-channel EEG data. The brain functional area mapping module is used to establish the mapping relationship between EEG electrodes and multiple brain functional areas, and calculate the final weight based on the electrode spatial mapping coefficient and lead quality factor to generate representative EEG signals for each brain functional area. The regional EEG feature extraction module is used to extract multimodal EEG features from representative EEG signals of each brain functional region and form a regional feature matrix. The Functional Connectivity and Network Analysis module is used to calculate the undirected functional connectivity characteristics, directed directional connectivity characteristics, and network topology characteristics of the intervals. The classification assessment module is used to input a fixed-dimensional input vector formed by the partition feature matrix, the undirected functional connectivity features of the partition, the directed directional connectivity features and the network topology features into the hierarchical classification model, and output the EEG classification probability results for auxiliary assessment of consciousness disorders. Reference parameter library and scoring module for calculating functional retention scores for each brain functional area; and The Results Visualization and Report Generation module is used to generate visualized analysis results and standardized reports.

[0008] In a third aspect, the present invention provides an electroencephalogram (EEG) analysis device, including a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the above-described method for EEG analysis of consciousness disorders.

[0009] This invention maps EEG electrodes to multiple brain functional zones defined by differences in brain region function, and generates representative EEG signals for each functional zone based on electrode spatial mapping coefficients and lead quality factors. This allows subsequent EEG analysis to move beyond relying on whole-brain averages and instead extract frequency domain features, time domain statistical features, complexity features, and intra-zone synchronicity features for each functional zone, forming a zone feature matrix. Furthermore, this invention calculates undirected functional connectivity features and directed directional connectivity features between brain functional zones as nodes, and extracts network topology features based solely on the undirected connectivity matrix. This enables the simultaneous characterization of local brain region activity and inter-zone network coordination, helping to uncover local functional preservation or reduction that is easily masked by traditional holistic EEG analysis.

[0010] This invention further employs a graded anomaly processing method that combines permanent bad sector identification and repair with transient bad segment removal, avoiding interference from repaired bad sector leads in the determination of transient bad segments. It reduces the uncertainty of model input dimensions by constructing a 248-dimensional fixed input vector using a 6×25-dimensional partition feature matrix, 45-dimensional undirected connectivity features, 30-dimensional directional connectivity features, and 23-dimensional undirected network topology features. It outputs EEG typing probabilities for auxiliary assessment of consciousness disorders through a first-level consciousness existence classification model and a second-level consciousness level classification model, and calculates the functional retention score of each brain functional region using a reference parameter library. Therefore, this invention improves the stability of the EEG data processing workflow, the consistency of model input, and the interpretability of analysis results, facilitating the generation of standardized and visualized EEG-assisted assessment reports.

[0011] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0013] Figure 1 This is an overall flowchart of an EEG analysis method for consciousness disorders based on brain functional partitioning according to the present invention; Figure 2 This is a structural block diagram of an EEG analysis system for consciousness disorders based on brain functional partitioning, according to the present invention. Figure 3 This is a schematic diagram illustrating the mapping relationship between EEG electrodes and brain functional areas in this invention; Figure 4 This is a flowchart of the graded preprocessing of EEG data in this invention; Figure 5 This is a flowchart illustrating the generation of representative EEG signals for a given region based on electrode spatial weights and lead quality weights in this invention. Figure 6 This is a flowchart illustrating the construction of the partitioned EEG feature matrix in this invention; Figure 7 This is a schematic diagram of a partition-level brain function connectivity network constructed using brain functional partitions as nodes in this invention; Figure 8 This is a schematic diagram of the reference parameter library construction and function retention scoring process in this invention; Figure 9 This is a schematic diagram illustrating the classification results, functional retention score, and visualization report generation in this invention. Detailed Implementation

[0014] The specific embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in these embodiments can be combined with each other.

[0015] Example 1: EEG Analysis Method for Consciousness Disorders Based on Brain Functional Partitioning like Figure 1 As shown, this embodiment provides a brain functional partitioning-based EEG analysis method for consciousness disorders, including the following steps: S100: Acquire multi-channel EEG data.

[0016] Acquire multichannel EEG data of the subject to be evaluated, wherein the multichannel EEG data is denoted as:

[0017] in, For the number of EEG leads, The number of sampling points. For the first The EEG leads were in the first The potential value at each sampling point.

[0018] In one specific embodiment, the EEG data uses 19-lead EEG data from the international 10-20 system, with leads including Fp1, Fp2, F3, F4, F7, F8, Fz, C3, C4, Cz, P3, P4, Pz, T3, T4, T5, T6, O1, and O2. The sampling frequency is from 250Hz to 1000Hz. To facilitate unified data processing across different devices, the original EEG data can be resampled to 250Hz or 500Hz.

[0019] EEG data can originate from real-time EEG acquisition devices or from offline EEG files, in formats including EDF, BDF, SET, or CSV. When reading EEG data, the system simultaneously reads the lead name, sampling frequency, acquisition duration, electrode impedance, and event marker information. If electrode impedance information is missing, the impedance quality factor for the corresponding lead is set to the default value of 1.

[0020] In one embodiment, the acquisition time is no less than 5 minutes; if the acquisition time exceeds 5 minutes, the system will prioritize continuous or non-continuous valid segments with a lower artifact ratio for subsequent analysis.

[0021] S200: Preprocesses EEG data and distinguishes between permanent and transient bad sectors. like Figure 4 As shown, the signal preprocessing module 200 processes the multi-channel EEG data acquired in step S100 in the following order: filtering, rereference, permanent bad sector identification, artifact removal, bad sector repair, transient bad segment removal, and data quality assessment. Permanent bad sectors and transient bad segments are identified using different criteria and processing methods. Permanent bad sectors are identified as abnormalities that persist in a single lead over a relatively long period, while transient bad segments are identified as short-term abnormalities occurring in multiple permanently valid leads within the same time window. The determination of transient bad segments does not use leads with permanent bad sector repair as a basis for judgment, to avoid interpolated repair signals affecting the time window quality assessment.

[0022] S2001: Filtering.

[0023] A bandpass filter ranging from 0.5Hz to 45Hz is used to remove low-frequency drift and high-frequency noise; a 50Hz power frequency notch filter is used to remove power frequency interference. If the acquisition environment uses a 60Hz power supply frequency, a 60Hz notch filter is used.

[0024] S2002: Rereference processing.

[0025] The filtered EEG data undergoes rereference processing. In one embodiment, a full-lead average reference is used, calculated as follows:

[0026] in, For reference only One lead signal, This represents the summation of all lead signals at the same moment.

[0027] When performing all-lead averaging, if the system has identified a clearly flat lead or an extremely noisy lead, that lead will not be included in the averaging calculation to avoid bad channels contaminating the other leads.

[0028] S2003: Permanent bad sector identification.

[0029] Permanent bad channels are used to indicate leads that exhibit flatness, saturation, abnormally high amplitude, abnormally low variance, abnormally high variance, abnormally high frequency contamination, or lack of consistency with other leads over a relatively long period. The EEG data after the first artifact removal is divided into multiple detection time windows. In one embodiment, the length of the detection time window is 4 seconds, and the overlap rate between adjacent detection time windows is 50%.

[0030] For any lead within any detection time window, if any of the following conditions are met, that lead is marked as an abnormal lead time window within that detection time window: 1. The continuous flatness time within the detection time window exceeds 1 second; 2. Within this detection time window, the proportion of sampling points with an absolute amplitude exceeding ±300μV exceeds 30%; 3. Within the detection time window, the variance of this lead is greater than 5 times the median variance of all leads within the same time window, or less than 0.2 times the median variance of all leads within the same time window; 4. Within this detection time window, the power in the 20Hz to 45Hz frequency band accounts for more than 0.60% of the total power in the 0.5Hz to 45Hz frequency band. 5. The average correlation coefficient between this lead and the other permanent candidate leads within the detection time window is less than 0.15.

[0031] The ratio of the number of abnormal time windows in each lead to the total number of detection time windows is calculated as the abnormal time window ratio for that lead. If the abnormal time window ratio exceeds 50%, the lead is determined to have permanent bad sectors.

[0032] Permanent bad sectors are identified in the EEG data after artifact removal. A lead is considered a permanent bad sector if it meets any of the following conditions: 1. The continuous flat time exceeds a fixed time; 2. The proportion of time during which the absolute amplitude exceeds the maximum voltage value exceeds a certain proportion; 3. The variance of this lead is more than a multiple of the median variance of all leads, or less than half of the median variance of all leads; 4. The average correlation coefficient between this lead and the other leads is less than a fixed value; 5. This lead was identified as an abnormal lead for more than half of the time windows.

[0033] Let the set of permanent bad sectors be denoted as The set of permanently effective leads is If the number of permanent bad sectors accounts for more than 25% of the total number of leads, that is:

[0034] The system will then output a "Insufficient effective leads" message and terminate the subsequent classification assessment. If the proportion of permanent bad sectors does not exceed 25%, the system will proceed to the bad sector repair step.

[0035] S2004: Artifact Removal.

[0036] Determining the set of permanently valid leads and permanent bad sectors Subsequently, artifact removal was performed on the EEG signals in the permanently valid lead set. Independent component analysis (ICA) was used to remove artifacts from electrooculography (EOG), electromyography (EMG), and electrocardiogram (ECG). Specifically, the multi-channel EEG signal was decomposed into multiple independent components, and the kurtosis, high-frequency energy proportion, and correlation coefficient with the frontal leads Fp1 / Fp2 of each independent component were calculated. An independent component was marked as an artifact component if it met any of the following conditions: 1. The correlation coefficient with lead Fp1 or Fp2 is greater than 0.6; 2. The power in the 20Hz to 45Hz frequency band accounts for more than 0.55% of the total power; 3. Kurtosis greater than 5; 4. The amplitude of a single component exceeds 5 times the median amplitude of all independent components.

[0037] After removing the marked artifacts, the EEG signal was reconstructed. S2005: Bad Sector Repair.

[0038] For leads determined to have permanent bad sectors, repair is performed using weighted interpolation of adjacent leads. For the [missing information - likely a specific lead or parameter]... The repair signal for each bad lead is denoted as:

[0039] in, For the first EEG signals from adjacent effective leads, The interpolation weights are for adjacent valid leads. Calculated based on the reciprocal of the lead spacing:

[0040] in, For the first The bad channel lead and the first Spatial distance between adjacent valid leads To prevent extremely small constants with a denominator of 0.

[0041] For 19-lead data, adjacent leads can be determined based on the spatial proximity of the international 10-20 system; for high-density EEG data, spherical spline interpolation can be used.

[0042] It should be noted that the purpose of bad sector repair is to maintain the channel integrity of the EEG data matrix, facilitating subsequent filtering, visualization, and zonal signal generation. However, in the subsequent zonal weight calculation, the original lead quality factor of the permanently damaged leads is reduced or set to zero, and the system prioritizes using valid leads to generate representative EEG signals for each zonal. In other words, the signal after bad sector repair is not processed as high-confidence raw information, but rather as low-confidence compensation information.

[0043] S2006: Instantaneous bad segment removal.

[0044] The restored EEG data is divided into multiple time windows. In one embodiment, a sliding time window with a length of 4 seconds and an overlap rate of 50% is used.

[0045] Unlike permanent bad sector detection, transient bad sector removal only applies to the set of permanently valid leads. The leads in the time window are used for evaluation, and the original signal of a permanent bad sector is not evaluated repeatedly. For each time window, if any of the following conditions are met, the time window is marked as a transient bad segment: 1. Among permanently valid leads, the proportion of leads with an absolute amplitude exceeding ±200μV is greater than 20% of the total number of permanently valid leads; 2. Within this time window, the proportion of high-frequency power from 20Hz to 45Hz to the total power exceeds 0.60; 3. The absolute amplitude of any temporary partition quality control signal exceeds ±200μV; wherein, the temporary partition quality control signal is only used for instantaneous bad segment determination, and is obtained by equal weighted averaging or median fusion of the permanent effective main leads in the corresponding brain functional partition, without using permanent bad channel repair leads, and is not used as the representative EEG signal of the brain functional partition in step S400. 4. During this time window, large slow wave fluctuations in the frontal leads and high-frequency electromyographic pollution of the whole brain were observed simultaneously.

[0046] Time windows marked as transiently faulty are discarded. Time windows not marked as transiently faulty are included as valid time windows in subsequent analysis.

[0047] Through the above method, this embodiment clearly distinguishes between permanent bad sectors at the lead level and transient bad sectors at the time window level. Permanent bad sectors are marked and repaired first, while the determination of transient bad sectors is based on permanently valid leads and the representative signal of the region. This prevents the entire time window from being incorrectly removed because a repaired bad sector lead triggers "any lead exceeds the limit" again.

[0048] It should be noted that the temporary partition quality control signal is only used to identify transient abnormalities at the time window level, and its calculation does not incorporate lead quality factors or final partition weights; the representative EEG signals of brain functional partitions in step S400 are the formal analysis signals generated after completing the S300 brain functional partition mapping and lead quality factor calculation. Therefore, there is no sequential dependency conflict between transient bad segment elimination and the generation of representative EEG signals of partitions.

[0049] S2007: Data quality assessment.

[0050] Calculate the quality score of EEG data :

[0051] in, To retain the number of effective time windows, For the total number of time windows, The number of permanently effective leads, For the total number of leads, This represents the minimum lead coverage among the six functional brain regions.

[0052] No. Lead coverage of each functional zone Calculate as follows:

[0053] in, For the first The first functional area and the first The fundamental spatial mapping coefficients between leads Indicates the first The value of a lead is determined by whether it is a permanently valid lead; if so, the value is 1, otherwise it is 0. Six functional zones The minimum value in.

[0054] like If the value is less than 0.60, a message "Insufficient EEG data quality" will be displayed, suggesting re-collection or extending the collection time. If the value is not less than 0.60, proceed to the next step.

[0055] S300: Establish the mapping relationship between EEG electrodes and brain functional areas.

[0056] like Figure 3 As shown, the cerebral cortex is divided into six functional regions: the prefrontal cortex (PFC), the motor cortex (MC), the somatosensory cortex (SC), the temporal lobe (TC), the occipital lobe (OC), and the posterior parietal lobe (PPC).

[0057] In one specific embodiment, a basic mapping relationship between EEG electrodes and brain functional areas is established based on the international 10-20 system. Specifically, the prefrontal cortex (PFC) area mainly corresponds to leads Fp1, Fp2, F3, F4, F7, F8, and Fz, with auxiliary leads C3 and C4, used to characterize EEG activity related to attention, executive function, and consciousness regulation; the motor cortex (MC) area mainly corresponds to leads C3, C4, and Cz, with auxiliary leads F3 and F4, used to characterize EEG activity related to motor planning and motor execution; the somatosensory cortex (SC) area mainly corresponds to leads C3, C4, Cz, P3, and P4, with auxiliary lead Pz, used to characterize EEG activity related to somatosensory and proprioceptive functions. The brain electrical activity of the temporal lobe is mainly represented by leads T3, T4, T5, and T6, with auxiliary leads F7 and F8. It is used to represent brain electrical activity related to auditory, language comprehension, and memory functions. The occipital lobe area (OC) mainly corresponds to leads O1 and O2, with auxiliary leads P3, P4, and Pz. It is used to represent brain electrical activity related to visual processing. The posterior parietal lobe area (PPC) mainly corresponds to leads P3, P4, and Pz, with auxiliary leads C3, C4, O1, and O2. It is used to represent brain electrical activity related to spatial cognition, multimodal integration, and self-related processing.

[0058] In the above mapping relationship, the primary corresponding lead is used to provide the main electrophysiological information of the corresponding brain functional area, while the auxiliary corresponding lead is used to supplement the EEG information of the boundary area or adjacent functional area of ​​the area. The same EEG lead can simultaneously serve as the primary corresponding lead of one brain functional area and the auxiliary corresponding lead of another brain functional area, and participates in the generation of representative EEG signals of the corresponding brain functional area with different weights according to its spatial mapping coefficient with different brain functional areas.

[0059] The aforementioned mapping does not simply assign an electrode uniquely to a brain region, but rather allows the same electrode to participate in the representative signal calculation of one or more functional regions with different weights, thereby avoiding the loss of functional information caused by hard segmentation when the spatial resolution of 19-lead EEG is limited. Since the motor cortex and somatosensory cortex regions are anatomically adjacent, and low-density scalp EEG signals exhibit spatial diffusion, leads C3, C4, and Cz can simultaneously contain motor-related and somatosensory-related EEG components. Therefore, the motor cortex and somatosensory cortex regions are allowed to share some major corresponding leads, and different representative EEG signals for each region are formed through different auxiliary corresponding leads and spatial mapping coefficients. The shared mapping is used to preserve the multifunctional EEG information contained in the limited leads in the central region, and does not imply that the shared leads can be completely separated within the sensor space. For any two brain functional regions with shared participating leads, the shared lead components are corrected before calculating the functional connectivity features between the two brain functional regions to reduce the influence of common lead input on the inter-regional connectivity strength.

[0060] For each functional zone and each lead Set the basic space mapping coefficients If lead Belongs to functional area The main corresponding leads, then Take 1.0; if lead Belongs to functional area The auxiliary corresponding lead, then Take 0.5; if lead With functional partitions If there is no corresponding relationship, then Take 0.

[0061] For high-density EEG data, the basic spatial mapping coefficient The distance can be calculated based on the spatial coordinates of the leads and the center point of the functional zones:

[0062] in, For the first The projection position of the first lead is related to the first... Spatial distance between the center points of each functional zone For the first Spatial expansion parameters for each functional partition. The distance can be 1.5 to 3 times the average distance between adjacent leads.

[0063] S400: Generates representative EEG signals for each brain functional region.

[0064] like Figure 5As shown, after completing the basic mapping, the final weight of each lead in each functional zone is calculated by combining the lead signal quality.

[0065] For the Calculate the lead quality factor for each lead. :

[0066] in, This is the impedance quality factor. For artifact quality factor, It is the variance stability factor. This is a permanent bad sector penalty factor.

[0067] When the system reads the lead impedance hour, Calculate according to the following formula:

[0068] in, Take 5kΩ, Use 50kΩ; `clip` restricts the calculated value to between 0 and 1. If no lead impedance is read, then... Take 1.

[0069] Artifact quality factor Calculate according to the following formula:

[0070] in, For the first The number of time windows in which each lead is identified as a transient anomaly. This represents the total number of time windows.

[0071] Variance stability factor Calculate according to the following formula:

[0072] in, For the first The signal variance of each lead This represents the median of the variance of all leads.

[0073] Permanent bad sector penalty factor Determined according to the following rules: like ,but ; like And if at least one permanent primary lead still exists within this functional zone, then ; like And if there is no permanently valid primary lead within this functional area, then .

[0074] Therefore, when a functional partition has a valid primary lead, the system does not rely on bad sector interpolation signals to generate a partition representative signal; only when a partition lacks a sufficient number of valid primary leads is a low-weight repair signal allowed to be used to avoid the complete loss of that partition.

[0075] For the The first functional area and the first Each lead, final partition weight Calculate according to the following formula:

[0076] in, Indicates the first The sum of the products of the basic spatial mapping coefficients and the lead quality factors of all leads participating in each functional zone.

[0077] No. Representative EEG signals from each functional area Calculate according to the following formula:

[0078] in, After pretreatment and necessary bad sector repair, the first One lead of EEG signal, For the first Representative EEG signals from each functional area.

[0079] Using the above method, the system converts the raw multi-channel EEG signals into representative EEG signals for six functional zones: .

[0080] The artifact quality factor is calculated based on the proportion of permanently valid leads marked as transient abnormal leads within the entire detection time window in S2006. For leads repaired for permanent bad sectors, the artifact quality factor is not recalculated based on the repaired signal; instead, the permanent bad sector penalty factor is directly applied. Therefore, the lead quality factor can reflect the true reliability of the original lead signal, avoiding the misidentification of the smoothed signal after interpolation repair as a high-quality signal.

[0081] S500: Extract regional EEG features and construct a regional feature matrix.

[0082] like Figure 6 As shown, feature extraction is performed on representative EEG signals from each brain functional region to form a region feature matrix.

[0083] For the Power spectral density was calculated for each brain functional region using the Welch power spectral estimation method. In the Welch method, the sub-window length is 2 seconds, the sub-window overlap rate is 50%, and the window function is a Hamming window.

[0084] Define the following frequency bands: Delta band: 0.5Hz to 4Hz; Theta band: 4Hz to 8Hz; Alpha band: 8Hz to 13Hz; beta band: 13Hz to 30Hz; gamma band: 30Hz to 45Hz.

[0085] No. The partition in the _th ... absolute power of each frequency band Calculate according to the following formula:

[0086] relative power Calculate according to the following formula:

[0087] in, It represents the sum of the absolute power of the five frequency bands: delta, theta, alpha, beta, and gamma.

[0088] Further calculations were performed on the slow wave ratio, fast-slow wave ratio, and spectral center characteristics:

[0089]

[0090]

[0091] in, To prevent extremely small constants with a denominator of 0, we can take... .

[0092] Spectral centroid Calculate according to the following formula:

[0093] Spectrum edge frequencies To achieve a cumulative power of 95% of the total power at the specified frequency.

[0094] For time-domain statistical characteristics, calculate the first... Root mean square value of representative EEG signals from each region Standard deviation Peak-to-peak value skewness and kurtosis .

[0095] For complexity features, calculate sample entropy. Permutation entropy and approximate entropy In sample entropy calculation, the embedding dimension... Choose 2, tolerance parameter Use 0.2 times the signal standard deviation; in the permutation entropy calculation, the embedding order is taken as 3 to 5, and the delay time is taken as 1.

[0096] For the synchronization characteristics within a partition, if the first... If a partition contains at least two valid participating leads, then the phase lock value (PLV) and coherence (Coh) are calculated pairwise for each participating lead in that partition.

[0097] For intra-partition synchronization characteristics, only the first... Permanently effective participating leads within each brain functional region are used for calculation, while permanent bad passage repair leads are not used for the calculation of synchronization characteristics within the region, in order to avoid artificially increasing synchronization due to common signal components between interpolated repair signals.

[0098] The first The number of permanently active leads within each brain functional region is denoted as _____. .when At that time, the phase lock value PLV and coherence Coh are calculated pairwise for each of the permanently active participating leads.

[0099] No. The first lead and the first The phase lock-in value between leads is calculated as follows:

[0100] in, and The first The first lead and the first The first lead in The instantaneous phase of each sampling point The number of sampling points. The unit is the imaginary number. The instantaneous phase is obtained by performing a Hilbert transform on the EEG signal of the corresponding frequency band.

[0101] No. The intra-regional phase synchronicity characteristics of each brain functional region are calculated as follows:

[0102] in, For the first A set of permanently active engagement leads within each brain functional region.

[0103] No. The first lead and the first The coherence between leads is denoted as , No. The intra-regional coherence characteristics of each brain functional region are calculated as follows:

[0104] when At that time, since there are no permanently valid lead pairs available for pairwise synchronization calculation, the brain functional region is not calculated. and The two features are set to missing values, but are not assigned the value 0 or deleted from the partition feature vector.

[0105] During the offline training phase, missing values ​​are imputed using the median of the corresponding synchronicity features in the training set, and this median is saved as model metadata. During the online evaluation phase, the imputed values ​​are retrieved from the median saved during offline training, without recalculating the imputed values ​​using data from the evaluation object, validation set, or test set. Thus, each brain functional region maintains a fixed 25-dimensional feature vector, and the six brain functional regions still form a fixed... Dimensional partition feature matrix.

[0106] No. Synchronization characteristics within functional partitions This is the average value of PLV for all lead pairs within this partition; This represents the average coherence of all lead pairs within this region.

[0107] In one specific embodiment, the first Feature vectors of brain functional regions It includes the following 25 features:

[0108]

[0109]

[0110]

[0111] in, to The characteristics are obtained by taking the logarithm of the absolute power of the five frequency bands. to The relative power characteristics of the five frequency bands are shown. , and As a power ratio characteristic, and As for the spectral distribution characteristics, , , , and Five time-domain statistical characteristics, , and As a complexity feature, and This refers to the synchronization characteristic within the partition.

[0112] After extracting the above 25 features from the six brain functional regions respectively, a dimensional partition feature matrix :

[0113] This partition feature matrix contains 150 partition EEG features, which are used for subsequent classification of consciousness impairment and scoring of functional retention.

[0114] S600: Constructing a regional-level brain functional connectivity network.

[0115] like Figure 7 As shown, a partition-level brain functional connectivity network is constructed using six brain functional regions as network nodes and the functional connectivity strength between regions as edges.

[0116] For any two functional partitions and Based on its representative EEG signals and Calculate the following connectivity features: 1. Phase Lock Value ; 2. Frequency domain coherence ; 3. Weighted phase lag index ; 4. Directional connection strength and .

[0117] The directional connectivity strength was calculated using the Granger causality algorithm. Specifically, an autoregressive model containing only historical signals from the target partition and a joint autoregressive model containing historical signals from both the source and target partitions were constructed. If the prediction residual of the target partition decreased after incorporating the historical signal from the source partition, the source partition was considered to have a directional influence on the target partition. Model order. Select between 1 and 200 using the AIC criteria.

[0118] For each pair of functional partitions, calculate the overall undirected connectivity strength. :

[0119] Norm represents normalization.

[0120] There are 15 undirected partition pairs among the six functional zones, resulting in 15 PLV connection features, 15 coherent connection features, and 15 wPLI connection features, for a total of 45 undirected connection features. Simultaneously, there are 30 directed partition pairs among the six functional zones, resulting in 30 directional connection features.

[0121] It consists of six functional zones. undirected connection matrix and Directed connection matrix For an undirected connectivity matrix , Take 0, For a directed connectivity matrix , Indicates the first The functional partitions up to the first The intensity of the information impact of each functional partition. Take 0.

[0122] To avoid unclear calculation benchmarks for network topology features, this embodiment explicitly stipulates that: by default, network topology features are based solely on the undirected connectivity matrix. Calculate the directed connectivity matrix. It is only used as input to the classification model for 30 directional connection features and does not participate in the calculation of default network topology features. If other embodiments require calculation based on the directed connection matrix... When calculating directed network topology features, these features should be defined separately as an additional feature group and not included in the fixed input dimension of this embodiment.

[0123] Based on undirected connection matrix Calculate network topology characteristics ,include: 1. Overall efficiency ; 2. Average clustering coefficient ; 3. Feature path length ; 4. Small World Attributes ; 5. Modularity ; 6. Node strength of the six partition nodes ; 7. Node degree of the six partition nodes ; 8. Betweenness centrality of the six partition nodes .

[0124] Therefore, the topological characteristics of undirected networks It includes: 5 global topological features + 6 node strength features + 6 node degree features + 6 betweenness centrality features = 23 topological features.

[0125] partition feature matrix Undirected connectivity features directional connectivity features and undirected network topology characteristics Concatenate to form a fixed-dimensional model input vector :

[0126] in, 150 dimensions 45 dimensions 30-dimensional It is 23-dimensional. Therefore, It is a 248-dimensional fixed input vector.

[0127] S700: Outputs consciousness disorder classification results based on fixed-dimensional feature screening and hierarchical classification model.

[0128] This embodiment employs a hierarchical classification model to assist in the assessment of brain function in patients with consciousness disorders using their EEG data. To facilitate understanding by those skilled in the art, this embodiment divides the implementation process of the hierarchical classification model into an offline training phase and an online assessment phase. The offline training phase is used to establish a training sample library, determine feature selection rules, train the hierarchical classification model, and save model metadata. The online assessment phase involves inputting the fixed-dimensional input vector of the patient to the trained hierarchical classification model and outputting the EEG typing probability results for the auxiliary assessment of consciousness disorders.

[0129] The offline training phase is not repeated every time the subject is evaluated, but rather performed before system deployment, during model updates, or after the reference sample library is expanded. The online evaluation phase is performed after obtaining the multi-channel EEG data of the subject and completing the processing described in steps S100 to S600.

[0130] I. Establishment of the training sample library during the offline training phase Establish a training sample library. Each training sample includes a segment of multichannel EEG data that meets quality requirements, along with a corresponding clinical label. The clinical label can be determined by at least two CRS-R scale assessments; if multiple assessments are inconsistent, the expert review result shall be used as the final label.

[0131] In one embodiment, the training sample library includes four types of samples: VS, MCS-, MCS+, and EMCS. Each type contains no fewer than 20 samples, or each type corresponds to no fewer than 60 valid EEG recordings, with each valid EEG recording lasting no less than 60 seconds. The training samples should use the same or uniformly calibrated EEG acquisition protocol as the subjects being evaluated, including the same or mappable lead system, sampling frequency, filtering range, and reference method. During model training and validation, the model is divided into training, validation, and test sets based on the patient; different EEG segments from the same patient must not appear simultaneously in the training and test sets. Feature selection, normalization parameter calculation, missing value imputation parameter calculation, and probability calibration parameter calculation are all performed only within the training set, and the parameters saved during the training phase are directly called in the validation and test sets. Feature selection results must not be recalculated using validation or test set data.

[0132] In one embodiment, the training samples are divided into training, validation, and test sets in a 7:1:2 ratio; when the sample size is insufficient, five-fold cross-validation by patient grouping is used. Within each fold, low-variance feature removal, redundant feature removal, mutual information ranking, and model parameter optimization are performed only within the training fold.

[0133] To evaluate model stability, at least four class classification accuracy, MCS and above state recognition sensitivity, VS recognition specificity, macro-average F1 score, AUC value, and calibration error should be calculated. If the maximum class probability in the test set is inconsistent with the actual label, or if the maximum class probability is lower than 0.45, the sample is marked as a low-confidence assessment sample, and the report will indicate that clinical review is required.

[0134] The inclusion criteria for training samples include: 1. Clinically diagnosed as a state of altered consciousness; 2. Complete at least one CRS-R scale assessment within 24 hours prior to EEG collection; 3. No sustained epileptic seizures or severe motor artifacts were observed during the EEG data collection process; 4. Quality score of EEG data after processing via step S200 Not less than 0.60.

[0135] The exclusion criteria for training samples include: 1. The patient was under deep anesthesia during the collection process; 2. The dosage of sedative drugs makes it impossible to determine the clinical state of consciousness; 3. The proportion of permanent bad sectors exceeds 25%; 4. The number of effective EEG time windows is less than 40% of the total number of time windows.

[0136] For each training sample in the training sample library, the same process described in S100 to S600 is followed to obtain the fixed-dimensional input vector corresponding to that training sample. The fixed-dimensional input vector It is formed by splicing together 150-dimensional regional EEG features, 45-dimensional undirected connectivity features, 30-dimensional directional connectivity features, and 23-dimensional undirected network topology features, for a total of 248 dimensions.

[0137] II. Fixed-dimensional feature selection during offline training To avoid uncertainty in the feature selection process, this embodiment adopts a fixed-rule feature selection process. For each training sample in the training sample library, a 248-dimensional fixed-dimensional input vector is extracted. In the training set, feature normalization, missing value processing, low variance feature deletion, redundant feature deletion, and mutual information sorting are performed uniformly.

[0138] First, robust normalization is performed on each feature in the training set:

[0139] in, For the first In the training samples, the th The original values ​​of each feature, These are the normalized eigenvalues. For the training set The median of the features, For the first Interquartile range of each feature To prevent extremely small constants with a denominator of 0, Can be taken .

[0140] Secondly, missing value handling is performed. If the missing percentage of a certain feature in the training set exceeds 20%, that feature is deleted; for the missing values ​​among the remaining features, the median of the corresponding feature in the training set is used for imputation. The missing values ​​also include intra-regional phase synchronization features and intra-regional coherence features that cannot be calculated because the number of permanently effective participating leads in a certain brain functional region is less than two. For these types of missing values, the median of the corresponding brain functional region and the corresponding feature location in the training set is used for imputation. For example, when an intra-regional phase synchronization feature is missing in the prefrontal cortex, the median of the same feature in the prefrontal cortex in the training set is used for imputation, instead of using the median of synchronization features from other brain functional regions.

[0141] Next, perform low-variance feature removal. If the variance of a feature in the training set is less than... If so, then delete that feature.

[0142] Then, redundant feature removal is performed. The Spearman correlation coefficient between any two features is calculated. .like If the mutual information values ​​of the two features with the classification label are equal, the feature with the lower missing proportion is retained; if the missing proportions are still equal, the feature with the smaller feature number is retained.

[0143] Mutual information value The calculation is performed as follows: For continuous features, first discretize the feature into 5 levels according to the quintiles based on the training set distribution, and then calculate the mutual information value between the discretized feature and the classification label:

[0144] in, Represents the feature level after discretization. Indicates category labels, This represents the joint probability of feature level and classification label. and These represent marginal probabilities.

[0145] For the first-level consciousness existence classification model, the classification labels are binary labels, where VS samples are labeled 0, and MCS-, MCS+, and EMCS samples are labeled 1. The top 80 features, sorted by mutual information value from highest to lowest, are retained as the input feature set for the first-level model. If fewer than 80 features remain after handling missing values, removing low-variance features, and removing redundant features, then all remaining features are retained.

[0146] For the two-level consciousness level classification model, the classification labels are three-category labels, where MCS-, MCS+, and EMCS serve as the three classification labels. The two-level model only uses MCS-, MCS+, and EMCS samples to calculate mutual information values. The features are sorted from highest to lowest mutual information value, and the top 80 features are retained as the input feature set for the two-level model. If fewer than 80 features remain after handling missing values, removing low-variance features, and removing redundant features, then all remaining features are retained.

[0147] Level 1 model input feature set and the input feature set of the second-level model After training, it is saved as model metadata. The model metadata includes at least normalization parameters, missing value imputation parameters, and... Feature index set The feature index set and probability calibration parameters are used. Therefore, during the online evaluation phase, corresponding features can be extracted according to the feature indexes and processing rules saved during the offline training phase, ensuring that the model input dimension, feature source, and processing rules remain fixed.

[0148] III. Training of hierarchical classification models during the offline training phase The hierarchical classification model in this embodiment includes a first-level consciousness existence classification model. and the two-level consciousness level classification model .

[0149] First-order consciousness existence classification model This is used to determine whether the subject under evaluation has detectable consciousness-related brain function activity, and its output is the probability of the presence of consciousness. The training labels for the Level 1 model are binary classification labels, where VS samples are labeled as 0, and MCS-, MCS+, and EMCS samples are labeled as 1.

[0150] Second-level consciousness classification model Used to output conditional probabilities among MCS-, MCS+, and EMCS, its output is , and The secondary model is trained using MCS-, MCS+, and EMCS samples, but VS samples are not used in the training of the secondary model.

[0151] In one specific embodiment, a first-order consciousness existence classification model and the two-level consciousness level classification model All models employ a gradient boosting tree model. Model parameters can be set as follows: tree depth 3 to 6, number of leaf nodes 15 to 63, learning rate 0.01 to 0.10, number of trees 100 to 500, minimum number of leaf node samples 5 to 30, and the loss function is cross-entropy loss. For imbalanced data, class weights are used for correction; the class weights are the normalized values ​​of the inverse of the number of samples in each class in the training set.

[0152] The model training process includes: 1. Extract 248-dimensional fixed-dimensional input vectors from the training samples in the training sample library according to S100 to S600. ; 2. Determine the input feature set of the first-level model and the input feature set of the second-level model ; 3. According to Input vector from the training samples (248 dimensions fixed dimension) Extracting the first-level model input vector and use Training a first-order consciousness existence classification model ; 4. According to 248-dimensional fixed-dimensional input vector from MCS-, MCS+, and EMCS training samples Extracting the input vector of the secondary model and use Training a Level 2 Consciousness Classification Model ; 5. Determine model parameters through five-fold cross-validation; 6. Use Platt scaling or isometric regression to calibrate the model output probabilities; 7. Save model parameters, normalization parameters, missing value imputation parameters, Feature index set Feature index set and probability calibration parameters.

[0153] Through the aforementioned offline training phase, the system obtains a fully trained classification model of the existence of first-level consciousness. Secondary level consciousness classification model In addition, the fixed feature selection rules and model metadata corresponding to the model.

[0154] IV. Fractal probability output during the online assessment phase When actually evaluating the object to be evaluated, the system does not rebuild the training sample library or retrain the hierarchical classification model. Instead, it calls the already trained and saved model and model metadata.

[0155] Specifically, the multichannel EEG data of the subjects to be evaluated are processed according to S100 to S600 to obtain a 248-dimensional fixed-dimensional input vector of the subjects to be evaluated. Subsequently, based on the normalization parameters and missing value imputation parameters saved during the offline training phase, the fixed-dimensional input vector of the object to be evaluated is... Perform the same data processing as during the training phase; then, based on the saved data... Feature index set and The feature index set is derived from the 248-dimensional fixed-dimensional input vector. Extracting the first-level model input vector from each and the input vector of the second-level model .

[0156] Input vector of the first-level model Input First-Level Consciousness Existence Classification Model To obtain the probability of the existence of consciousness ; Input vector of the second-level model Input Level 2 Consciousness Classification Model The conditional probabilities of MCS-, MCS+, and EMCS are obtained and denoted as follows: , and .

[0157] During the online assessment phase, regardless of the first-level consciousness existence classification model... The probability of the output consciousness existing. Is it below the first threshold? The system executes a two-level consciousness level classification model. ,get , and First threshold This is only used for interpreting the final results and providing confidence level information; it is not used to skip the calculation of the second-level model.

[0158] The final four probabilities are calculated as follows:

[0159]

[0160]

[0161]

[0162] in, , and A classification model for the second level of consciousness Output the conditional probability, and satisfy:

[0163] Final EEG typing results Determine as follows:

[0164] like If the probability of VS is high or the brain activity related to consciousness is low, the system will display the four probability categories: VS, MCS-, MCS+, and EMCS. If the probability of the largest category is less than 0.45, or the difference between the probability of the largest category and the probability of the second largest category is less than 0.10, the system will output the message "Uncertain, clinical review recommended".

[0165] First threshold This can be determined through a validation set. Specifically, candidate thresholds between 0.35 and 0.65 are iterated through in the validation set, and the threshold that best combines sensitivity and specificity for MCS and higher states is selected. If the training data is insufficient to support threshold optimization, then... Take 0.50.

[0166] S800: Construct a reference parameter library and calculate the functional retention score for each brain functional region. For example... Figure 8 As shown, this embodiment further outputs a functional retention score for each brain functional region, with a score range of 0 to 100. The functional retention score is calculated by combining the spectrum, complexity, synchronization, and network connectivity of each functional region.

[0167] S8001: Reference parameter library construction.

[0168] To avoid insufficient disclosure due to the lack of a reference parameter library construction method, this embodiment clarifies the construction method of the reference parameter library as follows.

[0169] The reference parameter library consists of EEG samples that have undergone clinical typing and annotation, including four categories: VS, MCS-, MCS+, and EMCS. Each category contains at least 20 samples, or at least 60 valid EEG segments per category, with each valid EEG segment lasting at least 60 seconds. Each sample undergoes the same process (S100 to S600) to extract a partition feature matrix. Connection characteristics directional connectivity features and topological features .

[0170] The inclusion criteria for the reference sample include: 1. Clinical diagnosis is performed by a neurologist or rehabilitation physician; 2. Complete the CRS-R scale assessment within 24 hours before and after EEG collection; 3. The same patient should undergo at least two behavioral assessments; 4. If the results of the two assessments are inconsistent, the final label shall be determined by two or more clinical experts. 5. EEG data quality score Not less than 0.60; 6. The acquisition protocol is consistent with the object to be evaluated, or lead mapping and sampling rate unification have been completed.

[0171] The exclusion criteria for the reference sample include: 1. The patient experienced a persistent epileptic seizure during the data collection process; 2. Deep sedation or anesthesia may have been present during the collection process; 3. Insufficient effective leads; 4. Insufficient effective time window; 5. Clinical labeling could not be confirmed.

[0172] The reference parameter library is managed hierarchically according to equipment model, lead system, and sampling protocol. If the number of reference samples under the same equipment model and lead system meets the above requirements, the same equipment reference parameter library shall be used first; if the same equipment reference samples are insufficient, cross-equipment reference parameter libraries shall be used, but resampling, bandwidth limitation, rereference, feature normalization, and lead mapping shall be uniformly completed before entering the parameter library.

[0173] For each brain functional region and each feature The following parameters are calculated using the reference parameter library: : No. The first partition The median of the features; : No. The first partition Interquartile range of each feature; : No. The first partition The 5th percentile of each feature; : No. The first partition The 95th percentile of the feature.

[0174] The above parameters are stored separately according to different consciousness level labels, or a "functional retention reference distribution" can be formed based on MCS and above samples, and a "functional decline reference distribution" can be formed based on VS samples. In one embodiment, the functional retention score preferentially uses the reference distribution formed by MCS+ and EMCS samples as a higher functional retention benchmark.

[0175] In addition to being stratified by device model, lead system, and sampling protocol, the reference parameter library can also be sub-libraries managed by age group, etiology type, acquisition status, and drug influence status. The etiology types include traumatic brain injury, hypoxic-ischemic brain injury, post-stroke consciousness impairment, and consciousness impairment caused by other reasons; the acquisition status includes quiet with eyes open, quiet with eyes closed, natural sleep, or a state of stimulation task; the drug influence status records at least the presence of sedative, antiepileptic, or anesthetic drugs within 24 hours prior to acquisition.

[0176] Each reference parameter library is assigned a version number. Version information includes sample source, sample quantity, acquisition device, lead system, preprocessing parameters, feature template, percentile parameters, and activation date. When the number of newly added reference samples exceeds 20% of the original reference library's sample quantity, or when the acquisition device, lead system, or preprocessing parameters change, the system recalculates the reference parameter library while retaining the old version's parameters to ensure the traceability of historical reports.

[0177] If the object to be evaluated cannot be matched with a reference parameter library of the same equipment, lead system and acquisition status, the system will only output a relative trend prompt and will not output a function retention score of 0 to 100 points, or mark it as "cross-library reference score" in the report.

[0178] S8002: Feature percentile normalization.

[0179] For the object to be evaluated, the first The first brain functional region Features Calculate percentile normalized values :

[0180] in, and From the reference parameter library, Pick .

[0181] For features that increase damage, such as delta relative power, theta relative power, slow wave / fast wave ratio, and delta / alpha ratio, their normalized values ​​are inverted:

[0182] For features that retain enhanced functionality, such as alpha relative power, beta relative power, sample entropy, permutation entropy, approximate entropy, intra-partition synchronicity, and node strength, their normalized values ​​remain unchanged. .

[0183] S8003: Sub-score calculation for partitions.

[0184] Calculate the sub-scores for the five categories: 1. Slow wave suppression score It is obtained by weighting the delta relative power, the theta relative power, and the inverted normalized value of the slow wave / fast wave ratio:

[0185] 2. Fastwave retains score It is obtained by weighting the normalized values ​​of the alpha relative power and the beta relative power:

[0186] 3. Complexity score It is obtained by weighting the normalized values ​​of sample entropy, permutation entropy, and approximate entropy:

[0187] 4. Integral scoring within the zone It is obtained by weighting the normalized values ​​of PLV_intra and Coh_intra:

[0188] 5. Scoring based on interval connections It is obtained by weighting the normalized values ​​of node strength, node degree, and betweenness centrality of the partition:

[0189] S8004: Function Retention Rating Calculation No. Functionality retention rating for each functional area Calculate according to the following formula:

[0190] Each sub-score ranges from 0 to 1. The higher the value, the closer the brain activity and network connectivity of that functional area are to the performance of samples with higher levels of consciousness in the reference library; The lower the value, the more significant the functional reduction in that functional area.

[0191] For the six functional brain regions, we obtain the following: , , , , ,

[0192] The system further calculates the overall brain function retention index. :

[0193] in, Assign weights to each functional partition. In one embodiment, Take 0.25, Take 0.20, Take 0.15, Take 0.15, Take 0.10, Let's take 0.15. The sum of the above weights is 1.

[0194] If the number of samples in the reference parameter library is insufficient to establish a stable percentile distribution, the system will not output a functional retention score of 0 to 100, but will only output the original features of the partition, the relative trend, and a "reference library insufficient" prompt; the system will then enable the formal scoring function after the reference parameter library meets the sample conditions described in S8001.

[0195] S900: Generates visualization results and standardized reports.

[0196] like Figure 9 As shown, the results visualization and report generation module 800 generates a standardized report based on the fractal probability, partition score, and connection matrix.

[0197] The report should include at least the following: 1. Basic information of EEG data, including sampling time, number of leads, effective duration, and data quality score; 2. Number of permanent bad sectors, proportion of transient bad sectors, and lead coverage in each region; 3. Representative signal quality of the six functional brain regions; 4. Scoring of the degree of functional retention in the six brain functional areas; 5. Probability of consciousness disorder classification, including the probability of VS, MCS-, MCS+ and EMCS; 6. Probability of the existence of primary consciousness and second-order conditional probability , , ; 7. Heatmap of regional connectivity matrix; radar map of functional retention of the prefrontal cortex (attention / executive), posterior parietal lobe (spatial cognition / multimodal integration), temporal lobe (auditory / language / memory), motor cortex (motor planning and execution), somatosensory cortex (somatosensory) and occipital lobe (visual processing).

[0198] 8. Regional brain functional network topology diagram; 9. Radar chart of functional preservation in the prefrontal cortex, posterior parietal lobe, temporal lobe, motor cortex, somatosensory cortex, and occipital lobe; 10. The system prompts a message stating, "The results are for auxiliary assessment of brain function and should be judged in conjunction with clinical behavioral assessments and imaging data."

[0199] In one embodiment, if a functional area score is significantly higher than the overall score, for example, if the functional area score is more than 20 points higher than the overall score, the system marks it as "local function retention indication"; if a functional area score is significantly lower than the overall score by more than 20 points, the system marks it as "local function decline indication". This method can highlight local function retention that is easily masked by traditional whole-brain averaging analysis.

[0200] In this embodiment, the thresholds for bad segment ratio, EEG data quality, instantaneous bad segment amplitude, low confidence threshold for classification probability, and probability difference threshold can be determined using a training sample library or a validation sample library. Specifically, the candidate threshold ranges are traversed in the validation set, and the effective sample retention rate, artifact removal rate, four-class classification accuracy, MCS and above recognition sensitivity, and VS recognition specificity are calculated under different thresholds. Provided that the effective sample retention rate is not lower than a preset requirement, the threshold with the highest comprehensive evaluation index is selected as the system default threshold.

[0201] In one embodiment, the permanent bad sector proportion threshold ranges from 15% to 35%, the EEG data quality threshold ranges from 0.50 to 0.75, the transient bad segment amplitude threshold ranges from ±150μV to ±300μV, the low confidence threshold for the classification probability ranges from 0.40 to 0.60, and the threshold for the difference between the highest and second highest probabilities ranges from 0.05 to 0.20. The system saves the final thresholds and their corresponding validation results as part of the model metadata.

[0202] Example 2: EEG Analysis System for Consciousness Disorders Based on Brain Functional Partitioning like Figure 2 As shown, this embodiment provides an EEG analysis system for consciousness disorders based on brain functional partitioning, including an EEG data acquisition / import module 100, a signal preprocessing module 200, a brain functional partitioning mapping module 300, a partitioned EEG feature extraction module 400, a functional connectivity and network analysis module 500, a classification assessment module 600, a reference parameter library and scoring module 700, and a result visualization and report generation module 800.

[0203] The EEG data acquisition / import module 100 is used to receive real-time acquired multi-channel EEG data or import offline EEG files, and read lead names, sampling frequencies, electrode impedances and acquisition time information.

[0204] The signal preprocessing module 200 is used to perform filtering, rereference, artifact removal, permanent bad sector identification, bad sector repair, transient bad segment removal, and data quality assessment. Among them, permanent bad sector identification is used to handle continuous anomalies at the lead level; transient bad segment removal is used to handle short-term anomalies at the time window level, and the two use different judgment objects and different removal rules.

[0205] The brain functional partition mapping module 300 is used to calculate the basic spatial mapping coefficient of each electrode corresponding to each brain functional partition based on the electrode name, spatial coordinates and functional partition mapping table, and calculate the final partition weight in combination with the lead quality factor to generate representative EEG signals for each brain functional partition.

[0206] The regional EEG feature extraction module 400 is used to independently extract 25-dimensional features for each brain functional region, including frequency domain features, time domain statistical features, complexity features, and intra-regional synchronicity features, and generate... Dimensional partition feature matrix.

[0207] The Functional Connectivity and Network Analysis Module 500 is used to calculate 45-dimensional undirected connectivity features, 30-dimensional directional connectivity features, and 23-dimensional undirected network topology features between the six brain functional regions, and form a 248-dimensional fixed model input vector.

[0208] The classification evaluation module 600 is used to convert the 248-dimensional model input vector into a first-level model input vector and a second-level model input vector according to the fixed-dimensional feature screening rules, and output the four probabilities of VS, MCS-, MCS+ and EMCS and the final classification result through the hierarchical classification model.

[0209] The reference parameter library and scoring module 700 are used to construct a reference parameter library based on labeled samples, calculate the characteristic distribution parameters of each brain functional region, and output the functional retention score of each brain functional region based on the percentile normalization method.

[0210] The Results Visualization and Report Generation Module 800 is used to generate zonal scoring radar charts, connectivity matrix heatmaps, brain functional network topology diagrams, and standardized assessment reports.

[0211] Example 3: Computer Equipment and Storage Media This embodiment provides a computer device, including a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the EEG analysis method for consciousness disorders based on brain functional partitioning as described in Embodiment 1.

[0212] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it performs the following steps: Acquire multi-channel EEG data; The EEG data is filtered, rereferenced, artifact removed, permanent bad sectors identified, bad sectors repaired, transient bad segments removed, and data quality assessed. Establish a mapping relationship between EEG electrodes and brain functional areas based on electrode names or electrode spatial coordinates; Representative EEG signals for each brain functional region are generated by combining electrode spatial mapping coefficients and lead quality factors. 25-dimensional EEG features were independently extracted from each brain functional region to form 3D partition feature matrix; The functional connectivity strength between brain functional regions is calculated using brain functional regions as nodes, and a region-level brain functional connectivity network is constructed. A 248-dimensional model input vector is formed based on 150-dimensional partition features, 45-dimensional undirected connectivity features, 30-dimensional directional connectivity features, and 23-dimensional undirected topological features; Generate the first-level model input vector and the second-level model input vector according to the fixed-dimensional feature selection rules; The probability of four classes—VS, MCS-, MCS+, and EMCS—is output through a first-level consciousness existence classification model and a second-level consciousness level classification model. The functional retention score of each brain functional region was calculated based on a reference parameter library; Generate visual reports.

[0213] Through the above implementation methods, this invention does not simply perform an overall average analysis of whole-brain EEG signals. Instead, it first converts the EEG signals into representative signals of functional brain functional regions, then extracts the features of each region and the inter-regional connectivity features, and finally outputs a classification of consciousness impairment and a score of functional preservation. Simultaneously, this invention clearly distinguishes between permanent bad channels and transient bad segments, fixes the dimensions of the regional feature matrix, clarifies the computational boundaries of directed connections and undirected topological features, avoids the problem of empty variables in hierarchical classification models, and discloses the construction method of the reference parameter library and the fixed feature selection rules, thereby enabling those skilled in the art to reproduce the complete brain functional region analysis process.

[0214] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. All 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 analyzing electroencephalograms (EEGs) of consciousness disorders, characterized in that, include: Acquire multichannel EEG data of the subject to be evaluated, and preprocess and assess the data quality of the multichannel EEG data; A mapping relationship between EEG electrodes and multiple brain functional regions is established based on EEG lead information. These multiple brain functional regions are partitions obtained according to the functional differences of brain regions. For each brain functional region, the final weight of the corresponding electrode in the brain functional region is calculated based on the electrode spatial mapping coefficient and the lead quality factor. The preprocessed multi-channel EEG signals are then weighted and fused according to the final weight to generate representative EEG signals for each brain functional region. Multimodal EEG features were extracted from representative EEG signals of each brain functional region to form a region feature matrix; Using brain functional regions as nodes, calculate the undirected functional connectivity features and directed directional connectivity features between regions. Construct an undirected connectivity matrix based on the undirected functional connectivity features between regions, and extract network topology features from the undirected connectivity matrix. The partition feature matrix, the undirected functional connectivity features of the partition, the directed directional connectivity features and the network topology features are concatenated into a fixed-dimensional input vector. After feature normalization and fixed-dimensional feature filtering, the vector is input into a hierarchical classification model, and the output is the EEG typing probability result used for auxiliary assessment of consciousness disorders. The functional retention scores for each brain functional region are calculated based on a reference parameter library, and visual analysis results and standardized reports are generated.

2. The method according to claim 1, characterized in that, The preprocessing includes filtering, rereference, bad sector identification, artifact removal, bad sector repair, and bad segment removal operations; The bad sector identification is permanent bad sector identification, used to identify persistent abnormalities at the lead level, and the bad segment removal is instantaneous bad segment removal, used to identify short-term abnormalities at the time window level. When performing transient bad segment removal, the judgment is based solely on the permanently valid leads and the temporary partition quality control signal generated by the permanently valid main leads, so as to exclude the influence of repaired permanently bad leads on the determination of transient bad segments.

3. The method according to claim 2, characterized in that, The identification of permanent bad sectors includes: if a lead meets at least one of the following conditions: the continuous flat time exceeds a preset time, the absolute amplitude exceeds a preset ratio, the lead variance is abnormal, the average correlation coefficient with other leads is lower than a preset threshold, or the abnormal time window ratio exceeds a preset ratio, then the lead is marked as a permanent bad sector. If the proportion of permanent bad sectors to the total number of leads exceeds the preset bad sector proportion threshold, the subsequent analysis will be terminated and a message indicating insufficient valid leads will be output. Leads marked as having permanent bad sectors are repaired using weighted interpolation of adjacent leads, and the repaired lead signals are used as low-confidence compensation signals.

4. The method according to claim 1, characterized in that, The lead quality factor is determined by the impedance quality factor, artifact quality factor, variance stability factor, and permanent bad sector penalty factor. When a permanent effective primary lead exists in a certain brain functional area, the weight of the permanent bad passage repair lead in the brain functional area is reset to zero. When a brain functional region does not have a permanently effective primary lead, a permanently damaged lead is allowed to participate in the generation of representative EEG signals for the brain functional region with a weight lower than that of a permanently effective lead.

5. The method according to claim 1, characterized in that, The undirected functional connectivity features of the subdivisions include the phase lock value, coherence and weighted phase lag index between each pair of the six brain functional regions, forming undirected connectivity features; The directional connectivity features include the strength of directional connectivity between each pair of the six brain functional regions, forming directional connectivity features; The network topology features are calculated solely based on the undirected connectivity matrix, including global efficiency, average clustering coefficient, feature path length, small-world property, modularity, node strength, node degree, and betweenness centrality of each brain functional region, thus forming the undirected network topology features.

6. The method according to claim 5, characterized in that, The fixed-dimensional input vector is formed by concatenating partitioned EEG features, undirected connectivity features, directional connectivity features, and undirected network topology features. The fixed-dimensional feature selection includes: robustly normalizing the fixed-dimensional input vector, handling missing values, deleting low-variance features, deleting redundant features whose absolute correlation coefficient is greater than a preset threshold, and sorting features based on mutual information values, while retaining the first-level model input feature set and the second-level model input feature set respectively.

7. The method according to claim 1, characterized in that, The hierarchical classification model includes a first-level consciousness existence classification model and a second-level consciousness level classification model; during the reasoning phase, regardless of whether the consciousness existence probability output by the first-level consciousness existence classification model is lower than the first threshold, the second-level consciousness level classification model is executed. The reference parameter library stores the median, interquartile range, 5th percentile, and 95th percentile of the feature distribution of reference samples. Based on the 5th percentile and 95th percentile, percentile normalization is performed on the regional EEG features and / or regional network features of the subject to be evaluated, and the functional retention score of each brain functional region is calculated by weighting multiple regional sub-scores.

8. A brainwave analysis system for consciousness disorders, characterized in that, include: The EEG data acquisition / import module is used to acquire multi-channel EEG data of the subject to be evaluated. The signal preprocessing module is used to filter, rereference, identify bad channels, remove artifacts, repair bad channels, remove bad segments, and evaluate data quality for the multi-channel EEG data. The brain functional area mapping module is used to establish the mapping relationship between EEG electrodes and multiple brain functional areas, and calculate the final weight based on the electrode spatial mapping coefficient and lead quality factor to generate representative EEG signals for each brain functional area. The regional EEG feature extraction module is used to extract multimodal EEG features from representative EEG signals of each brain functional region and form a regional feature matrix. The Functional Connectivity and Network Analysis module is used to calculate the undirected functional connectivity characteristics, directed directional connectivity characteristics, and network topology characteristics of the intervals. The classification assessment module is used to input a fixed-dimensional input vector formed by the partition feature matrix, the undirected functional connectivity features of the partition, the directed directional connectivity features and the network topology features into the hierarchical classification model, and output the EEG classification probability results for auxiliary assessment of consciousness disorders. Reference parameter library and scoring module are used to calculate the functional retention score of each brain functional area; as well as The Results Visualization and Report Generation module is used to generate visualized analysis results and standardized reports.

9. The system according to claim 8, characterized in that, The brain functional partition mapping module includes an electrode spatial coordinate reading unit, a functional partition mapping table, a partition spatial weight calculation unit, a lead quality weight calculation unit, and a partition representative EEG signal generation unit. The regional representative EEG signal generation unit is used to perform weighted fusion of preprocessed multi-channel EEG signals based on regional spatial weights and lead quality weights to generate representative EEG signals for the prefrontal lobe region, motor cortex region, somatosensory cortex region, temporal lobe region, occipital lobe region, and posterior parietal lobe region.

10. The system according to claim 9, characterized in that, The partitioned EEG feature extraction module is used to generate a partitioned feature matrix; The functional connection and network analysis module is used to generate undirected connection features, directional connection features, and undirected network topology features; The classification evaluation module is used to concatenate the partitioned EEG features, undirected connectivity features, directional connectivity features, and undirected network topology features after the partitioned feature matrix is ​​expanded into a fixed-dimensional input vector, and output the four classification probabilities of VS, MCS-, MCS+, and EMCS based on the fixed-dimensional feature screening and hierarchical classification model.

11. A storage medium, characterized in that, The storage medium stores a computer program executable by a processor, which, when executed, implements the EEG analysis method for consciousness disorders according to any one of claims 1 to 7.

Images