Chronic Doc patient grading diagnosis method and system

Abstract

The invention provides a chronic Doc patient grading diagnosis method and system. The method comprises the following steps: collecting multi-modal data from a chronic Doc patient; performing data cleaning on the multi-modal data to obtain cleaned data; extracting key feature vectors from the cleaning data of different modes by using a multi-set canonical correlation analysis algorithm; fusing the key feature vectors through a deep learning model, and optimizing weight distribution among the key feature vectors to obtain multi-modal feature vectors; training a preset grading diagnosis model based on the multi-modal data sample library; and inputting the multi-modal feature vector into a trained hierarchical diagnosis model to obtain a hierarchical diagnosis result. Through multi-modal data fusion, deep learning optimization and intelligent diagnosis model construction, scientificity, accuracy and clinical practicability of chronic Doc patient grading diagnosis are remarkably improved, and important support is provided for personalized treatment and intelligent diagnosis and treatment.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical treatment, and particularly to a method and system for grading diagnosis of chronic Doc patients. Background Art

[0002] The assessment of the consciousness level and prognosis prediction of patients with Chronic Disorder of Consciousness (Doc) are core issues in clinical treatment. However, the current diagnostic methods mainly rely on the analysis of single-modal data (such as scale scores, electroencephalograms, or imaging data), and there are the following problems:

[0003] Data singularity: Existing methods usually independently analyze data such as electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI), lacking the effective fusion of multi-modal data, resulting in low accuracy and stability of diagnostic results.

[0004] Strong subjectivity: Doctor experience and scale scores play a dominant role in the assessment, and misdiagnosis or missed diagnosis are likely to occur.

[0005] Lack of personalized treatment: Existing treatment plans are difficult to be dynamically adjusted according to the specific conditions of patients, resulting in limited treatment effects. Summary of the Invention

[0006] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a method and system for grading diagnosis of chronic Doc patients. By means of multi-modal data fusion, deep learning optimization, and the construction of an intelligent diagnostic model, the scientificity, accuracy, and clinical practicability of the grading diagnosis of chronic Doc patients are significantly improved, providing important support for personalized treatment and intelligent diagnosis and treatment.

[0007] To achieve the above purpose, the present invention provides the following solutions:

[0008] A method for grading diagnosis of chronic Doc patients, comprising:

[0009] Collect multi-modal data from chronic Doc patients; the multi-modal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical information;

[0010] Perform data cleaning on the multi-modal data to obtain cleaned data;

[0011] Extract key feature vectors from the cleaned data of different modalities by using the multi-set canonical correlation analysis algorithm;

[0012] Fuse the key feature vectors through a deep learning model, and optimize the weight distribution between the key feature vectors to obtain multi-modal feature vectors;

[0013] Train a preset hierarchical diagnosis model based on a multi-modal data sample library;

[0014] Input the multi-modal feature vector into the trained hierarchical diagnosis model to obtain a hierarchical diagnosis result.

[0015] Preferably, perform data cleaning on the multi-modal data to obtain cleaned data, including:

[0016] Group the multi-modal data according to a preset acquisition period to obtain multiple data groups;

[0017] Calculate the difference coefficient between the current data group and the previous data group in sequence;

[0018] Judge whether the value of the difference coefficient is within a preset range;

[0019] If the value of the difference coefficient is not within the preset range, remove the corresponding data group;

[0020] If the value of the difference coefficient is within the preset range, retain the corresponding data group until all data groups are traversed to obtain the cleaned data.

[0021] Preferably, the difference coefficient calculation formula is:

[0022]

[0023] where p X,Y is the difference coefficient, cov(X,Y) represents the covariance between the current data group X and the previous data group Y, α X represents the mean of the current data group X, and β Y represents the mean of the previous data group Y.

[0024] Preferably, use the multi-set canonical correlation analysis algorithm to extract key feature vectors from the cleaned data of different modalities, including:

[0025] Construct a joint covariance matrix C according to the cleaned data of different modalities; where C ij is the covariance matrix of the modality data set i and the modality data set j in the cleaned data; K is the number of modality data sets, and the dimension of each modality data set X k is m×d k where n is the number of samples, and d k is the feature dimension of the kth modality;

[0026] Solve the joint covariance matrix through generalized eigenvalue decomposition to obtain a projection matrix W k to maximize the correlation between modalities; the projected key feature vector Yk Expressed as: Y k = X k W k , k = 1, 2, …, K.

[0027] Preferably, the key feature vectors are fused through a deep learning model, and the weight distribution among the key feature vectors is optimized to obtain multi-modal feature vectors, including:

[0028] All the key feature vectors Y of all modalities k are weighted and fused to obtain multi-modal feature vectors; the expression of the multi-modal feature vectors is: where F is the multi-modal feature vector, and α k is the weight of the k-th modality, with the initial value being uniformly distributed; the weight of the k-th modality is dynamically optimized through a deep learning model.

[0029] Preferably, the deep learning model is a Transformer network.

[0030] Preferably, the multi-modal data sample library includes 800 - 1000 cases of patient data.

[0031] Preferably, the hierarchical diagnosis model is a CNN-LSTM combined model.

[0032] A hierarchical diagnosis system for chronic Doc patients, comprising:

[0033] A data acquisition module for acquiring multi-modal data from chronic Doc patients; the multi-modal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical information;

[0034] A data cleaning module for cleaning the multi-modal data to obtain cleaned data;

[0035] A feature extraction module for extracting key feature vectors from the cleaned data of different modalities by using a multi-set canonical correlation analysis algorithm;

[0036] A vector fusion module for fusing the key feature vectors through a deep learning model and optimizing the weight distribution among the key feature vectors to obtain multi-modal feature vectors;

[0037] A model training module for training a preset hierarchical diagnosis model based on a multi-modal data sample library;

[0038] A diagnosis module for inputting the multi-modal feature vectors into the trained hierarchical diagnosis model to obtain a hierarchical diagnosis result.

[0039] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0040] The present invention provides a method and system for grading and diagnosing chronic Doc patients. The method includes: collecting multimodal data from chronic Doc patients; the multimodal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical information; cleaning the multimodal data to obtain cleaned data; using a multi-set canonical correlation analysis algorithm to extract key feature vectors from the cleaned data of different modalities; fusing the key feature vectors through a deep learning model, and optimizing the weight distribution between the key feature vectors to obtain multimodal feature vectors; training a preset grading and diagnosis model based on a multimodal data sample library; inputting the multimodal feature vectors into the trained grading and diagnosis model to obtain a grading and diagnosis result. Through multimodal data fusion, deep learning optimization, and the construction of an intelligent diagnosis model, the present invention significantly improves the scientificity, accuracy, and clinical practicability of grading and diagnosing chronic Doc patients, providing important support for personalized treatment and intelligent diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0042] Figure 1 It is a flowchart of the method provided by the embodiment of the present invention;

[0043] Figure 2 It is a schematic structural diagram of the system provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0045] The purpose of the present invention is to provide a method and system for grading and diagnosing chronic Doc patients. Through multimodal data fusion, deep learning optimization, and the construction of an intelligent diagnosis model, the present invention significantly improves the scientificity, accuracy, and clinical practicability of grading and diagnosing chronic Doc patients, providing important support for personalized treatment and intelligent diagnosis and treatment.

[0046] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 It is a flowchart of the method provided by the embodiment of the present invention. As Figure 1 shown, the present invention provides a method for grading and diagnosing chronic Doc patients, including:

[0048] Step 100: Collect multimodal data from chronic Doc patients; the multimodal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical information;

[0049] Step 200: Clean the multimodal data to obtain cleaned data;

[0050] Step 300: Extract key feature vectors from the cleaned data of different modalities using the multi-set canonical correlation analysis algorithm;

[0051] Step 400: Fuse the key feature vectors through a deep learning model and optimize the weight distribution between the key feature vectors to obtain multimodal feature vectors;

[0052] Step 500: Train a preset grading and diagnosis model based on the multimodal data sample library;

[0053] Step 600: Input the multimodal feature vectors into the trained grading and diagnosis model to obtain a grading and diagnosis result.

[0054] Specifically, step 100 of this embodiment includes:

[0055] Step 101: Collection of brain imaging data

[0056] Brain imaging data is an important basis for the diagnosis of chronic Doc patients, mainly including magnetic resonance imaging (MRI) and functional magnetic resonance imaging (fMRI). MRI is used to obtain anatomical structure information of the patient's brain, such as the morphological changes of gray matter, white matter, and ventricles; fMRI is used to detect brain function activities and evaluate the patient's brain network connectivity and neural activity level. The collection process is usually carried out on a high-field-strength (such as 3T) magnetic resonance device, and the patient needs to remain stationary to reduce motion artifacts. The collection parameters (such as resolution, time series length) need to be optimized according to clinical needs, and at the same time, a standardized brain imaging collection protocol (such as DICOM format) is combined to ensure data quality and consistency.

[0057] Step 102: Collection of neuroelectrophysiological data

[0058] Neuroelectrophysiological data are mainly collected through electroencephalogram (EEG) or event-related potential (ERP) techniques, which are used to evaluate the patient's neural activity and level of consciousness. When collecting EEG, an electrode cap is worn on the patient's head, and electroencephalogram signals are recorded through multiple leads, focusing on the distribution and intensity of slow-wave activities (such as δ waves, θ waves) and high-frequency activities (such as γ waves). ERP, on the other hand, induces electroencephalogram responses through specific stimuli (such as sounds, lights) to analyze the patient's neural responses to external stimuli. During the collection process, highly sensitive amplifiers and filters are used to remove artifacts (such as electromyogram, eye movement artifacts), and electromagnetic shielding of the collection environment is ensured to improve the signal quality.

[0059] Step 103: Collection of biochemical indicators

[0060] Biochemical indicators are obtained through the detection of blood, urine, or cerebrospinal fluid, which are used to evaluate the patient's metabolic status and neurotransmitter levels. Blood tests can analyze the concentrations of inflammatory factors (such as C-reactive protein, interleukin), metabolites (such as glucose, lactate), and neurotransmitters (such as dopamine, glutamate); cerebrospinal fluid tests can provide more direct information on the metabolism of the central nervous system, such as the levels of β-amyloid protein and tau protein. During the collection process, strict aseptic operation specifications must be followed to ensure the integrity and accuracy of the samples, and standardized laboratory detection methods (such as ELISA, mass spectrometry) are used for quantitative analysis.

[0061] Step 104: Collection of clinical data

[0062] Clinical data include the patient's medical history, scores of consciousness assessment scales (such as Glasgow Coma Scale, CRS-R score), and treatment records, etc. When collecting the medical history, the cause of the disease (such as brain trauma, hypoxic encephalopathy), the course of the disease, and previous treatment conditions of the patient need to be recorded in detail; the consciousness assessment scale evaluates the patient's level of consciousness and response ability through standardized behavioral tests; the treatment records include information such as drug use, rehabilitation training, and surgical intervention. All clinical data need to be digitally stored through an electronic medical record system (EMR) to ensure the integrity and traceability of the data, providing a reliable basis for subsequent diagnosis and analysis.

[0063] Through the collection of the above four types of data in this embodiment, it is possible to comprehensively cover the brain structure, function, metabolism, and clinical manifestations of chronic Doc patients, providing a multi-dimensional scientific basis for subsequent hierarchical diagnosis and personalized treatment.

[0064] Preferably, the multi-modal data is cleaned to obtain cleaned data, including:

[0065] Group the multi-modal data according to a preset collection period to obtain multiple data groups;

[0066] Calculate the coefficient of difference between the current data group and the previous data group in sequence;

[0067] Determine whether the value of the coefficient of difference is within a preset range;

[0068] If the value of the coefficient of difference is not within the preset range, remove the corresponding data group;

[0069] If the value of the coefficient of difference is within the preset range, retain the corresponding data group until all data groups are traversed to obtain the cleaned data.

[0070] Preferably, the formula for calculating the coefficient of difference is:

[0071]

[0072] where p X,Y is the coefficient of difference, cov(X,Y) represents the covariance between the current data group X and the previous data group Y, α X represents the mean value of the current data group X, and β Y represents the mean value of the previous data group Y.

[0073] Specifically, step 200 of this embodiment includes:

[0074] Step 201: Group the collected multimodal data according to a preset collection period, and divide the data within a continuous time period into multiple data groups. Each data group includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical materials collected within this time period. The purpose of grouping is to facilitate subsequent difference analysis and data cleaning operations, and at the same time ensure the time continuity and logical consistency of the data. After grouping, the data groups are arranged in chronological order, ready for calculation of the coefficient of difference.

[0075] Step 202: Calculate the coefficient of difference between the current data group and the previous data group in sequence. The calculation of the coefficient of difference is based on the covariance and the change of the mean values of the two groups of data, and is used to measure the similarity or difference between the two groups of data. In this way, abnormal data groups can be effectively identified, such as data anomalies caused by malfunction of the acquisition device or external interference. After calculation, compare the coefficient of difference of each data group with the preset range to determine whether it meets the normal range.

[0076] Step 203: Screen the data groups according to the judgment result of the difference coefficient. If the difference coefficient of a certain data group is not within the preset range, it is considered that the data group is abnormal and needs to be removed from the data set; if the difference coefficient is within the preset range, the data group is retained. By traversing all data groups, abnormal data is gradually cleaned, and finally the cleaned multi-modal data is obtained. These cleaned data will be used as high-quality inputs for subsequent feature extraction and model training to ensure the accuracy and reliability of the diagnosis results.

[0077] Preferably, the multi-set canonical correlation analysis algorithm is used to extract key feature vectors from the cleaned data of different modalities, including:

[0078] Construct a joint covariance matrix C according to the cleaned data of different modalities; where C ij is the covariance matrix of modality data set i and modality data set j in the cleaned data; K is the number of modality data sets, and each modality data set X k has a dimension of n×d k , where n is the number of samples, and d k is the feature dimension of the kth modality;

[0079] Solve the joint covariance matrix through generalized eigenvalue decomposition to obtain the projection matrix W k , so as to maximize the correlation between modalities; the projected key feature vector Y k is expressed as: Y k =X k W k , k = 1, 2, …, K.

[0080] Specifically, step 300 of this embodiment includes:

[0081] Step 301: Construct a joint covariance matrix

[0082] First, extract the data set of each modality from the cleaned multi-modal data and calculate the covariance matrix between modalities. Each modality data set represents a specific type of data (such as brain imaging data, neuroelectrophysiological data, biochemical indicators, etc.). After cleaning and standardization, these data ensure the dimensional consistency of different modality data and the minimization of noise. Then, for each pair of modality data sets, calculate the covariance matrix between them, and the covariance matrix reflects the linear correlation between modality data. Combine all the covariance matrices between modalities to construct a joint covariance matrix C, which can comprehensively describe the correlation structure between multi-modal data and provide a basis for subsequent feature extraction.

[0083] Step 302: Solve the projection matrix to maximize the correlation between modalities

[0084] After constructing the joint covariance matrix, it is decomposed using the generalized eigenvalue decomposition method. The aim is to find a set of projection matrices such that different modal data has the maximum correlation in the projected feature space. The process of eigenvalue decomposition can be regarded as finding an optimal linear transformation that maps the original modal data into a common subspace. In this subspace, the data features of different modalities can be aligned and correlated, thereby maximizing the correlation between them. The core of this step is to optimize the projection matrix through eigenvalue decomposition so that the correlation information between modalities is fully retained while removing redundant information and noise.

[0085] Step 303: Extract the key feature vectors after projection

[0086] The datasets of each modality are linearly transformed through the projection matrix to obtain the key feature vectors after projection. These feature vectors are the representations of multi-modal data in the common subspace, which can capture the maximum correlation and the most discriminative feature information between modalities. The key feature vectors not only retain the main information of the original data but also eliminate the redundancy and noise between modalities through the projection process, enabling the data of different modalities to be fused and analyzed in a unified feature space. These feature vectors will be used as the input of the subsequent deep learning model for further optimization and fusion, thus supporting the construction of a hierarchical diagnosis and prognosis prediction model.

[0087] Through the above steps, this embodiment can effectively extract the key feature vectors of multi-modal data using the multi-set canonical correlation analysis algorithm, providing high-quality feature input for subsequent intelligent diagnosis and treatment.

[0088] Preferably, the key feature vectors are fused through a deep learning model, and the weight distribution between the key feature vectors is optimized to obtain multi-modal feature vectors, including:

[0089] The key feature vectors Y of all modalities k are weighted and fused to obtain multi-modal feature vectors; the expression of the multi-modal feature vectors is: where F is the multi-modal feature vector, and α k is the weight of the k-th modality, with an initial value of uniform distribution; the weight of the k-th modality is dynamically optimized through a deep learning model.

[0090] Preferably, the deep learning model is a Transformer network.

[0091] Furthermore, in the initial stage of fusing multi-modal key feature vectors, the weights α of all modalities kIs set to be evenly distributed. This initial allocation method ensures that the data of each modality has the same importance in the initial fusion process, avoiding biases caused by overly strong or weak features of a certain modality. The evenly distributed weights provide a fair starting point for subsequent dynamic optimization and ensure that the model can fully utilize the information of all modalities in the initial stage. Based on the initial weight allocation, the Transformer network performs a correlation analysis on the key feature vectors of different modalities through its multi-head self-attention mechanism. The multi-head self-attention mechanism can capture the global dependencies and potential interaction information between modalities, thereby identifying which modality features are more important for the current task (such as hierarchical diagnosis or prognosis prediction). Through this analysis, the model can dynamically adjust the weights of each modality, enabling modalities that contribute more to the task to obtain higher weights, while the weights of modalities that contribute less gradually decrease. Based on the correlation analysis, the Transformer network further optimizes the modality weights through its feed-forward neural network module. Specifically, the network dynamically adjusts its weight allocation according to the performance of each modality's features in the global feature fusion. This optimization process is achieved through the backpropagation algorithm. The model calculates the gradients based on the loss function (such as cross-entropy loss in classification tasks or mean squared error in regression tasks) and updates the network parameters through an optimizer (such as Adam), thereby gradually optimizing the allocation of modality weights. Finally, the weight α k Will reflect the actual contribution of each modality to the task. During the entire training process, the Transformer network continuously adjusts the modality weights dynamically according to the changes in the input data and the learning situation of the model. This dynamic adjustment mechanism ensures that the model can adapt to the individual differences of different patients. For example, for some patients, brain imaging data may better reflect their consciousness state, while for other patients, neuroelectrophysiological data may be more critical. Through this dynamic adjustment, the model can optimize the weight allocation in real time, enabling the finally generated multi-modal feature vectors to maximally capture the patient's consciousness state and prognosis information, thereby improving the accuracy and robustness of hierarchical diagnosis and prognosis prediction.

[0092] Preferably, the multi-modal data sample library includes 800 - 1000 cases of patient data.

[0093] Specifically, step 500 of this embodiment first extracts the multi-modal data of 800 - 1000 patients from the multi-modal data sample library, including brain imaging data (MRI, fMRI, PET, DTI, etc.), neuroelectrophysiological data (EEG, ERP, etc.), biochemical indicators (blood tests, etc.), and clinical information (medical records, scale scores, etc.). Standardize and preprocess this data to ensure data quality and consistency:

[0094] Data cleaning: Remove noise, outliers, and redundant information. For example, perform motion correction and artifact removal on imaging data, and use independent component analysis (ICA) to remove artifacts from EEG data.

[0095] Data formatting: Convert data of different modalities into a unified format (such as tensor form) for input into a deep learning model.

[0096] Data augmentation: Expand the sample size through data augmentation techniques (such as image rotation, cropping, random truncation of time series data, etc.) to improve the generalization ability of the model.

[0097] Data splitting: Divide the sample library into a training set (about 70%), a validation set (about 20%), and a test set (about 10%) to ensure the independence of model training and evaluation.

[0098] Secondly, the CNN-LSTM combined model combines the advantages of convolutional neural networks (CNN) and long short-term memory networks (LSTM), and is suitable for processing the spatial features and time series features of multimodal data; CNN is used to extract the spatial features of multimodal data, especially the local features and patterns in imaging data (such as MRI, fMRI). Design multiple convolutional layers and pooling layers to gradually extract high-level features. For example, the convolutional layer is used to extract low-level features such as edges and textures, and the deep convolutional layer extracts more complex patterns. The feature map output by the convolutional layer is converted into a feature vector through a fully connected layer and used as the input of the LSTM module. LSTM is used to process time series data (such as EEG, ERP, follow-up data, etc.) to capture the characteristics of the patient's dynamic changes. The input is the feature vector extracted by CNN and the time series data. LSTM captures long-term dependencies through memory units and gating mechanisms and extracts time dynamic features. The output of LSTM is a time series feature vector, which is combined with the spatial feature vector of CNN to form the final feature representation. The output feature vectors of CNN and LSTM are fused (such as concatenation or weighted average) to generate a comprehensive feature representation. The fused feature vector is input into a fully connected layer, and finally the hierarchical diagnosis result is output through the Softmax layer.

[0099] Based on the multimodal data sample library, train the CNN-LSTM combined model. The specific steps are as follows:

[0100] Use the cross-entropy loss function to calculate the error between the model prediction result and the true label, which is suitable for classification tasks. Use the Adam optimizer to update the parameters, combined with adaptive learning rate adjustment to improve the training efficiency.

[0101] The initial value of the learning rate is set to 0.001, and a learning rate decay strategy is adopted (e.g., decaying once every 10 epochs). A Dropout layer is added to the CNN and LSTM modules to prevent overfitting. L2 regularization is used to constrain the model parameters to further improve the generalization ability of the model. In this embodiment, the training set data is input into the model in batches (batch) for forward propagation and backward propagation, and the model parameters are gradually optimized. The performance of the model is evaluated on the validation set, and hyperparameters (such as the convolutional kernel size, the number of LSTM units, etc.) are adjusted to optimize the model performance. After training, the CNN-LSTM combined model is evaluated and applied to ensure its reliability and practicality in hierarchical diagnosis; in this embodiment, the hierarchical diagnosis accuracy, recall rate, F1 score and other indicators of the model are evaluated on the test set, and the goal is that the hierarchical evaluation accuracy is not less than 85%. The confusion matrix is used to analyze the performance of the model at different levels to identify possible misdiagnosis and missed diagnosis situations. In this embodiment, the trained model is deployed to the clinical diagnosis system, and the multi-modal data of the patient is input to output the hierarchical diagnosis result. Combining the hierarchical result of the model, a personalized treatment plan (such as maintenance treatment, drug treatment, nerve stimulation, etc.) is formulated for the patient. In actual applications, new data is continuously collected to fine-tune and update the model to improve the adaptability and robustness of the model.

[0102] Specifically, the CNN-LSTM combined model of this embodiment can make full use of the spatial features and temporal dynamic features of multi-modal data to achieve high-precision hierarchical diagnosis of chronic Doc patients. The design and training process of the model pays attention to data quality, feature extraction and optimization strategies to ensure the accuracy and clinical practicality of the hierarchical diagnosis result and provide a scientific basis for personalized treatment.

[0103] Corresponding to the above method, as Figure 2 shown, this embodiment also provides a hierarchical diagnosis system for chronic Doc patients, including:

[0104] A data acquisition module for acquiring multi-modal data from chronic Doc patients; the multi-modal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical materials;

[0105] A data cleaning module for cleaning the multi-modal data to obtain cleaned data;

[0106] A feature extraction module for extracting key feature vectors from the cleaned data of different modalities by using a multi-set canonical correlation analysis algorithm;

[0107] A vector fusion module for fusing the key feature vectors through a deep learning model and optimizing the weight distribution between the key feature vectors to obtain multi-modal feature vectors;

[0108] A model training module for training a preset hierarchical diagnosis model based on a multi-modal data sample library;

[0109] A diagnosis module for inputting a multi-modal feature vector into the trained hierarchical diagnosis model to obtain a hierarchical diagnosis result.

[0110] The beneficial effects of the present invention are as follows:

[0111] (1) By collecting multi-modal data (brain imaging data, neuroelectrophysiological data, biochemical indicators, and clinical information), and using the multi-set canonical correlation analysis algorithm to extract key feature vectors, the present invention solves the problems of unstable diagnosis results and high misdiagnosis rates caused by traditional single-modal data analysis methods; the deep learning model fuses and optimizes the weights of multi-modal feature vectors, fully exploiting the complementarity and potential correlation between multi-modal data, and significantly improving the refinement and accuracy of hierarchical diagnosis.

[0112] (2) Through a standardized data cleaning process and the multi-set canonical correlation analysis algorithm, the present invention solves the problems of data inconsistency and information loss existing in the fusion process of multi-source heterogeneous data (such as images, neuroelectrophysiology, biochemical indicators, etc.); the deep learning model further optimizes the weight distribution between feature vectors, breaking through the bottleneck of insufficient multi-modal data fusion in the prior art.

[0113] (3) Based on the hierarchical diagnosis result and the multi-modal data sample library, the present invention can accurately evaluate the patient's consciousness level and recovery potential, providing a scientific basis for patient grouping (such as maintenance treatment, drug treatment, non-surgical nerve stimulation, and neuromodulation surgery); through intelligent hierarchical diagnosis and prognosis prediction, it helps doctors formulate personalized treatment plans, avoiding over-treatment or delayed treatment, and reducing medical costs and social burdens.

[0114] (4) The multi-modal data sample library and intelligent fusion model constructed by the present invention provide a data-driven scientific basis for the diagnosis and treatment of chronic Doc, promoting the application of artificial intelligence in the diagnosis and treatment of chronic Doc; through an intelligent hierarchical diagnosis model, a new strategy for the intelligent diagnosis and treatment of chronic Doc is developed, providing important support for the rehabilitation care and improvement of the quality of life of patients.

[0115] (5) The accurate hierarchical diagnosis of the present invention can help the patient's family members understand the patient's recovery potential earlier, reducing emotional suffering and resource waste; for patients with no hope of recovery, accurate prognosis assessment can avoid unnecessary treatment and save medical resources.

[0116] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts among the various embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description in the method section.

[0117] In this article, specific examples are used to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present invention.

Claims

1. A method for grading diagnosis of chronic Doc patients, characterized in that: include: Collecting multimodal data from chronic Doc patients; the multimodal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators and clinical data; Performing data cleaning on the multimodal data to obtain cleaned data; Extracting key feature vectors from the cleaned data of different modalities using a multi-set canonical correlation analysis algorithm; The key feature vectors are fused through a deep learning model, and the weight distribution between the key feature vectors is optimized to obtain a multimodal feature vector; A preset hierarchical diagnosis model is trained based on a multimodal data sample library; The multimodal feature vector is input into the trained hierarchical diagnosis model to obtain a hierarchical diagnosis result.

2. The method for grading diagnosis of chronic Doc patients according to claim 1, characterized in that: Performing data cleaning on the multimodal data to obtain cleaned data includes: Grouping the multimodal data according to a preset collection period to obtain multiple data groups; Calculate the difference coefficient between the current data group and the previous data group in sequence; Determine whether the value of the coefficient of difference is within a preset range; If the value of the difference coefficient is not within the preset range, the corresponding data group is removed; If the value of the difference coefficient is within a preset range, the corresponding data group is retained until all data groups are traversed to obtain the cleaned data.

3. The method for grading diagnosis of chronic Doc patients according to claim 2, characterized in that: The coefficient of difference calculation formula is: Among them, p X,Y is the coefficient of difference, cov(X,Y) represents the covariance between the current data set X and the previous data set Y, α X represents the mean of the current data set X, β Y Represents the mean of the previous data set Y.

4. The method for grading diagnosis of chronic Doc patients according to claim 1, characterized in that: The key feature vectors are extracted from the cleaned data of different modes using a multi-set canonical correlation analysis algorithm, including: Constructing a joint covariance matrix C based on the cleaned data of different modalities; Among them, C ij is the covariance matrix of modal data set i and modal data set j in the cleaned data; K is the number of modal data sets, and each modal data set X k The dimension is n×d k , where n is the number of samples, d k is the characteristic dimension of the kth mode; The joint covariance matrix is ​​solved by generalized eigenvalue decomposition to obtain the projection matrix W k , in order to maximize the correlation between modes; the key eigenvector Y after projection k Expressed as: Y k =X k W k ,k=1,2,…,K。 5. The method for grading diagnosis of chronic Doc patients according to claim 4, characterized in that: The key feature vectors are fused through a deep learning model, and the weight distribution between the key feature vectors is optimized to obtain a multimodal feature vector, including: The key eigenvectors Y of all modes k Perform weighted fusion to obtain a multimodal feature vector; the expression of the multimodal feature vector is: Where, F is the multimodal feature vector, α k is the weight of the k-th mode, and its initial value is uniformly distributed; the weight of the k-th mode is dynamically optimized through a deep learning model.

6. The method for grading diagnosis of chronic Doc patients according to claim 5, characterized in that: The deep learning model is a Transformer network.

7. The method for grading diagnosis of chronic Doc patients according to claim 1, characterized in that: The multimodal data sample library includes 800-1000 patient data.

8. The method for grading diagnosis of chronic Doc patients according to claim 1, characterized in that: The hierarchical diagnosis model is a CNN-LSTM combined model.

9. A hierarchical diagnosis system for chronic Doc patients, characterized in that: include: A data acquisition module, used to collect multimodal data from chronic Doc patients; the multimodal data includes brain imaging data, neuroelectrophysiological data, biochemical indicators and clinical data; A data cleaning module, used to clean the multimodal data to obtain cleaned data; A feature extraction module, used to extract key feature vectors from the cleaned data of different modes using a multi-set canonical correlation analysis algorithm; A vector fusion module, used to fuse the key feature vectors through a deep learning model and optimize the weight distribution between the key feature vectors to obtain a multimodal feature vector; A model training module is used to train a preset hierarchical diagnosis model based on a multimodal data sample library; The diagnosis module is used to input the multimodal feature vector into the trained hierarchical diagnosis model to obtain a hierarchical diagnosis result.

Images