Method, system and device for identifying patient consciousness disorder based on a distributed perception automatic aggregation mechanism

Abstract

The application relates to the technical field of patient consciousness disorder identification, and provides a patient consciousness disorder identification method, system and device based on a distribution perception automatic aggregation mechanism, which comprises the following steps: acquiring multi-lead electroencephalogram signals of a patient to be identified for a consciousness disorder; performing pretreatment and segment division processing on the multi-lead electroencephalogram signals to obtain N electroencephalogram segments; adopting a deep learning encoder to extract deep feature representation of each electroencephalogram segment to obtain N feature vectors corresponding to the N electroencephalogram segments; adopting a distribution perception pooling mechanism to statistically process the N feature vectors to obtain a distribution vector; adopting an interpretable classifier to identify the consciousness disorder of the patient to be identified for a consciousness disorder according to the distribution vector to obtain a classification result corresponding to the patient to be identified for a consciousness disorder; the deep learning encoder and the interpretable classifier are connected in cascade, a constraint threshold optimization mechanism is introduced to update a decision threshold of classification identification in the interpretable classifier, and high-precision and interpretable consciousness disorder identification is realized.

Description

Technical Field

[0001] This application relates to the field of patient consciousness disorder recognition technology, and in particular to a method, system and device for patient consciousness disorder recognition based on a distributed perception automatic aggregation mechanism. Background Technology

[0002] Disorders of consciousness (DoC) are among the most serious clinical consequences of severe brain injury. The main categories of DoC include unresponsive wakefulness syndrome (UWS, formerly known as vegetative state) and minimally conscious state (MCS). Accurately differentiating between UWS and MCS directly determines the patient's treatment plan (such as arousal medications, neuromodulation, rehabilitation strategies) and prognosis, and is a core challenge in neurocritical care and rehabilitation medicine.

[0003] Currently, the clinical identification of patients with altered consciousness mainly relies on behavioral scale assessments, including the Coma Recovery Scale-Revised (CRS-R) and the Galveston Orientation and Adaptation Questionnaire (GOAQQ). The CRS-R is widely recognized as the gold standard, with rehabilitation medicine professionals conducting item-by-item behavioral tests. However, using behavioral scales to identify altered consciousness has three limitations: First, significant inter-assessment variability—the consistency of scores among different rehabilitation physicians is limited, especially in differentiating between UWS and MCS with an accuracy rate of only 63-77%; second, assessment limitations—some critically ill patients cannot cooperate with behavioral testing due to airway management (endotracheal intubation), muscle relaxation, or motor impairment, leading to false negatives; and third, difficulties in dynamic monitoring—CRS-R assessment is time-consuming and dependent on professionals, making it unsuitable for high-frequency bedside monitoring.

[0004] With the development of brain-computer interface technology, electroencephalography (EEG) has become a substitute / supplementary means for the identification of DoC (Disorders of Consciousness) due to its non-invasive, high-frequency acquisition, and low-cost characteristics. EEG captures patterns of activity within the brain (such as responsiveness, spectral characteristics, and functional connectivity), providing objective neurophysiological evidence when patients are unable to cooperate with behavioral tests. In recent years, EEG-based DoC identification methods have evolved from traditional machine learning methods based on handcrafted features (such as spectral power, functional connectivity + SVM) to end-to-end methods based on deep learning. However, existing methods for identifying the state of patients with disorders of consciousness based on EEG signals (especially for the classification of Unresponsive Wake-Up Syndrome (UWS) and Minimal Consciousness State (MCS) still suffer from limitations in clinical application due to factors such as individual patient differences, time-varying EEG characteristics, and uneven data distribution. These limitations include insufficient patient-level feature representation, lack of interpretability in the model decision-making process, and difficulty in balancing clinical sensitivity and generalization performance. Existing methods for identifying the state of patients with consciousness disorders based on EEG signals can be divided into methods based on handcrafted features and shallow machine learning, and end-to-end classification methods based on deep learning.

[0005] The method based on handcrafted features and shallow machine learning preprocesses EEG signals, then extracts manually designed features as patient state representations, and combines them with a traditional classifier to determine the state of consciousness. Specifically, it extracts time-domain, frequency-domain, and nonlinear features from the EEG signals, such as power spectral density (PSD), fractal dimension, and sample entropy, or further constructs functional connectivity features of brain regions (e.g., phase-locked value (PLV), mutual information (MI)). Then, it uses classification models such as Support Vector Machine (SVM) and Random Forest to classify UWS and MCS, thus identifying the patient's state of consciousness. The workflow of this method is: EEG preprocessing → handcrafted feature extraction (time-domain / frequency-domain / nonlinear features) → feature selection (e.g., LASSO) → shallow classifier (SVM / random forest) → decision output. However, this method suffers from limitations: the handcrafted feature design relies on prior experience, making it difficult to fully cover the complex discrimination patterns in EEG signals; furthermore, the feature extraction and classification processes are independent, preventing joint optimization, thus limiting its adaptability and generalization performance in complex clinical scenarios.

[0006] In recent years, deep learning methods have gradually become the main technical approach for EEG identification of disorders of consciousness. Existing deep learning-based end-to-end patient-level classification methods first input multiple EEG segments (epochs) from a single patient into a deep network (e.g., convolutional neural networks or recurrent neural networks) to obtain corresponding deep feature representations (embeddings). Then, methods such as mean pooling are used to fuse the features from multiple segments, forming a patient-level representation, and the final classification result is output through a fully connected layer or classifier. Specifically, the workflow of this deep learning-based end-to-end patient-level classification method is: EEG preprocessing → deep encoding network extracting epoch-level features → patient-level feature fusion → classifier → patient state output. Although this deep learning-based end-to-end patient-level classification method can achieve the integration of feature learning and classification processes, it still has the following shortcomings: First, the patient-level feature fusion method is difficult to fully express the internal EEG distribution characteristics of patients. Because there are often significant fluctuations and heterogeneities between different EEG segments, their changes are affected by the patient's awakening state, attention state, and dynamic changes in neural activity. The end-to-end patient-level classification method based on deep learning usually uses mean aggregation to aggregate features from multiple epochs, retaining only the overall central information, which is difficult to reflect the fluctuation characteristics and statistical distribution relationship between different EEG segments. This type of fusion method may lead to the loss of some discriminative information, thereby affecting classification performance and thus affecting the accuracy and stability of patient state identification. Second, the interpretability of the model decision-making process is insufficient. Existing end-to-end patient-level classification methods based on deep learning adopt an end-to-end structure. Their classification criteria are usually difficult to directly map to understandable EEG features or decision logic, making it difficult to provide traceable feature support for the classification results. This limits the clinical interpretability and application credibility of the model results, and makes it difficult to balance the needs of clinical sensitivity with the model's generalization ability. Third, existing end-to-end patient-level classification methods based on deep learning generally use overall classification performance or overall accuracy as optimization targets. When the class distribution is uneven and there are large differences among patients, it is difficult to effectively constrain the high sensitivity requirements that are of clinical concern, thus affecting the stability and applicability of the model in real patient groups.

[0007] In summary, while existing deep learning-based EEG identification methods can automatically extract deep EEG features, their patient-level fusion strategies are still relatively simple and cannot fully utilize the patient's internal EEG distribution information. They also suffer from insufficient model interpretability and limited clinical adaptability. Summary of the Invention

[0008] This application provides a method, system, and device for identifying patient consciousness disorders based on a distributed sensing automatic aggregation mechanism, which addresses the technical problems of insufficient patient-level feature expression, lack of interpretability of classification decisions, and limited generalization performance under high sensitivity requirements in existing EEG identification methods for consciousness disorders.

[0009] To achieve the above objectives, this application provides the following technical solution: On the one hand, a method for identifying patient consciousness disorders based on a distributed sensing automatic aggregation mechanism is provided, including the following steps: Acquire multi-lead EEG signals from patients to be identified as having impaired consciousness, and perform preprocessing and segmentation on the multi-lead EEG signals to obtain N EEG segments; A deep learning encoder is used to extract deep feature representations for each of the N EEG segments, resulting in N feature vectors corresponding to the N EEG segments. The N feature vectors are processed using a distribution-aware automatic aggregation mechanism to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state; An interpretable classifier is used to identify the consciousness disorder by the distribution vector of the patient to be identified, and the classification result of the corresponding patient to be identified is obtained. The deep learning encoder is cascaded with the interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification and recognition in the interpretable classifier.

[0010] Optionally, each feature vector includes D-dimensional features. A distribution-aware automatic aggregation mechanism is used to process the N feature vectors to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state, including: The mean and standard deviation corresponding to each dimension feature are calculated based on the number N of the feature vectors and each dimension feature of all the feature vectors; the percentile corresponding to each dimension feature of all the feature vectors is calculated using the Percentile function based on multiple preset threshold values ​​to obtain the percentile corresponding to each dimension feature at each preset threshold. Based on the mean, standard deviation and all percentages of each dimension feature, a basic statistical feature pool is determined corresponding to each dimension feature, the basic statistical feature pool including K basic feature data; Based on each basic feature data in all the basic statistical feature pools, determine the coefficient of variation corresponding to the basic feature data; The basic statistical feature pools of each dimensional feature are weighted and fused according to all the coefficients of variation to obtain the distribution manifold vector corresponding to each dimensional feature; All the aforementioned distribution manifold vectors are concatenated according to the order of their dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

[0011] Optionally, the constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; The deep learning encoder and the distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, the interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision threshold is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold; the classification accuracy corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and the sensitivity, all the candidate decision thresholds are filtered to obtain a better decision threshold that satisfies the sensitivity constraint threshold; Based on the classification accuracy, all the better decision thresholds are filtered using the argmax function to obtain the decision thresholds for classification and recognition in the interpretable classifier.

[0012] Optionally, before using the interpretable classifier to identify disorders of consciousness based on each candidate decision threshold and the patient distribution vectors of the M patients, the method for identifying the state of a patient with a disorder of consciousness further includes: data augmentation of the sample data for each patient; the content of data augmentation of the sample data for each patient includes: Based on the patient distribution vector of each patient, an adaptive synthetic sampling method is used to amplify the corresponding sample data to obtain the first training sample corresponding to the patient. Obtain random samples corresponding to the consciousness impairment category of each patient's sample data, and process the sample data and random samples of each patient using a linear interpolation method to obtain a second training sample corresponding to each patient; The first training sample and the second training sample of each patient are merged to obtain the corresponding patient's training sample as new sample data.

[0013] Optionally, an interpretable classifier is used to identify the patient with the disorder of consciousness using the distribution vector, and the classification results for the corresponding patient with the disorder of consciousness are as follows: The distribution vector of the patient to be identified as having a disorder of consciousness is input into the interpretable classifier, and the interpretable classifier outputs the original decision score of the patient to be identified as having a disorder of consciousness. Based on the original decision score, the predicted probability of the microconscious state of the patient to be identified as having a disturbance of consciousness is obtained. The classification result of the patient to be identified as having a disturbance of consciousness is obtained by comparing the predicted probability of the micro-consciousness state with the decision threshold. The classification results include plant state or minimal consciousness state.

[0014] Optionally, the patient consciousness disorder identification method based on the distribution-aware automatic aggregation mechanism further includes: outputting the classification results, the output of which includes the consciousness disorder category of the patient to be identified, the predicted probability of the consciousness disorder category, and the contribution value of each distribution feature in the distribution vector.

[0015] On the other hand, a patient consciousness disorder identification system based on a distribution-aware automatic aggregation mechanism is provided, including: a data acquisition module, a feature extraction module, a distribution vector module, and a disorder identification module; The data acquisition module is used to acquire multi-lead EEG signals of patients to be identified as having impaired consciousness, and to preprocess and segment the multi-lead EEG signals to obtain N EEG segments. The feature extraction module is used to perform deep feature representation extraction on each of the EEG segments using a deep learning encoder, to obtain N feature vectors corresponding to the N EEG segments. The distribution vector module is used to process the N feature vectors using a distribution-aware automatic aggregation mechanism to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state. The obstacle identification module is used to identify the consciousness disorder by using an interpretable classifier on the distribution vector of the patient to be identified as having a consciousness disorder, and to obtain the classification result of the corresponding patient to be identified as having a consciousness disorder. The deep learning encoder is cascaded with the interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification and recognition in the interpretable classifier.

[0016] Optionally, each feature vector includes D dimensional features, and the distribution vector module includes a calculation submodule, a feature pool determination submodule, a coefficient determination submodule, a weighted fusion submodule, and a splicing and construction submodule; The calculation submodule is used to calculate the mean and standard deviation corresponding to each dimension feature based on the number N of the feature vectors and each dimension feature of all the feature vectors; and to calculate the percentile corresponding to each dimension feature of all the feature vectors using the Percentile function based on multiple preset thresholds of different values, so as to obtain the percentile corresponding to each dimension feature under each preset threshold. The feature pool determination submodule is used to determine the basic statistical feature pool corresponding to each of the dimensional features based on the mean, the standard deviation and all the percentages of each dimensional feature. The basic statistical feature pool includes K basic feature data. The coefficient determination submodule is used to determine the coefficient of variation corresponding to each basic feature data in all the basic statistical feature pools. The weighted fusion submodule is used to perform weighted fusion processing on the basic statistical feature pool of each dimension feature according to all the coefficients of variation, so as to obtain the distribution manifold vector corresponding to each dimension feature; The splicing submodule is used to splice all the distribution manifold vectors in the order of dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

[0017] Optionally, the constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; The deep learning encoder and the distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, the interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision threshold is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold; the classification accuracy corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and the sensitivity, all the candidate decision thresholds are filtered to obtain a better decision threshold that satisfies the sensitivity constraint threshold; Based on the classification accuracy, all the better decision thresholds are filtered using the argmax function to obtain the decision thresholds for classification and recognition in the interpretable classifier.

[0018] On the other hand, a terminal device is provided, including a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the patient consciousness impairment identification method based on the distribution-aware automatic aggregation mechanism described above, according to the instructions in the program code.

[0019] This invention discloses a method, system, and device for identifying patient consciousness disorders based on a distribution-aware automatic aggregation mechanism. The method includes acquiring multi-lead EEG signals from a patient to be identified as having a consciousness disorder; preprocessing and segmenting the multi-lead EEG signals to obtain N fixed-length EEG segments; extracting deep feature representations for each EEG segment using a deep learning encoder to obtain N feature vectors corresponding to the N EEG segments; processing the N feature vectors using a distribution-aware automatic aggregation mechanism to obtain a distribution vector representing the heterogeneity of the patient's brain neural state; and using an interpretable classifier to identify the consciousness disorder based on the distribution vectors of the patient to be identified, obtaining the classification result for the corresponding patient. The deep learning encoder and the interpretable classifier are cascaded, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification in the interpretable classifier.

[0020] As can be seen from the above technical solutions, this application has the following advantages: The patient consciousness disorder identification method based on the distribution-aware automatic aggregation mechanism processes the acquired multi-lead EEG signals of patients through a deep learning encoder and a distribution-aware pooling mechanism to obtain a distribution vector characterizing the distribution features of the patient's EEG. Then, an interpretable classifier is used to identify the consciousness disorder by performing the distribution vector of the patient to be identified, thereby obtaining the classification result of the corresponding patient to be identified. This solves the technical problems of insufficient patient-level feature expression, lack of interpretability of classification decisions, and limited generalization performance under high sensitivity requirements in existing EEG identification methods for consciousness disorders.

[0021] This patient consciousness disorder identification system based on a distribution-aware automatic aggregation mechanism fully preserves the patient's internal EEG variation information through a distribution-aware automatic aggregation mechanism in the data acquisition module, feature extraction module, distribution vector module, and disorder identification module. An interpretable classifier provides traceable decision-making, and a constraint threshold optimization mechanism optimizes to meet clinical sensitivity requirements, achieving high-precision, interpretable, and clinically applicable identification of consciousness disorders. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the steps of the patient consciousness disorder identification method based on a distributed sensing automatic aggregation mechanism as described in the embodiments of this application. Figure 2 This is a flowchart illustrating a patient consciousness disorder identification method based on a distributed sensing automatic aggregation mechanism according to another embodiment of this application. Figure 3 This is a schematic diagram of the framework of a patient consciousness disorder recognition system based on a distributed sensing automatic aggregation mechanism, as described in another embodiment of this application. Figure 4 This is a schematic diagram of the terminal device described in an embodiment of this application. Detailed Implementation

[0024] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the description of the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0026] In the embodiments of this application, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application according to the specific circumstances.

[0027] Patent terminology used in this application: Multi-channel EEG refers to the collection of electrophysiological activities of different brain regions simultaneously acquired through multiple electrodes on the scalp. Compared with single-channel EEG, it can provide spatial distribution information and is the core data for locating brain functional abnormalities and studying brain network mechanisms.

[0028] LightGBM (Light Gradient Boosting Machine) is a high-performance gradient boosting decision tree (GBDT) machine learning framework designed for handling large-scale data. Through core technologies such as histogram algorithm, one-sided gradient sampling, and mutually exclusive feature binding, it significantly improves training speed and reduces memory consumption while maintaining model accuracy. It supports parallel and distributed computing and can directly process categorical features without one-hot encoding.

[0029] The Percentile() function is a statistical function used to calculate the k-th percentile of a dataset. It indicates that after sorting the data from smallest to largest, k% of the data are less than or equal to this value. It is often used to analyze data distribution, identify outliers, or set thresholds (such as the median, which is the 50th percentile).

[0030] A deep learning encoder is a core component in a neural network that maps high-dimensional input data to low-dimensional latent representations. It typically exists as part of an autoencoder, Transformer, or Variational Autoencoder (VAE).

[0031] `argmax` is a function that evaluates a function to its arguments (set). When we have another function y = f(x), if we have the result x0 = argmax(f(x)), it means that when x = x0, we get the maximum value of f(x) within its range. If there are multiple points that give f(x) the same maximum value, then the result of `argmax(f(x))` is a set of points.

[0032] Independent component analysis (ICA) is a statistical multivariate signal processing method whose core function is to separate independent components in a signal.

[0033] The flatten function is a tool used to flatten multidimensional data into one dimension.

[0034] Distribution-Aware Auto-Pooling (DAAP) automatically aggregates fragment-level features into a compact distribution vector representing the heterogeneity of patients' neural states by adaptively learning the weights of statistical dimensions.

[0035] Individual EEG Distribution Manifold: A low-dimensional parametric representation that fully preserves the statistical picture of a patient's internal EEG segments by constructing a multidimensional moment-quantile joint descriptor and introducing adaptive weights.

[0036] Constrained threshold optimization mechanism: This is a strategy that dynamically optimizes the decision threshold while satisfying sensitivity constraints, aiming to minimize the risk of missed diagnosis of microconsciousness state (MCS).

[0037] Heterogeneity refers to the differences that exist between different cells, tissues, or organs in an organism, and these differences may affect their function and behavior.

[0038] This application provides a method, system, and device for identifying patient consciousness disorders based on a distributed sensing automatic aggregation mechanism. It addresses the technical problems of insufficient patient-level feature expression, lack of interpretability in classification decisions, and limited generalization performance under high sensitivity requirements in existing EEG identification methods for consciousness disorders. This method, system, and device are used to classify and identify unresponsive wakefulness syndrome (UWS, also known as a vegetative state) and minimally conscious state (MCS, also known as a minimally conscious state). Through the distributed sensing automatic aggregation mechanism, it fully preserves the patient's internal EEG variation information; an interpretable classifier provides traceable decisions; and a constraint threshold optimization mechanism optimizes to meet clinical sensitivity requirements, achieving high-precision, interpretable, and clinically applicable identification of consciousness disorders.

[0039] Example 1: Figure 1 This is a flowchart illustrating the steps of the patient consciousness impairment identification method based on a distributed sensing automatic aggregation mechanism as described in an embodiment of this application. Figure 2 This is a flowchart illustrating the patient consciousness disorder identification method based on a distributed sensing automatic aggregation mechanism as described in an embodiment of this application.

[0040] like Figure 1 and Figure 2 As shown in the figure, this application provides a method for identifying patient consciousness disorders based on a distributed sensing automatic aggregation mechanism, including the following steps: S1. Acquire multi-lead EEG signals from patients to be identified as having impaired consciousness. Perform preprocessing and segmentation on the multi-lead EEG signals to obtain N fixed-length EEG segments.

[0041] It should be noted that in the process of identifying patients with impaired consciousness based on the distributed perception automatic aggregation mechanism, the multi-lead EEG signals of the patients to be identified with impaired consciousness are obtained from the database in step S1. Then, the obtained multi-lead EEG signals are preprocessed and segmented using preset rules to obtain N fixed-length EEG segments, which provide data for the subsequent identification of the state of impaired consciousness of the patients to be identified.

[0042] In this embodiment, acquiring multi-lead EEG signals from a patient requiring identification of altered consciousness can also be understood as using an EEG acquisition device to collect data from the patient and employing the international 10–20 system for electrode placement. The acquisition of multi-lead EEG signals can be performed in a resting state. The sampling rate on the EEG acquisition device is set to 250Hz, and the duration of a single sampling session is no less than 3 minutes.

[0043] It should be noted that EEG acquisition devices can be mobile terminals such as computers and industrial control computers.

[0044] To improve the stability and effectiveness of subsequent feature extraction and classification modeling of interpretable classifiers, and to eliminate the interference of environmental noise, electromyography artifacts, and baseline drift on subsequent deep learning feature extraction, a series of standardized preprocessing steps are performed on the multi-lead EEG signals. In this embodiment, the preprocessing and segmentation of the multi-lead EEG signals to obtain N fixed-length EEG segments includes: sequentially performing band filtering, notch filtering, artifact removal, segmentation, and bad segment removal on the multi-lead EEG signals.

[0045] It should be noted that band filtering refers to limiting the frequency band of multi-lead EEG signals, retaining the effective EEG frequency band of 0.5~45Hz as the first processed signal to suppress baseline drift and high-frequency noise. Notch filtering removes power frequency (50Hz) and environmental electromagnetic interference signals from the first processed signal to obtain the second processed signal. Artifact removal refers to using artifact detection, signal separation methods, independent component analysis (ICA), or regression methods to remove non-brain-origin artifact interference components such as electrooculography (EOG), electromyography (EMG), and electrocardiography (ECG) from the second processed signal to obtain the third processed signal, ensuring the signal-to-noise ratio. Segment segmentation refers to dividing the pre-processed continuous multi-lead EEG signal (such as the third processed signal) into multiple fixed-length EEG segments according to a preset time window, and multiple EEG segments constitute the EEG sample set of the corresponding patient. X Segment segmentation can also be understood as dividing the pre-processed, continuous third-processed signal into independent EEG segments (Epochs) of fixed length (e.g., 6 seconds). Bad segment removal is performed on the EEG sample set. X Abnormal segments with amplitude values ​​exceeding a preset threshold are removed to ensure data quality.

[0046] In this embodiment, each EEG segment contains 1500 sampling points, corresponding to a length of approximately 6 seconds. Assuming a single patient yields N EEG segments after preprocessing, an EEG sample set is formed. X ,Right now ;in, C The number of brainwave channels can be 19. T The number of sampling points corresponding to a single EEG segment is 1500. In the real number space, the surface data are continuous real-valued signals. x i For the first i A fragment of brainwave.

[0047] In other embodiments, the patient consciousness disorder identification method based on the distribution-aware automatic aggregation mechanism can preprocess multi-lead EEG signals using methods such as independent component analysis and regression analysis. This method can segment the preprocessed multi-lead EEG signals using segmentation and baseline correction, as well as bad electrode and bad segment removal / interpolation. For example, continuous multi-lead EEG signals are segmented into EEG segments of independent data of a specified duration according to event markers. Baseline correction is then used to eliminate initial potential shifts between different trials. All EEG segments are screened using spectral analysis and temporal amplitude checks to identify poorly contacted bad electrodes, which are then repaired using spherical interpolation. EEG segments severely contaminated by noise and unrepairable are directly removed; the proportion of bad segments is typically controlled to be within 10% of the total trials.

[0048] S2. A deep learning encoder is used to extract deep feature representations for each EEG segment, resulting in N feature vectors corresponding to the N EEG segments.

[0049] It should be noted that step S2 involves taking the set of EEG samples obtained in step S1. X Each EEG segment x i The deep learning feature encoding network, which is fed into the deep learning encoder, performs end-to-end feature encoding on the input EEG segments to fully extract the spatiotemporal discriminative patterns in the EEG signals of each segment. This is done using a convolutional neural network structure based on an improved version of EEGNet, and outputs a fixed-dimensional low-dimensional embedding representation. e i Its input representation The overall feature extraction process can be represented as: ; In the formula, f () indicates a feature extraction network.

[0050] In this embodiment, the deep learning encoder includes a spatial-temporal joint feature extraction module, a temporal dynamic modeling module, and a channel attention enhancement module, which are cascaded in sequence.

[0051] It should be noted that the spatial-temporal joint feature extraction module captures the local temporal waveform features of a single lead through temporal convolution, and then models the spatial topological relationships between leads through depthwise separable convolution, obtaining a temporal feature map with spatial context information. The temporal dynamic modeling module uses a multi-layer convolutional neural network to further perform temporal dynamic modeling of spatial features, capturing long-range temporal dependencies, and reduces temporal resolution through averaging and aggregation operations to enhance feature stability. The channel attention enhancement module introduces a channel attention mechanism, generating weight coefficients for each channel through global averaging and aggregation and fully connected layers, and adaptively recalibrating the feature map composed of temporal features. This channel attention enhancement module can automatically enhance sensitive brain regions related to consciousness state discrimination and suppress noisy channels, thereby outputting a more robust feature representation. Finally, through flattening and mapping with fully connected layers, each EEG segment is encoded into a fixed-dimensional feature vector. N feature vectors constitute a feature vector set.

[0052] In other embodiments, the convolutional neural network structure based on EEGNet may include a spatial feature extraction module, a temporal feature extraction module, and a feature enhancement module connected in sequence.

[0053] In this embodiment, the spatial feature extraction module uses temporal convolution and depthwise separable convolution to model the spatial features of EEG signals from EEG segments, specifically: By performing time-dimensional local feature extraction on the EEG signal of the EEG segment through the first two-dimensional convolution, the local frequency band information and the first feature of the short-time waveform structure of the EEG signal in the EEG segment are obtained. A depth-separable switch is used to extract spatial features from the first feature of each EEG segment, thus obtaining the spatial features corresponding to each EEG segment.

[0054] It should be noted that the first two-dimensional convolution is a temporal convolution, and the depthwise separable convolution is a depthwise separable convolution. The first two-dimensional convolution contains 8 convolutional kernels, each with a size of 1×64, and is symmetrically padded in the temporal dimension (padding=32). The depthwise separable convolution contains 16 spatial convolutional kernels, each with a size of C=19×1, and uses grouped convolution (groups=8). The expression for extracting the first feature using the first two-dimensional convolution is: ; The expression for extracting spatial features using depthwise separable convolution is: ; ; In the formula,W t The temporal convolution kernel for the first two-dimensional convolution. This represents the convolution operation. BN This represents batch normalization, and σ() is a nonlinear activation function (ELU). In order to be with the first i The first feature corresponding to each EEG segment W s For spatial convolution kernels, and S i All are related to the first i Spatial features corresponding to individual EEG segments.

[0055] In other embodiments, the deep learning feature encoding network of the deep learning encoder may also employ other deep learning encoding networks (such as Transformer, Recurrent Neural Network (RNN), Self-Supervised Representation Learning Model, etc.) to complete the learning of EEG signal representations and generate feature embeddings.

[0056] In this embodiment, the temporal feature extraction module is used to perform temporal dynamic modeling of the spatial features of the EEG segment using a second two-dimensional convolution, thereby obtaining the local temporal dependence and dynamic temporal features of the EEG signal in the EEG segment; and to perform downsampling processing on the temporal features through average pooling operation to obtain the temporal features corresponding to the EEG segment.

[0057] It should be noted that the second 2D convolution contains 16 convolutional kernels, each 1×16 in size, and is symmetrically padded (padding=8) in the time dimension. The pooling window size for the average pooling operation is 1×8. The expression for extracting time dimension features using the second 2D convolution is: ; The expression for extracting time features using average pooling is: ; ; In the formula, W temp The temporal convolution kernel for the second two-dimensional convolution. In order to be with the first i The time-dimensional features corresponding to each EEG segment, where Pool() is the average pooling function. and T i All are related to the first i Temporal features corresponding to each EEG segment. These temporal features are temporal features that reduce temporal resolution and enhance stable temporal semantic representation.

[0058] In this embodiment of the application, the feature enhancement module includes: A bottleneck structure is used to compress and reconstruct the temporal features of each EEG segment, resulting in a feature map corresponding to the EEG segment. The feature map of each EEG segment is enhanced using a channel attention mechanism to obtain the feature enhancement map corresponding to the EEG segment. The feature enhancement map of each EEG segment is flattened and mapped to obtain the feature vector corresponding to each EEG segment.

[0059] It should be noted that the bottleneck structure contains three convolutional layers: the first, second, and third convolutional layers. The first convolutional layer uses a 1×1 kernel to compress the temporal features of each EEG segment, obtaining compressed features. The second convolutional layer uses a 1×16 kernel to extract low-dimensional features from the compressed features of each EEG segment, obtaining low-dimensional features. The third convolutional layer uses a 1×1 kernel to restore the channel dimension of the low-dimensional features, obtaining a feature map corresponding to the EEG segment, thus achieving dimensionality reduction modeling and information reconstruction of the temporal features. The expression for the first convolutional layer is: ; The expression for the second convolutional layer is: ; The expression for the third convolutional layer is: ; In the formula, W 1 represents the 1×1 convolution kernel of the first convolutional layer. W 2 represents the 1×16 convolution kernel of the second convolutional layer. W 3 represents the 1×1 convolution kernel of the third convolutional layer. In order to be with the first i Compression features corresponding to each EEG segment In order to be with the first i Low-dimensional features corresponding to individual EEG segments In order to be with the first i Feature maps corresponding to individual EEG segments.

[0060] In this embodiment, the channel attention mechanism includes compressing the feature map using global average pooling to extract global statistical information for each channel; generating channel weights from all global statistical information using a two-layer fully connected network; and weighting and labeling the channel weights of each feature channel in the feature map to obtain a feature enhancement map corresponding to the EEG segment.

[0061] It should be noted that the global statistical response value of the c-th channel is calculated using the first formula of global average pooling. The first formula is: ; In the formula, For the first i The EEG segment in the first c Global statistical response values ​​on each feature channel For the first c Each feature channel is located at (j, k) eigenvalues ​​at that location H and W The feature enhancement module outputs the spatial and temporal dimensions of the feature map, respectively. H×W This represents the total number of elements in the feature map of this feature channel. Channel weights are generated from all global statistics using the channel weight formula of a two-layer fully connected network. The channel weight formula is: ; In the formula, u i For the first i Global statistical response value of individual EEG segments, w i For the first i Channel weights for each EEG segment. A weighted average formula is used to assign channel weights to each feature channel of the feature map. The weighted average formula is: ; In this context, ⊙ represents a channel-wise multiplication operation, which is equivalent to adaptively enhancing key brain region features and suppressing redundant or noisy channels. For the first i Feature enhancement map of a single EEG segment.

[0062] In this embodiment, a flattening formula is used to flatten the feature enhancement map of each EEG segment to obtain one-dimensional data corresponding to each EEG segment; a fixed-dimensional embedding formula is used to map the one-dimensional data of each EEG segment to obtain a feature vector corresponding to the EEG segment.

[0063] It should be noted that the flattening formula is: ; The fixed-dimensional embedding formula is: ; In the formula, F latte ( ) is the flatten function. f i For the first i One-dimensional data of a single EEG segment e i For the first i Each EEG segment is independently encoded as a feature vector. W f For the convolution kernel of the fully connected layer, D=64 is the feature output dimension. b These are the mapping coefficients. The N feature vectors from N EEG segments constitute the feature vector set. E ,Right now: .

[0064] S3. The distribution-aware automatic aggregation mechanism is used to process the N feature vectors to obtain the distribution vectors that characterize the heterogeneity of the patient's brain neural state.

[0065] It should be noted that processing N feature vectors typically employs fixed statistical rules (such as taking only the mean, or fixed concatenation of the mean, standard deviation, 75th percentile P75, and 90th percentile P90) for feature aggregation. This method suffers from strong prior dependence and insufficient flexibility: for some patients, the mean may be the most discriminative; while for others (such as patients with minimally conscious MCS), higher-tailed quantiles (such as P90) may be more critical. Fixed statistical rules cannot adaptively adjust to the individualized EEG distribution characteristics of different patients. Step S3 of this patient consciousness disorder identification method based on distribution-aware automatic aggregation mechanism involves processing the patient's feature vector set output in step S2. E This study employs a distribution-aware auto-pooling (DAAP) mechanism for patient-level statistical aggregation to construct a distribution vector characterizing the heterogeneity of patients' brain neural states, providing data for identifying patient consciousness disorders based on this mechanism. In this embodiment, rules or formulas are set on the DAAP mechanism to process N feature vectors. The DAAP mechanism includes: calculating the mean, standard deviation, p1th percentile, p2th percentile, ..., pqth percentile, ..., p1th percentile for each channel dimension of the N feature vectors, where p1 > pq > p2 > p1. The mean, standard deviation, and p1 percentiles are then weighted and fused to form an individual EEG distribution manifold vector as a single distribution vector. This distribution vector is represented as... If the output dimension feature D=64, a distribution vector of patient-level distribution features with dimension 256 is obtained. The distribution-aware automatic aggregation mechanism aims to automatically discover and weightedly fuse statistical descriptive quantities that best characterize the heterogeneity of the patient's internal neural states.

[0066] S4. An interpretable classifier is used to identify the distribution vector of the patient to be identified as having a disorder of consciousness, and the classification result of the patient to be identified as having a disorder of consciousness is obtained. Among them, the deep learning encoder is cascaded with the interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold of classification and identification in the interpretable classifier, so as to achieve the identification of disorder of consciousness with high sensitivity.

[0067] It should be noted that step S4 involves inputting the distribution vector of the patient to be identified with altered consciousness obtained in step S3 into the interpretable classifier. This outputs the predicted result and corresponding probability value of whether the patient belongs to a vegetative state (UWS) or a minimally conscious state (MCS), thus achieving the identification of patient altered consciousness based on a distribution-aware automatic aggregation mechanism. In this embodiment, the interpretable classifier is either a gradient boosting classifier based on decision tree ensemble or an interpretable classifier based on a linear model. Specifically, the interpretable classifier can be the LightGBM interpretable classifier. The distribution features extracted from the distribution vector by the deep encoder are cascaded with the LightGBM interpretable classifier, rather than directly connected to a fully connected layer. LightGBM, based on decision tree ensemble, allows the classification decision for each sample to be traced back to the specific splitting path and feature contribution. LightGBM naturally supports feature importance ranking, providing a foundation for subsequent clinical insights into which EEG patterns are most sensitive.

[0068] Furthermore, interpretable classifiers can employ the SHAP (SHapley Additive exPlanations) additive interpretation model to quantify the contribution of each statistical feature (such as the standard deviation of a channel, percentile, etc.) in the output distribution vector to the classification result. For example, an interpretable classifier can output: "Channel 2 has the highest contribution from the standard deviation, suggesting increased variability in EEG activity in this region," thus providing clinicians with an understandable basis for decision-making.

[0069] In other embodiments, the interpretable classifier may also employ interpretable classifiers such as Support Vector Machine (SVM) or Random Forest.

[0070] In this embodiment of the application, the patient consciousness disorder identification method based on the distribution perception automatic aggregation mechanism performs preprocessing, deep feature extraction, patient-level distribution feature construction and classification identification of the multi-lead EEG signals of the patient to be identified for consciousness disorder in steps S1 to S4, thereby realizing the automatic identification of the patient's consciousness state.

[0071] This application provides a method for identifying patient consciousness disorders based on a distribution-aware automatic aggregation mechanism. The method includes acquiring multi-lead EEG signals from a patient to be identified as having a consciousness disorder; preprocessing and segmenting the multi-lead EEG signals to obtain N fixed-length EEG segments; using a deep learning encoder to extract deep feature representations from each EEG segment to obtain N feature vectors corresponding to the N EEG segments; processing the N feature vectors using a distribution-aware automatic aggregation mechanism to obtain a distribution vector representing the heterogeneity of the patient's brain neural state; and using an interpretable classifier to identify the consciousness disorder based on the distribution vectors of the patient to be identified, obtaining the classification result for the corresponding patient. The deep learning encoder and the interpretable classifier are cascaded, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification in the interpretable classifier. This method, based on EEG signals, processes multi-lead EEG signals from patients to obtain distribution vectors characterizing the distribution features of the patient's EEG signals. Then, an interpretable classifier is used to identify the distribution vectors of patients with consciousness disorders to obtain the classification results. This method solves the technical problems of insufficient patient-level feature expression, lack of interpretability of classification decisions, and limited generalization performance under high sensitivity requirements in existing EEG identification methods for consciousness disorders.

[0072] In one embodiment of this application, each feature vector includes D dimensional features. A distribution-aware automatic aggregation mechanism is used to process the N feature vectors to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state, including: The mean and standard deviation corresponding to each dimension feature are calculated based on the number N of feature vectors and each dimension feature of all feature vectors. The percentile function is used to calculate the percentile corresponding to each dimension feature of all feature vectors based on multiple preset thresholds of different values. Based on the mean, standard deviation and all percentiles of each dimension feature, a basic statistical feature pool corresponding to each dimension feature is determined. The basic statistical feature pool includes K basic feature data. Based on each basic feature data in all basic statistical feature pools, determine the coefficient of variation corresponding to the basic feature data; The basic statistical feature pools for each dimension feature are weighted and fused based on all the coefficients of variation to obtain the distribution manifold vector corresponding to each dimension feature; All distribution manifold vectors are concatenated according to the order of dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

[0073] It should be noted that in obtaining the distribution vector, a multidimensional moment-quantile joint descriptor is first constructed: for each feature dimension of each feature vector, the mean and standard deviation of N feature vectors are calculated to represent the central tendency and dispersion of EEG activity, respectively; then, percentile density modeling is performed: for each feature dimension, the corresponding percentile projection is calculated using a preset threshold set of quantiles to capture the long-tailed distribution characteristics and extreme state fluctuations of EEG features; individual distribution vector fusion: the mean, standard deviation, and all percentile projections are concatenated in dimensional order to form an individual EEG distribution manifold vector, which serves as the distribution vector; wherein, the individual EEG distribution manifold vector, through low-dimensional parameterization, fully preserves the statistical complete picture of the patient's internal EEG segments, achieving adaptive aggregation from the segment level to the individual level. In this embodiment, the preset threshold set includes asymmetric percentiles on both sides of the median and high-tailed partial sites, and the preset threshold set includes p1, p2, ... pq, ..., p1, etc., such as 10%, 25%, 50%, 75%, 90%, and 95%, etc. Asymmetric percentiles on either side of the median (e.g., 25%, 50%, 75%) are used to characterize the skewness of EEG distribution to distinguish the differences in neural activity patterns under different states of consciousness; high-tailed percentiles (90%, 95%) are used to characterize abnormal fluctuations to capture intermittent, weak signals of consciousness.

[0074] In the embodiments of this application, in the process of calculating the mean and standard deviation of each dimension feature, the mean corresponding to the dimension feature can be obtained by calculating based on the number of feature vectors N and each dimension feature of the feature vector; the standard deviation corresponding to the dimension feature can be obtained by calculating based on the number of feature vectors N, each dimension feature of the feature vector and the corresponding mean.

[0075] It should be noted that for the set of feature vectors E Statistical calculations are performed on each channel dimension separately, and multiple statistics are independently extracted for each dimension feature. The mean of each dimension feature describes the overall distribution position of the dimension feature. The standard deviation of each dimension feature describes the dispersion of the dimension feature distribution. The mean of each dimension feature is calculated using the mean formula based on the number of feature vectors N and the mean of each feature vector. The standard deviation of each dimension feature is calculated using the standard deviation formula based on the number of feature vectors N, the mean of each feature vector, and the corresponding mean. The mean formula is: ; The formula for standard deviation is: ; In the formula, d For the first d Features in each dimension e i，d In order to be with the first iThe first EEG segment corresponds to the feature vector of the first... d Features in each dimension m d For the first d The mean of each feature dimension, σ d For the first d The standard deviation of each dimension feature.

[0076] In this embodiment of the application, let the set of dimensional features of the distribution vector corresponding to the patient be . After sorting the samples in ascending order, we get Then the qth percentile is defined as This indicates that there are samples. q% The observed value is not greater than this value, describing the statistical characteristics of the high interval in the nth dimension feature. Among them, Percentile () represents the Percentile function.

[0077] It should be noted that if the preset threshold can be selected as 10%, 25%, 50%, 75%, 90%, or 95%, then the corresponding [number] [value]... d The percentiles of each dimension feature are denoted as P 10 d , P 25 d , P 50 d , P 75 d , P 90 d and P 95 d ,like , Etc. Distribution vector CDF controls the feature dimensionality (avoiding dimensionality explosion caused by directly using all epochembedding) while preserving the statistical full picture of intra-patient variation.

[0078] In this embodiment of the application, if the first d The basic statistical feature pool of each dimension feature is denoted as... S d Basic statistical feature pool S d It includes central tendency indicators, dispersion indicators, and distribution pattern indicators. Among them, the basic statistical feature pool comprehensively describes the distribution pattern of this feature dimension across all N EEG segments.

[0079] It should be noted that the central trend indicator can use the mean. m d The degree of dispersion can be expressed using the standard deviation σ. d Distribution patterns can be represented using multiple percentiles, such as... P10 d , P 25 d , P 50 d , P 75 d , P 90 d and P 95 d In other embodiments, the distribution pattern indicators can also be selected as skewness and kurtosis of the data. In this embodiment, in the basic statistical feature pool... S d In this context, the distribution pattern index can be selected from relatively representative percentiles, such as the basic statistical feature pool. S d =[ m d , σ d , P 25 d , P 75 d , P 90 d These five basic feature data constitute the pool, k∈K=5. It should be understood that the specific composition of the basic statistical feature pool is not limited to this; any statistic that can characterize the distribution characteristics of EEG features can be included in the pool, and will be automatically filtered and weighted by the subsequent adaptive importance weighting mechanism.

[0080] In this embodiment of the application, in order to automatically identify which statistics are most discriminative for the current sample data, the patient consciousness disorder identification method based on the distribution-aware automatic aggregation mechanism, during the process of obtaining the distribution vector, performs the following for each basic feature data. s k ∈ S d The coefficient of variation (CV) of the basic feature data across N EEG segments is calculated and used as the importance prior, Uk. Then, the CV corresponding to each basic feature data in all basic statistical feature pools is calculated using the variability formula. The variability formula is: U k = CV ( s k =Std( s k ) / (|Mean( s k )|+ e ); In the formula, s k For the first kBasic feature data; e This is the smoothing factor, typically taking a value of 1×e. -8 or 1×e -6 ;Std( s k ) is the first of the D-dimensional features k The standard deviation of the basic feature data, Mean( s k ) is the first of the D-dimensional features k The mean of the basic feature data, U k and CV ( s k All of them are the first k The coefficient of variation of the basic feature data.

[0081] It should be noted that the smoothing factor serves to prevent division by zero errors when the absolute value of the mean in the denominator is 0 (division by zero protection). To capture extreme responses (such as large fluctuations), EEG responses in patients with minimally conscious MCS are often intermittent. In this case, high-tailed percentiles (such as...) P 90 d The coefficient of variation may vary greatly across all segments. U k (Very high), this abnormal fluctuation is precisely the key feature that distinguishes Minimal Consciousness (MCS) from Unresponsive Awakening Syndrome (UWS). At this time, the high coefficient of variation... U k Corresponding to high weighting. To capture a stable baseline (e.g., small fluctuations are important), for some patients, their state of consciousness may be determined by the level of overall EEG activity (e.g., patients with unresponsive wakefulness syndrome (UWS) have extremely low and stable overall activity). In this case, the feature varies very little across all segments (coefficient of variation). U k (Very low), this high stability itself is a strong signal. At this point, the low coefficient of variation... U k This might correspond to high weights (or, through MLP learning, map low coefficients of variation to high importance). Coefficient of variation U k The larger the value, the greater the fluctuation of the basic characteristic data in different samples (e.g., the percentile of patients with minimally conscious MCS). P 90 d (Vast fluctuations) This type of uncertainty or instability is itself a characteristic of high discriminative power and should be given high weight. Coefficient of variation U kThe smaller the value, the more stable the basic feature data (for example, the mean of unresponsive wakefulness syndrome (UWS) patients is very low and stable). This kind of stability is also a strong signal and is retained. In this embodiment, the coefficient of variation of each basic feature data is calculated to transform the inherent statistical characteristics of the data into features. Then, through subsequent adaptive importance weighting, the neural network automatically learns from the training data to determine whether to reward "features with large fluctuations" or "features with small fluctuations".

[0082] In this embodiment, during the process of obtaining the distribution vector, the adaptive weight coefficient of each basic feature data is first calculated based on the coefficient of variation of each basic feature data; then, the adaptive weight coefficient of each basic feature data is weighted with the basic feature data in the basic statistical feature pool of each dimension feature to obtain the distribution manifold vector of the corresponding dimension feature; finally, all dimension features are... d =1, ..., D distribution manifold vectors V d After splicing, the final K×D-dimensional individual EEG distribution manifold vector is formed as the distribution vector.

[0083] It should be noted that, in the process of calculating the adaptive weight coefficients for each basic feature data, the coefficient of variation U=[ of the basic statistical feature pool] can be used. U 1, ..., U k ,..., U K The input is fed into a lightweight fully connected network (MLP) or attention layer, and normalized using the Softmax function to obtain a weight set W = [ ]. w 1, ..., w k ,..., w K ], that is, W=[ w 1, ..., w k ,..., w K =Softmax(MLP(U)). In this embodiment, if a patient belongs to the Minimal Consciousness State (MCS) in the test set, and its characteristic manifestations are the basic feature data. P 90 d Drastic fluctuations will automatically improve the correlation with basic feature data. P 90 d The corresponding adaptive weighting coefficients amplify this anomalous signal. If a patient in the test set belongs to unresponsive wakefulness syndrome (UWS) and their characteristic features are characterized by extremely low and stable mean values, the baseline feature data will be automatically increased. md Corresponding adaptive weight coefficients are used to reinforce this stable signal. The distribution-aware automatic aggregation mechanism achieves true adaptability, independent of manually defined rules (such as only looking at basic feature data). P 90 d Or just look at the basic feature data m d Instead of relying on statistical methods, the study dynamically determines which statistical descriptive measures (whether quantiles reflecting extreme fluctuations or mean reflecting overall levels) are most critical for assessing a patient's state of consciousness, based on the statistical distribution characteristics of the patient's current EEG data.

[0084] In this embodiment, the process of obtaining the distribution manifold vector of the dimensional feature involves first multiplying the adaptive weight coefficient of each basic feature data with the basic feature data corresponding to each dimensional feature to obtain the statistical feature; then, all statistical features of each dimensional feature are summed to obtain the distribution manifold vector of the corresponding dimensional feature. V d =Σ( w k × s k ), k From 1 to K .

[0085] It should be noted that the distributed sensing automatic aggregation mechanism has the functions of de-priorification, individual adaptability, and interpretability preservation. Specifically, de-priorification means that computation is not required without human experience specifying it. P 75 d or P 90 d Instead of relying on data distribution, it automatically selects the most effective statistical description. Individual adaptability refers to the ability to dynamically adjust the emphasis of feature expression based on the degree of EEG heterogeneity among different patients, achieving true adaptive aggregation. Interpretability preservation means that although the weights are automatically generated, the output is still a clear combination of statistics. Through the SHAP value of subsequent interpretable classifiers (such as LightGBM), it is still possible to trace which statistics (and their automatic weights) drove the decision.

[0086] To ensure high sensitivity (low false negative rate) in clinical applications, a constraint threshold optimization mechanism is introduced. Addressing the high risk of missed diagnoses in minimally conscious states (MCS) during the identification of disorders of consciousness, this constraint threshold optimization mechanism does not solely aim for overall accuracy but prioritizes clinical safety. In one embodiment of this application, the constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; A deep learning encoder and a distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, an interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision thresholds is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to each candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold. The classification accuracy corresponding to each candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and sensitivity, all candidate decision thresholds are screened to obtain the optimal decision threshold that satisfies the sensitivity constraint threshold. Based on the classification accuracy, the argmax function is used to filter all the better decision thresholds to obtain the decision thresholds for classification recognition in the interpretable classifier.

[0087] It should be noted that the sensitivity constraint threshold can be chosen as 0.9. In determining the decision threshold for classification and recognition in the interpretable classifier, several arbitrary candidate decision thresholds can be set. i First, the MCS prediction probability of each patient's sample data is compared with the current candidate decision threshold. When the prediction probability is not less than... i If the condition is met, it is classified as MCS; otherwise, it is classified as UWS. The consistency of the classification results obtained under this candidate threshold is compared with the true labels of the patient's sample data to obtain the classification evaluation data for all sample data under the corresponding candidate decision threshold. In this embodiment, the sensitivity corresponding to the candidate decision threshold is calculated using a sensitivity formula based on the first correctly predicted sample number and the first incorrectly predicted sample number for each candidate decision threshold; the classification accuracy corresponding to the candidate decision threshold is calculated using a classification accuracy formula based on the first correctly predicted sample number, the first incorrectly predicted sample number, the second correctly predicted sample number, and the second incorrectly predicted sample number for each candidate decision threshold. The sensitivity formula is: ; The formula for classification accuracy is: ; In the formula, TP(θ () represents the candidate decision threshold i The number of the first correct prediction samples, FN(θ() represents the candidate decision threshold i The number of first mispredicted samples, TN(θ () represents the candidate decision threshold i The number of the second correct prediction samples, FP(θ () represents the candidate decision threshold i The second number of incorrectly predicted samples, Recall( i) Candidate decision threshold i Sensitivity, Accuracy i) Candidate decision threshold i The classification accuracy.

[0088] In this embodiment of the application, in order to reduce the risk of missed detection of MCS patients, while meeting the sensitivity constraint threshold... t Given the given conditions, the optimal threshold is selected from all candidate decision thresholds to obtain the decision threshold, i.e.: ; In the formula, i To explain the decision threshold for classification and recognition in the classifier, t The threshold is used as a sensitivity constraint. This patient consciousness disorder identification method based on a distributed sensing automatic aggregation mechanism ensures high recall and reduces the risk of missed diagnoses for MCS patients through a constraint threshold optimization mechanism, while also taking into account the overall balance of classification performance. The output classification results of patients to be identified with consciousness disorders are more in line with clinical application needs.

[0089] In one embodiment of this application, during the process of determining the classification decision threshold in the interpretable classifier, the specificity corresponding to the candidate decision threshold can be calculated using a specificity formula based on the number of correctly predicted samples and the number of incorrectly predicted samples for each candidate decision threshold. The specificity is then used to comprehensively evaluate the classification performance of the interpretable classifier under different candidate decision threshold conditions. The specificity formula is: ; In the formula, Specificity( i) Candidate decision threshold i Specificity.

[0090] In one embodiment of this application, before using an interpretable classifier to identify disorders of consciousness based on each candidate decision threshold and the patient distribution vectors of M patients, the method for identifying the state of patients with disorders of consciousness further includes: data augmentation of the sample data for each patient; the content of data augmentation of the sample data for each patient includes: Based on the patient distribution vector of each patient, an adaptive synthetic sampling method is used to amplify the corresponding sample data to obtain the first training sample corresponding to the patient; Obtain random samples corresponding to the consciousness disorder category of each patient's sample data, and process them using linear interpolation based on each patient's sample data and random samples to obtain the second training sample corresponding to the patient; The first and second training samples of each patient are merged to obtain the corresponding patient's training sample as new sample data.

[0091] It should be noted that, since the number of samples of different consciousness states in patient EEG signal data is often imbalanced, in order to reduce the impact of class distribution differences on the classification and recognition training of the interpretable classifier, the training set sample data is first subjected to class balancing processing during the training phase of the interpretable classifier. In this embodiment, the first training sample can be represented as... In the formula, CFDs m Indicates the first m The distribution vector at the patient level corresponding to each patient. y m This indicates the corresponding state of consciousness label. M This represents the number of training samples (e.g., the number of patients). The adaptive synthetic sampling method can be an adaptive synthetic sampling approach (such as Adaptive Synthetic Sampling, ADASYN). The ADASYN method uses minority class samples as a base and adaptively determines the number of amplified samples based on the class distribution in the sample's neighborhood, generating new synthetic samples in the feature space. More synthetic samples are allocated to minority class samples with a higher proportion of majority class samples in the neighborhood; fewer synthetic samples are allocated to minority class samples with a lower proportion of majority class samples. In this way, samples that are difficult to distinguish near the classification boundary receive more attention, thereby improving the classification model's ability to learn minority class samples.

[0092] In this embodiment of the application, during the process of obtaining the second training sample of the patient, from the first m Sample data from one patient X m A random sample is randomly selected from the nearest neighbors of a minority class. x z And new synthetic samples are generated using linear interpolation. x new As the second training sample, i.e. In the formula, λ represents a random number that follows a uniform distribution between 0 and 1.

[0093] It should be noted that the patient training samples are used as new sample data. . D ADASYN This represents a synthetic sample generated from all patients. xnew The collection is constructed by using sample data from patients tagged with MCS (Minimally Conscious State) as a minority category for sample amplification to avoid data leakage and assessment bias.

[0094] In one embodiment of this application, during the process of obtaining the second training sample of the patient, for the first... m A small number of patient categories X m According to its k Number of samples belonging to the majority class among the nearest neighbors m Calculate the difficulty coefficient for the corresponding patient. The formula for calculating the difficulty coefficient is: ; In the formula, k Represents the number of nearest neighbor samples. r m Reflecting the sample X m The classification coefficient. When r m A higher value indicates a greater number of majority class samples surrounding the sample, meaning it is closer to the class boundary and requires more synthetic samples to be generated. The difficulty coefficient of all minority class samples is normalized using the following formula: ; In the formula, N m This represents the total number of samples in the minority category. This represents the difficulty coefficient after normalization. Let G be the total number of minority class samples that need to be generated, then the sample... x i The corresponding number of synthetic samples generated is: ; In the formula, g m Indicates the assignment to the sample X m The number of synthetic samples. This shows that samples with greater classification difficulty will generate more synthetic samples, thus allowing interpretable classifiers to focus more on complex samples near class boundaries.

[0095] In other embodiments, in addition to using ADASYN for data augmentation, SMOTE, random oversampling, or cost-sensitive learning methods may be used to mitigate class imbalance in patient sample data.

[0096] In one embodiment of this application, an interpretable classifier is used to identify the distribution vector of the patient to be identified as having a disorder of consciousness, and the classification result for the corresponding patient to be identified as having a disorder of consciousness includes: The distribution vector of the patient to be identified with impaired consciousness is input into the interpretable classifier, and the interpretable classifier outputs the original decision score of the patient to be identified with impaired consciousness. Based on the original decision scores, the predicted probability of the microconsciousness state of the patient to be identified as having a disturbance of consciousness is obtained. The classification results of the patient to be identified with the disorder of consciousness are obtained by comparing the predicted probability of the minimal state of consciousness with the decision threshold. The classification results include either plant state or minimal consciousness state.

[0097] It should be noted that the classification and recognition of the interpretable classifier belongs to the ensemble learning model based on gradient boosting decision trees. In this embodiment, if the interpretable classifier is the LightGBM interpretable classifier, the classification and recognition of the LightGBM interpretable classifier are obtained through stepwise iterative training of multiple decision trees. Each decision tree divides the input feature space (such as the mathematical space where the distribution features of the CFD distribution vector are located, which is actually a 256-dimensional space composed of all possible values ​​of the distribution features of the patient distribution vector) into segments, and outputs the final classification result through a weighted combination method.

[0098] In the embodiments of this application, the classification and recognition of the LightGBM interpretable classifier specifically includes: tree-based decision path expression capability, feature contribution analysis capability, training results of classification and recognition based on the LightGBM interpretable classifier, efficient training and inference capability, and adaptability to high-dimensional features.

[0099] In this embodiment of the application, the decision path representation capability based on tree structure refers to the fact that for each input distribution vector, its classification result is jointly determined by the leaf node paths of multiple decision trees, and the decision process can be traced and analyzed based on the node splitting characteristics.

[0100] In this embodiment, the feature contribution analysis capability based on the training results of the LightGBM interpretable classifier refers to the quantitative analysis of the contribution of statistical features of each dimension in the patient-level distribution feature representation vector CFD to the classification output, thereby characterizing the influence mechanism of different distribution statistics on the classification results of unresponsive wakefulness syndrome (UWS, also known as vegetative state) and minimally conscious state (MCS, also known as minimally conscious state). m d , σ d , P 25 d , P 75 d ,P 90 d ] is used as a distribution vector for explanation, for example: for the th m The distribution vector of the patient-level distribution characteristics of a sample of patients is represented as follows: ; In the formula, the first l The contribution of the distribution characteristics of the dimensional CFD distribution vector to the classification and recognition output of the LightGBM interpretable classifier is defined as follows: And expressed as a Shapley additive interpretation model: ; In the formula, This represents the classification output for the m-th patient. This represents the baseline output for classification and recognition. Indicates the first l The marginal contribution of CFD statistical characteristics to classification results; further, for M patients, the... l The global influence weight of the distribution characteristics of a dimensional distribution vector CFD is defined as its average absolute contribution value, and the calculation of the dimensional distribution vector CFD is performed on the dimensional distribution vector CFD. l The formula for the average absolute contribution of the distribution characteristics of a dimensional vector CFD is: ; The relative influence weights of the distribution characteristics of each distribution vector in CFD are obtained by normalization, and the distribution vector of the first distribution vector is calculated. l The formula for the relative influence weights of the dimensional distribution features is: ; In the formula, The relative importance of distribution features used to characterize the distribution vector CFD for the UWS / MCS classification task.

[0101] In this embodiment, efficient training and inference capability refers to the use of a histogram-based split point search strategy and a leaf-first growth strategy to improve the classification and recognition training efficiency of the LightGBM interpretable classifier and reduce computational complexity.

[0102] In this embodiment, the adaptability to high-dimensional features refers to the ability to adapt to the high-dimensional patient-level distributed feature vector (CFD) input generated in step S3 and map it to the classification space to output the classification result of the patient's state of consciousness. For the first... m For each patient, the classification decision function is expressed as: ; In the formula, f t () indicates the first classification result in the LightGBM interpretable classifier. tThe output function of each decision tree s m The raw decision scores for classification recognition by the LightGBM interpretable classifier.

[0103] Furthermore, based on the original decision scores s m By mapping the original decision scores to probability outputs using the Sigmoid function, the predicted probability of the patient belonging to the minimum state of consciousness (MCS, also known as minimally conscious state) is obtained. P ( MCS|CFD m The Sigmoid function is: ; Correspondingly, the predicted probability that the m-th patient belongs to unresponsive wakefulness syndrome (UWS), also known as a vegetative state, is: ; In the formula, P() represents the posterior probability output obtained by the LightGBM interpretable classifier based on the CFD distribution vector input for classification recognition.

[0104] In the embodiments of this application, the predicted probability of the micro-consciousness state of the patient to be identified as having a disorder of consciousness is judged based on a decision threshold, thereby obtaining the final classification result. For the first... m The predicted probability of minimally conscious states in patients with pending consciousness disorders. P m Represented as: ; Predicting the probability of subtle states of consciousness using a thresholding function P m The mapping is to the classification result, and the thresholding function is: ; In the formula, i The decision threshold for classification and recognition. This indicates that the predicted classification result is MCS. This indicates that the predicted classification result is UWS. i definition: .

[0105] In one embodiment of this application, the patient consciousness disorder identification method based on the distribution-aware automatic aggregation mechanism further includes: outputting the classification results, the output of which includes the consciousness disorder category of the patient to be identified, the predicted probability of the consciousness disorder category, and the contribution value of each distribution feature in the distribution vector.

[0106] It should be noted that the contribution values ​​of each distribution feature in the distribution vector can be understood as follows: a higher mean in the first channel indicates enhanced α-band activity related to MCS. These features are then sorted according to their contribution, and the top K feature dimensions with the highest contribution are selected as the key discriminant feature set. The key discriminant features include the mean, standard deviation, and quantile feature dimensions from the distribution features of the CFD distribution vector.

[0107] In the embodiments of this application, sample data from 109 patients (71 UWS and 38 MCS) are used as an example. The accuracy rate of the automatic aggregation mechanism for identifying patient consciousness disorders based on the distribution perception mechanism is 83.4%, while the accuracy rate of the control group (using only the mean aggregation) without DAP is approximately 65%, representing a relative improvement of approximately 28% (an absolute improvement of approximately 18 percentage points). The LightGBM interpretable classifier naturally supports feature importance ranking and can output the contribution of each distribution vector CFD channel to the MCS / UWS decision. By combining the signal channel mapping relationship of the acquired EEG signals, it is possible to trace back to specific EEG leads and specific statistics (e.g., a high standard deviation of the first signal channel in the central region → tending towards MCS), providing clinicians with an understandable basis for decision-making. The interpretable classifier not only outputs classification results but also outputs decision-making rationale, enabling doctors to verify rationality, discover abnormal patterns, accumulate clinical experience, and thus build trust in the classification results.

[0108] Example 2: Figure 3 This is a schematic diagram of the framework of the patient consciousness disorder recognition system based on the distributed sensing automatic aggregation mechanism described in the embodiments of this application.

[0109] like Figure 3 As shown, this application provides a patient consciousness disorder recognition system based on a distribution-aware automatic aggregation mechanism, including: a data acquisition module 100, a feature extraction module 200, a distribution vector module 300, and a disorder recognition module 400; The data acquisition module 100 is used to acquire multi-lead EEG signals from patients to be identified as having impaired consciousness, and to preprocess and segment the multi-lead EEG signals to obtain N fixed-length EEG segments. The feature extraction module 200 is used to extract deep feature representations for each EEG segment using a deep learning encoder, and obtain N feature vectors corresponding to N EEG segments. The distribution vector module 300 is used to process N feature vectors using a distribution-aware automatic aggregation mechanism to obtain a distribution vector that characterizes the heterogeneity of the patient's brain neural state. The obstacle recognition module 400 is used to identify the consciousness disorder by using an interpretable classifier to identify the distribution vector of the patient to be identified as having a consciousness disorder, and to obtain the classification result of the corresponding patient to be identified as having a consciousness disorder. In this system, a deep learning encoder is cascaded with an interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification and recognition in the interpretable classifier.

[0110] It should be noted that the content of each module in the device of Embodiment 2 corresponds to the content of each step in the method of Embodiment 1. Therefore, the content of each module of the patient consciousness disorder recognition system based on the distributed perception automatic aggregation mechanism will not be repeated in this embodiment. This patient consciousness disorder recognition system based on the distributed perception automatic aggregation mechanism fully preserves the patient's internal EEG variation information through the distributed perception automatic aggregation mechanism of the data acquisition module, feature extraction module, distribution vector module, and disorder recognition module. An interpretable classifier provides traceable decision-making, and a constraint threshold optimization mechanism optimizes to meet clinical sensitivity requirements, achieving high-precision, interpretable, and clinically applicable recognition of consciousness disorders. The data acquisition module 100, feature extraction module 200, distribution vector module 300, and disorder recognition module 400 can be deployed on the same computing device (such as a hospital workstation or portable EEG analysis terminal), or they can be deployed in a distributed manner (e.g., data acquisition and encoding of the data acquisition module 100 and feature extraction module 200 are deployed on the same device, while the aggregation and classification of the distribution vector module 300 and disorder recognition module 400 are deployed on a cloud server).

[0111] In this embodiment, each feature vector includes D dimensional features, and the distribution vector module 300 includes a calculation submodule, a feature pool determination submodule, a coefficient determination submodule, a weighted fusion submodule, and a concatenation submodule. The calculation submodule is used to calculate the mean and standard deviation corresponding to each dimension feature based on the number N of feature vectors and each dimension feature of all feature vectors; and to calculate the percentile corresponding to each dimension feature of all feature vectors using the Percentile function based on multiple preset threshold values ​​to obtain the percentile corresponding to each dimension feature at each preset threshold. The feature pool determination submodule is used to determine the basic statistical feature pool corresponding to each dimension feature based on the mean, standard deviation and all percentiles of each dimension feature. The basic statistical feature pool includes K basic feature data. The coefficient determination submodule is used to determine the coefficient of variation corresponding to each basic feature data in all basic statistical feature pools. The weighted fusion submodule is used to perform weighted fusion processing on the basic statistical feature pool of each dimension feature based on all the coefficients of variation, so as to obtain the distribution manifold vector corresponding to each dimension feature; The splicing submodule is used to splice all the distribution manifold vectors in the order of dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

[0112] In one embodiment of this application, the constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; A deep learning encoder and a distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, an interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision thresholds is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to each candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold. The classification accuracy corresponding to each candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and sensitivity, all candidate decision thresholds are screened to obtain the optimal decision threshold that satisfies the sensitivity constraint threshold. Based on the classification accuracy, the argmax function is used to filter all the better decision thresholds to obtain the decision thresholds for classification recognition in the interpretable classifier.

[0113] Example 3: Figure 4 This is a schematic diagram of the terminal device described in an embodiment of this application.

[0114] like Figure 4 As shown, this application provides a terminal device, including a processor and a memory; Memory is used to store program code and transfer the program code to the processor; The processor is used to execute the aforementioned patient consciousness impairment identification method based on the distribution-aware automatic aggregation mechanism according to the instructions in the program code.

[0115] It should be noted that the processor is used to execute the steps in the above-described embodiments of the patient consciousness impairment recognition method based on the distributed perception automatic aggregation mechanism, according to the instructions in the program code. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described system / device embodiments.

[0116] For example, a computer program can be divided into one or more modules / units, one or more of which are stored in memory and executed by a processor to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a terminal device.

[0117] Terminal devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. Terminal devices may include, but are not limited to, processors and memory. Those skilled in the art will understand that this does not constitute a limitation on the terminal device, which may include more or fewer components than illustrated, or combinations of certain components, or different components. For example, a terminal device may also include input / output devices, network access devices, buses, etc.

[0118] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0119] Memory can be an internal storage unit of a terminal device, such as a hard drive or RAM. Memory can also be an external storage device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal and external storage units. Memory is used to store computer programs and other programs and data required by the terminal device. Memory can also be used to temporarily store data that has been output or will be output.

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0122] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying patient consciousness impairment based on a distributed sensing automatic aggregation mechanism, characterized in that, Includes the following steps: Acquire multi-lead EEG signals from patients to be identified as having impaired consciousness, and perform preprocessing and segmentation on the multi-lead EEG signals to obtain N EEG segments; A deep learning encoder is used to extract deep feature representations for each of the N EEG segments, resulting in N feature vectors corresponding to the N EEG segments. The N feature vectors are processed using a distribution-aware automatic aggregation mechanism to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state; An interpretable classifier is used to identify the consciousness disorder by the distribution vector of the patient to be identified, and the classification result of the corresponding patient to be identified is obtained. The deep learning encoder is cascaded with the interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification and recognition in the interpretable classifier.

2. The method for identifying patient consciousness impairment based on a distributed sensing automatic aggregation mechanism according to claim 1, characterized in that, Each feature vector includes D-dimensional features. A distribution-aware automatic aggregation mechanism is used to process the N feature vectors to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state, including: The mean and standard deviation corresponding to each dimension feature are calculated based on the number N of the feature vectors and each dimension feature of all the feature vectors; the percentile corresponding to each dimension feature of all the feature vectors is calculated using the Percentile function based on multiple preset threshold values ​​to obtain the percentile corresponding to each dimension feature at each preset threshold. Based on the mean, standard deviation and all percentages of each dimension feature, a basic statistical feature pool is determined corresponding to each dimension feature, the basic statistical feature pool including K basic feature data; Based on each basic feature data in all the basic statistical feature pools, determine the coefficient of variation corresponding to the basic feature data; The basic statistical feature pools of each dimensional feature are weighted and fused according to all the coefficients of variation to obtain the distribution manifold vector corresponding to each dimensional feature; All the aforementioned distribution manifold vectors are concatenated according to the order of their dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

3. The method for identifying patient consciousness impairment based on a distributed sensing automatic aggregation mechanism according to claim 1, characterized in that, The constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; The deep learning encoder and the distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, the interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision threshold is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold; the classification accuracy corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and the sensitivity, all the candidate decision thresholds are filtered to obtain a better decision threshold that satisfies the sensitivity constraint threshold; Based on the classification accuracy, all the better decision thresholds are filtered using the argmax function to obtain the decision thresholds for classification and recognition in the interpretable classifier.

4. The method for identifying patient consciousness impairment based on a distributed sensing automatic aggregation mechanism according to claim 3, characterized in that, Before identifying disorders of consciousness using the interpretable classifier based on each candidate decision threshold and the patient distribution vectors of the M patients, the method for identifying the state of a patient with a disorder of consciousness further includes: data augmentation of the sample data for each patient; the content of data augmentation of the sample data for each patient includes: Based on the patient distribution vector of each patient, an adaptive synthetic sampling method is used to amplify the corresponding sample data to obtain the first training sample corresponding to the patient. Obtain random samples corresponding to the consciousness disorder category of each patient's sample data, and process the sample data and random samples of each patient using linear interpolation to obtain a second training sample corresponding to each patient; The first training sample and the second training sample of each patient are merged to obtain the corresponding patient's training sample as new sample data.

5. The method for identifying patient consciousness impairment based on a distributed sensing automatic aggregation mechanism according to any one of claims 1-4, characterized in that, The distribution vector of the patient to be identified as having a disorder of consciousness is used to identify the disorder of consciousness, and the classification results for the corresponding patient to be identified as having a disorder of consciousness are as follows: The distribution vector of the patient to be identified as having a disorder of consciousness is input into the interpretable classifier, and the interpretable classifier outputs the original decision score of the patient to be identified as having a disorder of consciousness. Based on the original decision score, the predicted probability of the microconscious state of the patient to be identified as having a disturbance of consciousness is obtained. The classification result of the patient to be identified as having a disturbance of consciousness is obtained by comparing the predicted probability of the micro-consciousness state with the decision threshold. The classification results include plant state or minimal consciousness state.

6. The method for identifying patient consciousness impairment based on a distribution-aware automatic aggregation mechanism according to any one of claims 1-4, characterized in that, It also includes: outputting the classification results, the output of which includes the consciousness disorder category of the patient to be identified, the predicted probability of the consciousness disorder category, and the contribution value of each distribution feature in the distribution vector.

7. A patient consciousness impairment identification system based on a distributed sensing automatic aggregation mechanism, characterized in that, include: Data acquisition module, feature extraction module, distribution vector module, and obstacle recognition module; The data acquisition module is used to acquire multi-lead EEG signals of patients to be identified as having impaired consciousness, and to preprocess and segment the multi-lead EEG signals to obtain N EEG segments. The feature extraction module is used to perform deep feature representation extraction on each of the EEG segments using a deep learning encoder, to obtain N feature vectors corresponding to the N EEG segments. The distribution vector module is used to process the N feature vectors using a distribution-aware automatic aggregation mechanism to obtain a distribution vector characterizing the heterogeneity of the patient's brain neural state. The obstacle identification module is used to identify the consciousness disorder by using an interpretable classifier on the distribution vector of the patient to be identified as having a consciousness disorder, and to obtain the classification result of the corresponding patient to be identified as having a consciousness disorder. The deep learning encoder is cascaded with the interpretable classifier, and a constraint threshold optimization mechanism is introduced to update the decision threshold for classification and recognition in the interpretable classifier.

8. The patient consciousness impairment identification system based on a distributed sensing automatic aggregation mechanism according to claim 7, characterized in that, Each feature vector includes D dimensional features, and the distribution vector module includes a calculation submodule, a feature pool determination submodule, a coefficient determination submodule, a weighted fusion submodule, and a splicing and construction submodule. The calculation submodule is used to calculate the mean and standard deviation corresponding to each dimension feature based on the number N of the feature vectors and each dimension feature of all the feature vectors; and to calculate the percentile corresponding to each dimension feature of all the feature vectors using the Percentile function based on multiple preset thresholds of different values, so as to obtain the percentile corresponding to each dimension feature under each preset threshold. The feature pool determination submodule is used to determine the basic statistical feature pool corresponding to each of the dimensional features based on the mean, standard deviation and all percentiles of each dimensional feature. The basic statistical feature pool includes K basic feature data. The coefficient determination submodule is used to determine the coefficient of variation corresponding to each basic feature data in all the basic statistical feature pools. The weighted fusion submodule is used to perform weighted fusion processing on the basic statistical feature pool of each dimension feature according to all the coefficients of variation, so as to obtain the distribution manifold vector corresponding to each dimension feature; The splicing submodule is used to splice all the distribution manifold vectors in the order of dimensional features to obtain a K×D dimensional individual EEG distribution manifold vector as the distribution vector.

9. The patient consciousness impairment identification system based on a distributed sensing automatic aggregation mechanism according to claim 7, characterized in that, The constraint threshold optimization mechanism includes: Obtain sensitivity constraint thresholds, several candidate decision thresholds, and sample data from M patients; The deep learning encoder and the distribution-aware automatic aggregation mechanism are used to process the sample data of each patient to obtain the patient distribution vector corresponding to each patient; Based on each candidate decision threshold and the patient distribution vectors of M patients, the interpretable classifier is used to identify disorders of consciousness, and classification evaluation data corresponding to the candidate decision threshold is obtained. The classification evaluation data includes the number of first correct prediction samples of vegetative state, the number of first incorrect prediction samples of vegetative state, the number of second correct prediction samples of minimally conscious state, and the number of second incorrect prediction samples of minimally conscious state. The sensitivity corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples and the number of first incorrectly predicted samples for each candidate decision threshold; the classification accuracy corresponding to the candidate decision threshold is calculated based on the number of first correctly predicted samples, the number of first incorrectly predicted samples, the number of second correctly predicted samples, and the number of second incorrectly predicted samples for each candidate decision threshold. Based on the sensitivity constraint threshold and the sensitivity, all the candidate decision thresholds are filtered to obtain a better decision threshold that satisfies the sensitivity constraint threshold; Based on the classification accuracy, all the better decision thresholds are filtered using the argmax function to obtain the decision thresholds for classification and recognition in the interpretable classifier.

10. A terminal device, characterized in that, Including the processor and memory; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute, according to the instructions in the program code, the patient consciousness impairment identification method based on the distribution-aware automatic aggregation mechanism as described in any one of claims 1-6.

Images