Prognosis prediction method and system for chronic Doc patient

Abstract

The invention provides a prognosis prediction method and system for a chronic Doc patient, relates to the technical field of clinical prognosis evaluation, and is characterized in that multi-modal data such as neuroimaging, neuroelectrophysiological data, biochemical indexes and clinical data are collected, standardized processing and feature extraction are carried out, and spatial features, time sequence features and key features are fused. And in combination with the trained consciousness state grading evaluation model and the prognosis prediction model, generating a feature level data set, and accurately predicting the patient consciousness state grading and prognosis development trend. According to the method, complementary information of multi-modal data is effectively integrated, the accuracy and reliability of prognosis prediction are improved, and scientific support is provided for clinical decision making, personalized treatment and medical resource optimization.

Description

Technical Field

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

[0002] The consciousness assessment and prognosis prediction of patients with Chronic Disorder of Consciousness (Doc) are core issues in clinical treatment. However, existing single-modal data analysis methods are difficult to meet complex clinical needs, resulting in low accuracy of diagnosis and prognosis assessment, high misdiagnosis rates, and a lack of personalized treatment strategies.

[0003] In recent years, multi-modal brain data fusion technology has gradually become a research hotspot. By integrating multi-dimensional indicators such as brain nerve circuits, metabolism, connectivity, electrophysiology, and biochemistry, hidden information can be deeply mined to reveal the diversity of brain network connectivity and neural activities. However, there are still few domestic and foreign studies on multi-modal brain data fusion models, especially in the hierarchical assessment and prognosis prediction of chronic Doc patients, and no standardized and intelligent solutions have been formed. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a prognosis prediction method and system for chronic Doc patients, which can accurately predict the consciousness state grading and prognosis development trend of patients, provide scientific decision-making support for clinicians, optimize personalized treatment plans, improve the accuracy and reliability of prognosis prediction, and thus enhance the rehabilitation effect of patients and the utilization efficiency of medical resources.

[0005] To achieve the above purpose, the present invention provides the following solutions:

[0006] A prognosis prediction method for chronic Doc patients, comprising:

[0007] Collect multi-modal data of chronic Doc patients, and perform standardization processing on each type of the multi-modal data respectively to obtain a standardized data set; the multi-modal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators, and clinical data;

[0008] Extract features from each type of data in the standardized data set to obtain spatial features corresponding to the neuroimaging data, time series features corresponding to the neuroelectrophysiological data, and key features corresponding to the biochemical indicators and clinical data;

[0009] Perform feature fusion on the spatial features, the time series features, and the key features to obtain a fused feature vector;

[0010] Input the fused feature vector into the trained consciousness state grading evaluation model to obtain a classification result and the features of the middle layer of the model;

[0011] Use the classification result as a label and generate a feature-level data set in combination with the features of the middle layer of the model;

[0012] Input the feature-level data set into the trained patient prognosis prediction model to obtain a prediction result.

[0013] Preferably, the process of the standardization processing includes: denoising, normalization, and format conversion.

[0014] Preferably, perform feature extraction on various types of data in the standardized data set to obtain the spatial features corresponding to neuroimaging data, the time series features corresponding to neuroelectrophysiological data, and the key features corresponding to biochemical indicators and clinical data, including:

[0015] For neuroimaging data, use a convolutional neural network to extract spatial features;

[0016] For neuroelectrophysiological data, use time-frequency analysis to extract time series features;

[0017] For biochemical indicators and clinical data, use a random forest algorithm to extract key features.

[0018] Preferably, for neuroelectrophysiological data, use a time-frequency analysis method to extract time series features, including:

[0019] Use the short-time Fourier transform method to extract time-frequency features; the expression of the time-frequency features is: where x(τ) is the original neuroelectrophysiological data, w(τ - t, σ) is the time window function, t is the time point, representing the local spectrum of the signal at different time periods, f is the frequency, representing the frequency component of the signal, X(t, f) is the time-frequency feature, representing the frequency distribution of the signal at each time point, and σ is the width of the window function, dynamically adjusted as: α is an adjustment parameter for controlling the sensitivity of the window width to the signal amplitude;

[0020] Extract the power spectral density according to the time-frequency features and normalize the power spectral density to obtain the normalized power spectral density; the expression of the power spectral density is: P(t, f) = |X(t, f)| 2 ; where P(t, f) is the power spectral density, representing the energy intensity of the signal at time t and frequency f; the expression of the normalized power spectral density is: where f min and f max are respectively the minimum and maximum values of the analysis frequency range, and P norm(t, f) is the normalized power spectral density;

[0021] Nonlinear feature analysis is performed according to the short-time multi-scale entropy to obtain the frequency-weighted multi-scale entropy; wherein, the expression of the short-time multi-scale entropy is wherein, E(t, τ) is the short-time multi-scale entropy at time t and time scale τ, which is used to quantify the complexity of the signal and reflect the dynamic change pattern of the neuroelectrophysiological signal. τ is the time scale, representing the size of the window, and p i (τ) is the probability distribution of the signal at time scale τ; the expression of the frequency-weighted multi-scale entropy is: E f (t, τ) = ∑ f w(f)E(t, τ, f); wherein, E(t, τ, f) is the short-time multi-scale entropy at frequency f, w(f) is the frequency weight function, which is adaptively adjusted according to the importance of a specific frequency band, and E f (t, τ) is the frequency-weighted multi-scale entropy;

[0022] The normalized power spectral density is transformed into a time-frequency weighted feature; the expression of the time-frequency weighted feature is: wherein, f β is the frequency weighting term. When β > 0, the high-frequency components are emphasized. When β < 0, the low-frequency components are emphasized. -F(t) is the time-frequency weighted feature, representing the projection of the time-frequency feature in the time dimension; σ'(t) is the signal local standard deviation of the neuroelectrophysiological data, and μ(t) is the signal local mean of the neuroelectrophysiological data;

[0023] The time-frequency feature, the normalized power spectral density, the frequency-weighted multi-scale entropy, and the time-frequency weighted feature are subjected to multi-dimensional feature splicing to obtain the fused time series feature.

[0024] Preferably, the key features include: blood glucose, lactate dehydrogenase, C-reactive protein, medical history, behavioral assessment scale, and medication record.

[0025] Preferably, the fused feature vector is input into the trained consciousness state grading evaluation model to obtain the classification result and the intermediate layer features of the model, including:

[0026] Construct an initial deep model based on a multi-layer perceptron;

[0027] Input the preset sample fusion feature vector and the sample awareness state level into the initial deep model to obtain a trained awareness state grading evaluation model; the loss function of the awareness state grading evaluation model is the cross-entropy loss function; the optimizer of the cross-entropy loss function is the Adam optimizer; the network structure of the awareness state grading evaluation model includes: an input layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence;

[0028] Input the fusion feature vector into the trained awareness state grading evaluation model to obtain a classification result and the features of the middle layer of the model; the features of the middle layer of the model are the output of the second fully connected layer, representing the deep features of the fusion feature vector.

[0029] Preferably, input the feature-level data set into the trained patient prognosis prediction model to obtain a prediction result, including:

[0030] Obtain a preset feature-level data sample set;

[0031] Construct an initial convolutional neural network;

[0032] Train the initial convolutional neural network according to the feature-level data sample set to obtain a trained classifier;

[0033] Connect a trained LSTM neural network after the classifier to obtain the patient prognosis prediction model;

[0034] Input the feature-level data set into the patient prognosis prediction model to obtain a prediction result.

[0035] Preferably, the prediction result includes: the recovery potential of the patient's awareness state, the prediction of the recovery time, the possibility of disease deterioration, and the specific prognosis grade.

[0036] Preferably, the prognosis grade includes: complete recovery, partial recovery, and no recovery.

[0037] A prognosis prediction system for chronic Doc patients, including:

[0038] A data normalization module for collecting multi-modal data of chronic Doc patients and performing normalization processing on each type of the multi-modal data to obtain a normalized data set; the multi-modal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators, and clinical data;

[0039] A feature extraction module for extracting features from each type of data in the normalized data set to obtain the spatial features corresponding to the neuroimaging data, the time series features corresponding to the neuroelectrophysiological data, and the key features corresponding to the biochemical indicators and clinical data;

[0040] A feature fusion module, configured to perform feature fusion on the spatial feature, the time series feature, and the key feature to obtain a fused feature vector;

[0041] A classification module, configured to input the fused feature vector into a trained consciousness state grading evaluation model to obtain a classification result and intermediate layer features of the model;

[0042] A feature-level generation module, configured to use the classification result as a label and generate a feature-level data set in combination with the intermediate layer features of the model;

[0043] A prediction module, configured to input the feature-level data set into a trained patient prognosis prediction model to obtain a prediction result.

[0044] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0045] The present invention provides a method and system for predicting the prognosis of chronic Doc patients. The method includes: collecting multimodal data of chronic Doc patients, and respectively performing standardization processing on each type of the multimodal data to obtain a standardized data set; the multimodal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators, and clinical data; performing feature extraction on each type of data in the standardized data set to obtain a spatial feature corresponding to the neuroimaging data, a time series feature corresponding to the neuroelectrophysiological data, and a key feature corresponding to the biochemical indicators and clinical data; performing feature fusion on the spatial feature, the time series feature, and the key feature to obtain a fused feature vector; inputting the fused feature vector into a trained consciousness state grading evaluation model to obtain a classification result and intermediate layer features of the model; using the classification result as a label and generating a feature-level data set in combination with the intermediate layer features of the model; inputting the feature-level data set into a trained patient prognosis prediction model to obtain a prediction result. By integrating multimodal data (including neuroimaging, neuroelectrophysiological data, biochemical indicators, and clinical data), the present invention realizes comprehensive feature extraction and fusion analysis of chronic Doc patients, and fully utilizes the complementary information of spatial features, time series features, and key features. Combining the trained consciousness state grading evaluation model and prognosis prediction model, the present invention can accurately predict the consciousness state grading and prognosis development trend of patients, provide scientific decision-making support for clinicians, optimize personalized treatment plans, improve the accuracy and reliability of prognosis prediction, and thus improve the rehabilitation effect of patients and the utilization efficiency of medical resources. Description of the Drawings

[0046] 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 for use 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 also be obtained based on these drawings.

[0047] Figure 1 It is the flowchart of the method provided by the embodiment of the present invention;

[0048] Figure 2 It is the schematic diagram of the system structure provided by the embodiment of the present invention. Specific embodiments

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. 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.

[0050] The purpose of the present invention is to provide a prognosis prediction method and system for chronic Doc patients, which can accurately predict the grading of the patient's consciousness state and the prognosis development trend, provide scientific decision-making support for clinicians, optimize personalized treatment plans, improve the accuracy and reliability of prognosis prediction, and thus improve the patient's rehabilitation effect and the utilization efficiency of medical resources.

[0051] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will further describe the present invention in detail in conjunction with the drawings and specific embodiments.

[0052] Figure 1 It is the flowchart of the method provided by the embodiment of the present invention. As Figure 1 shown, the present invention provides a prognosis prediction method for chronic Doc patients, including:

[0053] Step 100: Collect multi-modal data of chronic Doc patients, and perform standardization processing on each type of multi-modal data to obtain a standardized data set; the multi-modal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators, and clinical data;

[0054] Step 200: Extract features from each type of data in the standardized data set to obtain spatial features corresponding to neuroimaging data, time series features corresponding to neuroelectrophysiological data, and key features corresponding to biochemical indicators and clinical data;

[0055] Step 300: Fusing spatial features, time series features and key features to obtain a fused feature vector;

[0056] Step 400: Input the fused feature vector into the trained consciousness state classification assessment model to obtain the classification result and the model intermediate layer features;

[0057] Step 500: Use the classification results as labels and combine them with the model's intermediate layer features to generate a feature-level dataset;

[0058] Step 600: Input the feature-level data set into the trained patient prognosis prediction model to obtain a prediction result.

[0059] Preferably, the standardization process includes: denoising, normalization and format conversion.

[0060] Specifically, step 100 of this embodiment includes:

[0061] Step 101: First, for the multimodal data of chronic Doc patients, neuroimaging data (such as MRI, fMRI, DTI, etc.), neuroelectrophysiological data (such as EEG, ERP, etc.), biochemical indicators (such as blood glucose, lactate dehydrogenase, C-reactive protein, etc.) and clinical data (such as medical history, behavioral assessment scale, medication records, etc.) are collected. During the data collection process, the reliability and consistency of the data source must be ensured, and the data of different modalities must be preliminarily checked to eliminate obvious outliers and incomplete records, laying the foundation for subsequent processing.

[0062] Step 102: Next, each type of data is denoised. For neuroimaging data, filtering algorithms (such as Gaussian filtering or wavelet transform) can be used to remove image noise; for neuroelectrophysiological data, bandpass filters can be used to remove power frequency interference and low-frequency drift; for biochemical indicators and clinical data, statistical methods can be used to detect and remove outliers. After denoising, the data is normalized, for example, the pixel values ​​of the imaging data are normalized to the [0,1] interval, the amplitude of the neuroelectrophysiological data is standardized to zero mean and unit variance, and the biochemical indicators are scaled according to their normal range to ensure that data from different modalities are comparable.

[0063] Step 103: Finally, format conversion is performed to unify the data format. Neuroimaging data can be converted into standardized NIfTI or DICOM formats, neuroelectrophysiological data can be saved in EDF or MAT formats, and biochemical indicators and clinical data can be organized into structured tables (such as CSV or JSON formats). Through denoising, normalization and format conversion, a standardized multimodal data set is obtained, providing high-quality data input for subsequent feature extraction and analysis.

[0064] Preferably, feature extraction is performed on various types of data in the standardized dataset to obtain spatial features corresponding to neuroimaging data, time series features corresponding to neuroelectrophysiological data, and key features corresponding to biochemical indicators and clinical data, including:

[0065] For neuroimaging data, a convolutional neural network is used to extract spatial features;

[0066] For neuroelectrophysiological data, time-frequency analysis is used to extract time series features;

[0067] For biochemical indicators and clinical data, a random forest algorithm is used to extract key features.

[0068] Preferably, for neuroelectrophysiological data, a time-frequency analysis method is used to extract time series features, including:

[0069] The short-time Fourier transform method is used to extract time-frequency features; the expression of the time-frequency features is: where x(τ) is the original neuroelectrophysiological data, w(τ - t, σ) is the time window function, t is the time point, representing the local spectrum of the signal at different time periods, f is the frequency, representing the frequency component of the signal, X(t, f) is the time-frequency feature, representing the frequency distribution of the signal at each time point, and σ is the width of the window function, which is dynamically adjusted as: α is an adjustment parameter used to control the sensitivity of the window width to the signal amplitude; exemplary, the adjustment parameter is used to dynamically control the width of the time window to balance the time resolution and frequency resolution, so as to more accurately capture the frequency distribution characteristics of the signal at different time points. The specific adjustment method in this embodiment is to dynamically change the window width according to the frequency component of the signal: for high-frequency signals, a narrower time window is used to obtain higher time resolution, so as to more precisely capture the rapidly changing signal characteristics; for low-frequency signals, a wider time window is used to improve the frequency resolution, so as to more clearly describe the distribution of frequency components. This dynamic adjustment usually involves setting an adjustment parameter, and adaptively optimizing the width of the time window according to the range of signal frequencies and the importance of the analysis task, ensuring a relatively balanced time-frequency feature expression in different frequency bands.

[0070] The power spectral density is extracted according to the time-frequency features, and the power spectral density is normalized to obtain the normalized power spectral density; the expression of the power spectral density is: P(t, f) = |X(t, f)| 2 ; where P(t, f) is the power spectral density, representing the energy intensity of the signal at time t and frequency f; the expression of the normalized power spectral density is: where f min and f max are respectively the minimum and maximum values of the analysis frequency range, and P norm(t, f) is the normalized power spectral density;

[0071] Nonlinear feature analysis is performed according to the short-time multi-scale entropy to obtain the frequency-weighted multi-scale entropy; wherein, the expression of the short-time multi-scale entropy is wherein, E(t, τ) is the short-time multi-scale entropy at time t and time scale τ, which is used to quantify the complexity of the signal and reflect the dynamic change pattern of the neuroelectrophysiological signal. τ is the time scale, representing the size of the window, and p i (τ) is the probability distribution of the signal at time scale τ; the expression of the frequency-weighted multi-scale entropy is: E f (t, τ) = ∑ f w(f)E(t, τ, f); wherein, E(t, τ, f) is the short-time multi-scale entropy at frequency f, w(f) is the frequency weight function, which is adaptively adjusted according to the importance of specific frequency bands, and E f (t, τ) is the frequency-weighted multi-scale entropy; Exemplarily, the frequency weight function assigns weights to different frequency components to reflect the importance of specific frequency bands in signal analysis, and then weights and adjusts the multi-scale entropy. The specific implementation method is to customize or learn the weight distribution according to the requirements of the target task and the characteristics of the signal. For example, in the analysis of neuroelectrophysiological signals, certain frequency bands (such as alpha waves, beta waves, gamma waves, etc.) may be closely related to the patient's consciousness state or pathological characteristics. Therefore, higher weights can be assigned to these frequency bands, while lower weights can be assigned to other irrelevant or noisy frequency bands. This weight assignment can be preset by expert experience or adaptively adjusted through a data-driven method (such as optimizing the weight parameters using training data) to ensure that the frequency-weighted multi-scale entropy can more effectively capture the key dynamic features and complexity of the signal.

[0072] Convert the normalized power spectral density into a time-frequency weighted feature; the expression of the time-frequency weighted feature is: wherein, f β is the frequency weighting term. When β > 0, the high-frequency components are emphasized. When β < 0, the low-frequency components are emphasized. F(t) is the time-frequency weighted feature, which represents the projection of the time-frequency feature in the time dimension; σ'(t) is the signal local standard deviation of the neuroelectrophysiological data, and μ(t) is the signal local mean of the neuroelectrophysiological data;

[0073] Perform multi-dimensional feature splicing on the time-frequency feature, the normalized power spectral density, the frequency-weighted multi-scale entropy, and the time-frequency weighted feature to obtain the fused time series feature.

[0074] Preferably, the key features include: blood glucose, lactate dehydrogenase, C-reactive protein, medical history, behavioral assessment scale, and medication record.

[0075] Specifically, step 200 of this embodiment includes:

[0076] Step 201: Extraction of spatial features of neuroimaging data

[0077] For neuroimaging data (such as MRI, fMRI, DTI, etc.), first perform preprocessing, including format conversion (such as NIfTI or DICOM format), normalization, and denoising. Then, use a convolutional neural network (CNN) to extract spatial features, capturing local texture, shape, and structural features in the image data through multi-layer convolution and pooling operations. Finally, compress the high-dimensional spatial features into low-dimensional feature vectors through fully connected layers or global average pooling layers as the spatial features of the neuroimaging data. These features can reflect the spatial differences in the brain structure and function of the patient.

[0078] Step 202: Extraction of time series features of neuroelectrophysiological data

[0079] For neuroelectrophysiological data (such as EEG or ERP), first perform preprocessing, including filtering and denoising (such as removing power frequency interference and low-frequency drift) and normalization. Then, use the short-time Fourier transform (STFT) to perform time-frequency analysis on the signal, extracting the frequency distribution features of the signal at different time points. By dynamically adjusting the time window width, balance the time resolution and frequency resolution, and capture the time-frequency characteristics of the signal, providing a basis for subsequent analysis.

[0080] Step 203: Extraction of normalized power spectral density

[0081] According to the time-frequency analysis results, calculate the energy distribution (power spectral density) of the signal in time and frequency, and perform normalization on it to make it reflect the relative energy distribution of the signal in the frequency range. The normalized power spectral density can more intuitively represent the energy intensity of the signal at different frequencies, providing a standardized input for subsequent feature fusion.

[0082] Step 204: Nonlinear feature analysis and frequency-weighted multi-scale entropy extraction

[0083] Use short-time multi-scale entropy to analyze the complexity of the signal and quantify the dynamic change patterns of neuroelectrophysiological signals. Combine with the frequency weight function to further extract frequency-weighted multi-scale entropy, capturing the importance and complexity of the signal in different frequency bands. These nonlinear features can reflect the dynamic change laws and frequency characteristics of neuroelectrophysiological signals, providing support for the comprehensive description of time series features.

[0084] Step 205: Calculation of time-frequency weighted features

[0085] The normalized power spectrum density is combined with frequency weights to calculate time-frequency weighted features, which are used to adjust the weights of high-frequency or low-frequency components. In addition, by combining the local mean and standard deviation of the signal, the ability to describe time-frequency features is further enhanced. The time-frequency weighted features can reflect the frequency projection of the signal in the time dimension, providing important information for the fusion of time series features.

[0086] Step 206: Multi-dimensional feature concatenation and time series feature fusion

[0087] The time-frequency features, normalized power spectrum density, frequency weighted multi-scale entropy, and time-frequency weighted features are concatenated multi-dimensionally to form the fused time series features. Through the concatenation operation, the unique information of each feature is retained, while the overall ability to describe neuroelectrophysiological signals is enhanced. These fused features can capture the dynamic change patterns, frequency distributions, and non-linear complexities of the signals, providing high-quality time series feature inputs for subsequent prognosis prediction.

[0088] Step 207: Extraction of key features of biochemical indicators and clinical data

[0089] The biochemical indicators (such as blood glucose, lactate dehydrogenase, C-reactive protein, etc.) and clinical data (such as medical history, behavioral assessment scales, and medication records) are preprocessed, including filling missing values, standardization, etc. Then, the random forest algorithm is used to extract key features. By evaluating the importance of the features, the features that contribute the most to the target task are selected, such as blood glucose level, inflammatory indicator (C-reactive protein), patient medical history, and behavioral scores. These key features are combined with neuroimaging and neuroelectrophysiological features, providing important references for multi-modal data fusion.

[0090] Furthermore, in step 300 of this embodiment, various types of features are preprocessed to ensure that their dimensions, ranges, and scales are consistent (such as through normalization or standardization). Then, the spatial features (such as the features extracted from neuroimaging), time series features (such as the time-frequency features of neuroelectrophysiological signals), and key features (such as the features extracted from biochemical indicators and clinical data) are concatenated in a certain order to form a high-dimensional feature vector. If the spatial features are more important, they can be weighted and then fused by setting weights. Finally, the fused feature vector integrates the key information of multi-modal data, providing a unified input for subsequent modeling.

[0091] Exemplarily, in this embodiment, a random forest model is utilized to assign an importance score to each feature. This score represents the average contribution of the feature to the prediction of the target variable across all decision trees. The higher the score, the more important the feature is to the prediction result of the model. The importance scores of the features can be directly extracted from the attributes of the model. According to the extracted feature importance scores, all features are sorted in descending order to obtain the importance ranking of the features. The features ranked at the top are those that contribute the most to the prediction of the target variable, while the features ranked at the bottom may be redundant or irrelevant. To verify the rationality of the feature ranking, the model can be retrained by gradually removing the features with low importance and observing the changes in the model performance. If the model performance does not significantly decline after removing certain features, it indicates that these features contribute less to the task and the feature set can be further optimized. In this way, this embodiment can not only determine the importance of the features but also provide a basis for subsequent feature selection and model optimization, thereby improving the efficiency and prediction performance of the model.

[0092] Preferably, the fused feature vector is input into the trained consciousness state grading evaluation model to obtain a classification result and intermediate layer features of the model, including:

[0093] Construct an initial deep model based on a multi-layer perceptron;

[0094] Input the preset sample fused feature vector and the sample consciousness state level into the initial deep model to obtain the trained consciousness state grading evaluation model; the loss function of the consciousness state grading evaluation model is the cross-entropy loss function; the optimizer of the cross-entropy loss function is the Adam optimizer; the network structure of the consciousness state grading evaluation model includes: an input layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence;

[0095] Input the fused feature vector into the trained consciousness state grading evaluation model to obtain a classification result and intermediate layer features of the model; the intermediate layer features of the model are the output of the second fully connected layer to represent the deep features of the fused feature vector.

[0096] Specifically, step 400 of this embodiment includes:

[0097] Step 401: Construct an initial deep model based on a multi-layer perceptron

[0098] First, design a multi-layer perceptron (MLP) as the initial deep model for the grading evaluation of the consciousness state. The network structure of this model includes an input layer, two fully connected layers (the first fully connected layer and the second fully connected layer), and an output layer. The dimension of the input layer is consistent with the size of the fused feature vector, and it is used to receive the feature vector after multi-modal data fusion. The first fully connected layer and the second fully connected layer are respectively used to extract high-order features, and the output layer is used to generate classification results. An activation function (such as ReLU) can be added after each layer to enhance the non-linear expression ability of the model, and regularization (such as Dropout) can be added to prevent overfitting.

[0099] Step 402: Train the grading evaluation model of the consciousness state

[0100] Input the preset sample fused feature vector and the corresponding sample consciousness state level (label) into the initial deep model for training. During the training process, use the cross-entropy loss function as the objective function to measure the difference between the model prediction result and the true label. To optimize the model parameters, select the Adam optimizer, which can dynamically adjust the learning rate, has a fast convergence speed and good robustness. Through multiple rounds of iterative training, the model gradually learns the mapping relationship between the fused feature vector and the consciousness state level, and finally obtains the trained grading evaluation model of the consciousness state.

[0101] Step 403: Input the fused feature vector for inference

[0102] Input the new fused feature vector into the trained grading evaluation model of the consciousness state, and the model generates a classification result through forward propagation. The classification result represents the consciousness state level of the patient (such as fully conscious, partially conscious or unconscious, etc.), providing a basis for clinicians to evaluate the patient's current consciousness state. At the same time, during the inference process, the output of the middle layer (i.e., the second fully connected layer) of the model will also be extracted as the deep feature of the fused feature vector.

[0103] Step 404: Extract the features of the middle layer of the model

[0104] The features of the middle layer of the model are the output of the second fully connected layer, representing the high-order features extracted from the fused feature vector after passing through the deep model. These features can capture the deep associations and complex patterns of multi-modal data and have higher expression ability. The features of the middle layer can be used as the input for subsequent analysis, such as for the training of a patient prognosis prediction model or further feature analysis. This method can not only provide the classification result of the consciousness state, but also provide rich feature information for subsequent tasks through the features of the middle layer.

[0105] Further, in step 500 of this embodiment, the classification result of the trained consciousness state grading evaluation model is used as a label to represent the consciousness state level of each sample. Then, the intermediate layer features of the model (i.e., the output of the second fully connected layer) are extracted. These features are high-order features after the fusion feature vector is extracted by the deep model and can reflect the deep association and complex patterns of multimodal data. The intermediate layer features of each sample are combined with the corresponding classification result (label) to form a new dataset, that is, the feature-level dataset. The samples of this dataset consist of intermediate layer feature vectors, and the labels are the classification results of consciousness state grading.

[0106] The feature-level dataset mainly includes two parts: one is the deep feature vector extracted from the intermediate layer of the model as the input feature of the dataset; the other is the classification result of the consciousness state grading model as the label of the dataset. This dataset can be used for the subsequent training of the patient prognosis prediction model because the intermediate layer feature vector contains high-order information of multimodal data, and the classification result as a label can guide the prognosis prediction model to learn the relationship between the patient's consciousness state and prognosis. This construction method of the feature-level dataset can effectively improve the performance and generalization ability of the prognosis prediction model.

[0107] Preferably, inputting the feature-level dataset into the trained patient prognosis prediction model to obtain a prediction result, including:

[0108] Obtain a preset feature-level data sample set;

[0109] Construct an initial convolutional neural network;

[0110] Train the initial convolutional neural network according to the feature-level data sample set to obtain a trained classifier;

[0111] Connect a trained LSTM neural network after the classifier to obtain the patient prognosis prediction model;

[0112] Input the feature-level dataset into the patient prognosis prediction model to obtain a prediction result.

[0113] Preferably, the prediction result includes: the recovery potential of the patient's consciousness state, the prediction of the recovery time, the possibility of disease deterioration, and the specific prognosis grade.

[0114] Preferably, the prognosis grade includes: complete recovery, partial recovery, and no recovery.

[0115] Specifically, step 600 of this embodiment includes:

[0116] Step 601: Obtain a feature-level data sample set

[0117] First, collect and organize the feature-level data sample set, which consists of the deep feature vectors extracted from the middle layer of the model and the classification result labels. These feature vectors contain high-order information of multi-modal fusion features and can reflect the patient's consciousness state and related complex patterns. Through preprocessing (such as normalization, standardization, etc.), ensure the quality and consistency of the feature-level data sample set. Then, divide the data set into a training set, a validation set, and a test set for subsequent training and evaluation of the patient prognosis prediction model.

[0118] Step 602: Construct an initial convolutional neural network and train the classifier

[0119] Construct an initial convolutional neural network (CNN) to further extract local patterns and high-order features from the feature-level data sample set. The input layer of the CNN is consistent with the dimension of the feature vector. The convolutional layer extracts the local correlations of the features through filters, and the pooling layer is used for dimensionality reduction and enhancing the robustness of the model. Input the feature-level data sample set into the CNN for training, use cross-entropy as the loss function, and optimize the model parameters through the Adam optimizer. After training, obtain a trained classifier that can preliminarily classify the feature-level data and extract the deep features related to the patient's prognosis.

[0120] Step 603: Connect the LSTM neural network to construct the prognosis prediction model

[0121] Connect an LSTM neural network after the trained CNN classifier to capture the temporal dynamic patterns and sequential dependencies of the feature-level data. The LSTM can handle the long-term dependency problems in sequential data and is very suitable for analyzing the temporal evolution of features during the patient's prognosis process. Use the high-order features extracted by the CNN as the input of the LSTM, and through training the LSTM network, learn the dynamic change rules of the patient's consciousness state and the time series features related to the prognosis. Finally, the combination of the CNN and the LSTM constitutes a complete patient prognosis prediction model.

[0122] Step 604: Input the feature-level data set and generate the prediction results

[0123] Input the feature-level data set into the trained patient prognosis prediction model, and the model generates the prediction results through forward propagation. The prediction results include the recovery potential of the patient's consciousness state (such as the probability of complete recovery or partial recovery), the prediction of the recovery time (such as the time range required for expected recovery), the possibility of disease deterioration (such as the deterioration probability or risk level), and the specific prognosis grade (such as the grading score or classification result). These prediction results can provide comprehensive prognosis information for clinicians to help formulate personalized treatment plans and intervention measures, and at the same time provide a scientific decision-making basis for the patient's family members.

[0124] Corresponding to the above method, this embodiment also provides a prognosis prediction system for chronic Doc patients, including:

[0125] A data normalization module, configured to collect multimodal data of chronic Doc patients, and perform normalization processing on each type of the multimodal data respectively to obtain a normalized data set; the multimodal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators, and clinical information;

[0126] A feature extraction module, configured to extract features from each type of data in the normalized data set to obtain spatial features corresponding to the neuroimaging data, time series features corresponding to the neuroelectrophysiological data, and key features corresponding to the biochemical indicators and clinical information;

[0127] A feature fusion module, configured to perform feature fusion on the spatial features, the time series features, and the key features to obtain a fused feature vector;

[0128] A classification module, configured to input the fused feature vector into a trained consciousness state grading evaluation model to obtain a classification result and intermediate layer features of the model;

[0129] A feature-level generation module, configured to use the classification result as a label and generate a feature-level data set in combination with the intermediate layer features of the model;

[0130] A prediction module, configured to input the feature-level data set into a trained patient prognosis prediction model to obtain a prediction result.

[0131] The beneficial effects of the present invention are as follows:

[0132] By combining a trained consciousness state grading evaluation model and a prognosis prediction model, the present invention can accurately predict the consciousness state grading and prognosis development trend of patients, provide scientific decision-making support for clinicians, optimize personalized treatment plans, improve the accuracy and reliability of prognosis prediction, and thus enhance the rehabilitation effect of patients and the utilization efficiency of medical resources.

[0133] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0134] In this article, specific examples are used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for helping to 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 and application scope. To sum up, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for predicting the prognosis of patients with chronic Doc, characterized in that: include: Collecting multimodal data of chronic Doc patients, and standardizing each type of the multimodal data to obtain a standardized data set; The multimodal data include: neuroimaging data, neuroelectrophysiological data, biochemical indicators and clinical data; Extracting features of various types of data in the standardized data set to obtain spatial features corresponding to neuroimaging data, time series features corresponding to neuroelectrophysiological data, and key features corresponding to biochemical indicators and clinical data; Performing feature fusion on the spatial feature, the time series feature and the key feature to obtain a fused feature vector; Inputting the fused feature vector into a trained consciousness state classification assessment model to obtain classification results and model intermediate layer features; The classification results are used as labels and combined with the model's intermediate layer features to generate a feature-level data set; The feature-level data set is input into a trained patient prognosis prediction model to obtain a prediction result.

2. The method for predicting prognosis of chronic Doc patients according to claim 1, characterized in that: The standardization process includes: denoising, normalization and format conversion.

3. The method for predicting the prognosis of chronic Doc patients according to claim 1, characterized in that: Feature extraction is performed on each type of data in the standardized data set to obtain spatial features corresponding to neuroimaging data, time series features corresponding to neuroelectrophysiological data, and key features corresponding to biochemical indicators and clinical data, including: For neuroimaging data, convolutional neural networks are used to extract spatial features; For neuroelectrophysiological data, time-frequency analysis is used to extract time series features; For biochemical indicators and clinical data, random forest algorithm was used to extract key features.

4. The method for predicting prognosis of chronic Doc patients according to claim 3, characterized in that: For neurophysiological data, time-frequency analysis methods are used to extract time series features, including: The time-frequency features are extracted by short-time Fourier transform; the expression of the time-frequency features is: Among them, x(τ) is the original neurophysiological data, w(τ-t,σ) is the time window function, t is the time point, which represents the local spectrum of the signal in different time periods, f is the frequency, which represents the frequency component of the signal, X(t,f) is the time-frequency feature, which represents the frequency distribution of the signal at each time point, and σ is the width of the window function, which is dynamically adjusted as follows: α is an adjustment parameter used to control the sensitivity of the window width to the signal amplitude; The power spectrum density is extracted according to the time-frequency feature, and the power spectrum density is normalized to obtain the normalized power spectrum density; the expression of the power spectrum density is: P(t,f)=|X(t,f)| 2 ; Wherein, P(t,f) is the power spectral density, which represents the energy intensity of the signal at time t and frequency f; the expression of the normalized power spectral density is: Among them, f min and f max are the minimum and maximum values ​​of the analysis frequency range, P norm (t,f) is the normalized power spectral density; According to the short-time multi-scale entropy, nonlinear characteristic analysis is performed to obtain the frequency-weighted multi-scale entropy; wherein the expression of the short-time multi-scale entropy is: Among them, E(t,τ) is the short-term multiscale entropy at time t and time scale τ, which is used to quantify the complexity of the signal and reflect the dynamic change pattern of the neural electrophysiological signal. τ is the time scale, indicating the size of the window, and p i (τ) is the probability distribution of the signal on the time scale τ; the expression of the frequency-weighted multiscale entropy is: E f (t,τ)=∑ f w(f)E(t,τ,f); where E(t,τ,f) is the short-term multiscale entropy at frequency f, w(f) is the frequency weight function, which is adaptively adjusted according to the importance of a specific frequency band, and E f (t,τ) is the frequency-weighted multi-scale entropy; The normalized power spectrum density is converted into a time-frequency weighted feature; the expression of the time-frequency weighted feature is: Among them, f β is the frequency weighting term. When β>0, the high-frequency component is emphasized. When β<0, the low-frequency component is emphasized. -F(t) is the time-frequency weighted feature, which represents the projection of the time-frequency feature on the time dimension. σ'(t) is the local standard deviation of the signal of the neuroelectrophysiological data, and μ(t) is the local mean of the signal of the neuroelectrophysiological data; The time-frequency features, the normalized power spectral density, the frequency-weighted multi-scale entropy and the time-frequency weighted features are multi-dimensionally concatenated to obtain fused time series features.

5. The method for predicting prognosis of chronic Doc patients according to claim 3, characterized in that: The key features include: blood glucose, lactate dehydrogenase, C-reactive protein, medical history, behavioral assessment scales and medication records.

6. The method for predicting the prognosis of chronic Doc patients according to claim 3, characterized in that: The fused feature vector is input into the trained consciousness state classification assessment model to obtain the classification result and the model intermediate layer features, including: Build an initial deep model based on a multi-layer perceptron; Input the preset sample fusion feature vector and the sample consciousness state level into the initial deep model to obtain a trained consciousness state grading evaluation model; the loss function of the consciousness state grading evaluation model is a cross entropy loss function; the optimizer of the cross entropy loss function is an Adam optimizer; the network structure of the consciousness state grading evaluation model includes: an input layer, a first fully connected layer, a second fully connected layer and an output layer connected in sequence; The fused feature vector is input into a trained consciousness state grading assessment model to obtain a classification result and a model intermediate layer feature; the model intermediate layer feature is the output of the second fully connected layer to represent the deep features of the fused feature vector.

7. The method for predicting the prognosis of chronic Doc patients according to claim 3, characterized in that: The feature-level data set is input into the trained patient prognosis prediction model to obtain prediction results, including: Obtain a preset feature-level data sample set; Construct an initial convolutional neural network; Training the initial convolutional neural network according to the feature-level data sample set to obtain a trained classifier; Connecting the trained LSTM neural network after the classifier to obtain the patient prognosis prediction model; The feature-level data set is input into the patient prognosis prediction model to obtain a prediction result.

8. The method for predicting prognosis of chronic Doc patients according to claim 1, characterized in that: The prediction results include: the potential for recovery of the patient's consciousness state, the prediction of recovery time, the possibility of disease progression, and the specific prognostic level.

9. The method for predicting prognosis of chronic Doc patients according to claim 8, characterized in that: The prognostic levels include: complete recovery, partial recovery, and no recovery.

10. A prognosis prediction system for chronic Doc patients, characterized in that: include: A data standardization module is used to collect multimodal data of chronic Doc patients and standardize each type of the multimodal data to obtain a standardized data set; the multimodal data includes: neuroimaging data, neuroelectrophysiological data, biochemical indicators and clinical data; A feature extraction module is used to extract features from various types of data in the standardized data set to obtain spatial features corresponding to neuroimaging data, time series features corresponding to neuroelectrophysiological data, and key features corresponding to biochemical indicators and clinical data; A feature fusion module, used for fusing the spatial feature, the time series feature and the key feature to obtain a fused feature vector; A classification module, used for inputting the fused feature vector into a trained consciousness state classification assessment model to obtain a classification result and a model intermediate layer feature; A feature-level generation module, used to use the classification results as labels and combine them with the model's intermediate layer features to generate a feature-level data set; The prediction module is used to input the feature-level data set into a trained patient prognosis prediction model to obtain a prediction result.

Images