Newborn brain injury grading system based on elastic net and crow optimization fusion algorithm

Abstract

The present application belongs to the technical field of grading detection of neonatal brain injury, and particularly relates to a neonatal brain injury grading system based on elastic net and crow optimization fusion algorithm; wherein a data preprocessing module receives text data of original neonatal electroencephalogram signals, and processes the data into independent samples; a feature extraction module uses quantitative electroencephalogram analysis to obtain quantitative information from the independent sample data, extracts qEEG related features and mne related features of amplitude, spectral density, connectivity and burst interval for each independent sample, and performs standardization processing; a feature preliminary screening module uses a feature selection algorithm based on elastic net regression to preliminarily select features related to neonatal brain injury grading results from the independent sample data; an optimal feature subset establishment module uses a crow optimization algorithm to further select the features selected by the feature preliminary screening module, establishes an optimal feature subset strongly related to neonatal brain injury results; and a model grading decision module uses a support vector machine (SVM) machine learning model to predict neonatal brain injury grading for the optimal features established by the optimal feature subset establishment module, and outputs the prediction results; the present application can improve the accuracy of neonatal brain injury grading diagnosis, screen out features strongly related to the results, find characteristic markers of neonatal brain injury, and provide effective auxiliary support for clinical medical diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of neonatal brain injury severity grading detection technology, specifically involving a neonatal brain injury grading system based on an elastic network and crow optimization fusion algorithm. Background Technology

[0002] Neonatal encephalopathy (NE) is a complex neurological disorder that causes brain damage in newborns. In high-income countries, the incidence of NE is approximately 0.5–3 per 1,000 live births, while in low- and middle-income countries it affects millions of premature and full-term infants. Neonatal brain damage carries the risk of neurological sequelae such as seizures, cerebral palsy, intellectual disability, and even death. Developing neonatal brain damage manifests as changes in baseline electroencephalogram (EEG). Continuous resting-state EEG monitoring can characterize and quantify brain activity and is widely used to help physicians diagnose the severity of neonatal brain damage.

[0003] Therapeutic hypothermia and other novel treatment strategies necessitate initiation before secondary brain injury begins. Therefore, accurate assessment of the extent of brain injury in newborns is crucial to ensure prompt initiation of therapeutic intervention during functional monitoring. While continuous EEG monitoring for several hours is convenient in the neonatal intensive care unit (NICU), multichannel neonatal EEGs contain complex spatiotemporal information that is difficult for non-EEG specialists to interpret.

[0004] In recent years, machine learning has developed rapidly. It can automatically improve its own performance and continuously extract important information from experimental data efficiently. It has been widely used in the diagnosis and research of neonatal brain injury and other neurological diseases, such as epilepsy seizure prediction, mood transformation analysis, and brain-computer interface (BCI).

[0005] For the grading and prediction of neonatal brain injury severity, most existing studies are trained on relatively small electroencephalogram (EEG) datasets due to institutional procedures and data protection laws, making it extremely difficult to share medical data between institutions in practice. Secondly, most researchers use morphological pattern features and clinical medical features, which cannot guarantee the identification of features highly correlated with the prediction results. Furthermore, in feature selection, most researchers use statistical methods to derive the association between relevant features and prediction results, which may overlook potential correlations between features, making rapid and accurate diagnosis of neonatal brain injury grading difficult.

[0006] Fusion feature selection is a feature selection algorithm suitable for high-dimensional and feature-related datasets. Its core idea is to fully leverage the advantages of different feature selection algorithms. By combining two different feature selection methods, their strengths complement each other, achieving the selection of the optimal feature subset. The fused feature selection algorithm significantly improves both subset evaluation and classification accuracy. Currently, this method is mainly used in industrial applications and has not yet been widely applied in the field of neonatal brain injury grading and diagnosis. Summary of the Invention

[0007] To overcome the aforementioned problems, this invention provides a neonatal brain injury grading system based on a fusion algorithm of Elastic Regression (ENRA) and Crow Optimization (CSA). This system utilizes an Elastic Regression Feature Selection Algorithm (ENRA), a Crow Optimization Algorithm (CSA), and a Support Vector Machine (SVM) machine learning model. It includes a data preprocessing module, a feature extraction module, a preliminary feature screening module, an optimal feature subset establishment module, and a model grading decision module. The system processes, performs machine learning on, and classifies the raw EEG dataset of newborns, thereby improving the accuracy of neonatal brain injury grading diagnosis. Simultaneously, it filters out features strongly correlated with the results, identifying characteristic biomarkers of neonatal brain injury and providing effective auxiliary support for clinical medical diagnosis.

[0008] A neonatal brain injury grading system based on the fusion algorithm of elastic net and crow optimization is characterized by including a data preprocessing module, a feature extraction module, a preliminary feature screening module, an elastic net regression module, an optimal feature subset establishment module, and a model grading decision module. The data preprocessing module is used to receive text data of the neonatal raw EEG signals and perform artifact removal, missing value filling, and data slicing into independent samples.

[0009] The feature extraction module uses quantitative electroencephalography (QEEG) analysis to obtain quantitative information from the independent sample data after data preprocessing. For each independent sample, it extracts qEEG-related features and MNE-related features of amplitude, spectral density, connectivity correlation, and pulse train interval, and performs standardization processing. The specific content is as follows:

[0010] Step 1: The quantitative EEG analysis method uses the NEURAL_py_EEG package in Python to decompose each independent sample output from the data preprocessing module into four frequency bands: δ (0.5-4Hz), θ (4-7Hz), α (7-13Hz), and β (13-30Hz). Features are calculated within a short duration window, and this window is shifted over time. The window length is 64 seconds with 50% overlap. Features are estimated for each channel of each short duration window sample. For the feature values ​​of multiple short duration window samples calculated in each channel under each frequency band, the median is used to summarize the feature values ​​of each channel. For the feature values ​​summarized for all channels under each frequency band, the median of all channels is taken as the feature value of the same feature in each frequency band, and is recorded as the multiple first features obtained.

[0011] Then, using the mne package in Python, the 5-minute independent sample data processed by the data preprocessing module is imported. Using the mne.features module, the selected_funcs parameter is set to the specific features selected. The mne package will automatically process each independent sample and extract a series of time domain, frequency domain and complexity features according to the specified function, which are denoted as the multiple second features obtained.

[0012] Step 2: Each independent sample contains multi-dimensional feature samples; the data standardization process uses the Min-Max standardization method to standardize each feature sample. The formula for the Min-Max standardization method is as follows:

[0013]

[0014] Where: x′ represents the standardized one-dimensional feature value, x represents the original feature value in the independent sample, min(x) is the minimum value of the feature value in all independent samples, and max(x) is the maximum value of the feature value in all independent samples;

[0015] The preliminary feature screening module uses a feature selection algorithm based on elastic network regression to initially select features related to the neonatal brain injury grading results from the independent sample data after the feature extraction module. The specific content is as follows:

[0016] Step 1: Reconstruct the n-dimensional EEG feature data from the independent samples extracted by the feature extraction module into a feature data matrix X (feature vector x = (x1, x2, ..., x...)). n This adapts it to the processing of feature selection algorithms.

[0017]

[0018] Where: x m,nThe value representing the nth feature on the mth independent sample is called the feature value.

[0019] Step two: The feature data matrix X is pre-screened using the elastic network regression module to obtain the largest feature sample combination after screening by the elastic network regression module, which is then reconstructed into a new feature data matrix Y and feature vector y. s The selected feature is represented by 1 ≤ s ≤ n;

[0020]

[0021] Where: y m,s This represents the value of the s-th feature on the m-th independent sample after filtering.

[0022] The Elastic Net Regression module uses the Elastic Net Regression loss function to filter features, as follows:

[0023] Step 1: Input a certain number of selected neonatal EEG signal samples into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data processed by the feature extraction module into a feature data matrix X. Use the feature data matrix X of all independent samples as the training set.

[0024] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set;

[0025] Step 3: Input the labeled training set into the Elastic Regression module for training. The Elastic Regression loss function of the Elastic Regression module is calculated using the following formula:

[0026]

[0027] Where: M is the number of all independent samples, y i It is the actual label value for each individual sample. λ1 is the predicted label value for each independent sample, i.e., normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, respectively; λ1 is the regularization parameter, α is the mixture parameter, p is the number of features, and β is the number of features. j It is the regression coefficient of the j-th feature;

[0028] Step four: Repeat step three a specified number of times to obtain the trained elastic network regression module; the elastic network regression module sets feature importance coefficients for the multidimensional features in the feature data matrix X, and through continuous training, the regression coefficients β of some unimportant features will become 0. By continuously training, these features with coefficients of 0 can be eliminated; when new unlabeled sample data is input, the feature combinations in these sample data that are the same as the feature combinations retained by the elastic network regression module on the training set are directly retained as the new feature data matrix Y;

[0029] The optimal feature subset determination module uses the Raven optimization algorithm to further select features from the initial feature screening module, establishing the optimal feature subset that is strongly correlated with neonatal brain injury outcomes. The specific details are as follows:

[0030] Within the framework of a population-based crow optimization algorithm, consider a population of 10 crows, where each crow represents a new combination of features;

[0031] The crow optimization algorithm uses a specific fitness function to determine the quality of each crow's search result. By traversing and searching all feature samples in the maximum feature sample combination Y, it finally determines the subset of feature samples that are identical to the feature sample set left during training as the strongly correlated feature combination for neonatal brain injury grading. The result is then reconstructed into a new feature data matrix Z and feature vector z. r The selected feature is 1 ≤ r ≤ s;

[0032]

[0033] Where: z m,r This represents the value of the r-th feature on the m-th independent sample after being filtered by the optimal feature subset determination module;

[0034] The training process of the optimal feature subset determination module is as follows:

[0035] Step 1: Select a certain number of newborn EEG signal samples and input them into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data after processing by the feature extraction module into a feature data matrix X. Use the feature data matrix Y of all independent samples after preliminary feature screening as the training set.

[0036] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set;

[0037] Step 3: Calculate the fitness function `fitness()` using the Raven optimization algorithm on the labeled training set as follows:

[0038]

[0039]

[0040] Where: Acc is the classification accuracy calculated using the nearest neighbor classifier, A is a weighting factor with values ​​in [0, 1], and L... t It is the number of features that each crow selects from the combination of the largest feature samples Y, 1≤t≤s, L s It is the number of features in the maximum feature sample combination Y output by the feature preliminary screening module;

[0041] Step four, repeat step three a specified number of times to obtain the trained best feature subset establishment module. During the training process, a subset of feature combinations will be selected from the feature data matrix Y of all independent samples. These feature combinations are fixed z. New input samples can directly retain the same feature combination Z as this combination.

[0042] The new feature data matrix Z is input into the model grading and decision module for brain injury label prediction. The model grading and decision module uses the support vector machine (SVM) machine learning model to classify each independent sample, and then votes on the classification results of all independent samples. The category with the most votes is designated as the final classification result of the newborn's EEG signal sample.

[0043] The model grading decision module uses a support vector machine (SVM) machine learning model to predict the grading of neonatal brain injury based on the best features established by the best feature subset establishment module, and outputs the prediction results:

[0044] For an input sample feature vector z, its class is determined by the sign of the decision function: that is, if w i ·z+b i If the value is greater than 0, then the independent sample belongs to category i; otherwise, it belongs to other categories.

[0045] The class with the largest decision function value is selected as the predicted class: For a new input independent sample, four decision functions f are calculated by four binary classifiers respectively. i The (z) value, where the category corresponding to the largest decision function value will be output as the prediction result for the independent sample of the new input.

[0046] The model classification and decision module adopts the Support Vector Machine (SVM) machine learning model, and its structure is as follows:

[0047] f i (z)=w i ·z+b i

[0048]

[0049] Where: w i It is a weight vector associated with class i, C is a regularization parameter used to balance the weights that minimize the loss and maximize the margin, y j In a neonatal brain injury classification task, the corresponding label values ​​for a specific category j are: 0, 1, 2, 3, f i (z) is the decision function for each category i, which calculates the score for predicting whether a sample belongs to category i based on the input feature vector z. The feature vector z = (z1, z2, ..., z...) r ) is the input feature vector, b i It is the bias term of the corresponding classifier;

[0050] The Support Vector Machine (SVM) machine learning model performs classification using multiple classification planes. Each classification plane employs a one-to-one strategy, treating each class as a separate category and distinguishing it from all other classes. All other classes are then merged into a single class. The constraints of the classification planes are as follows:

[0051] y i (w i ·z+b i )≥1

[0052] Where: y i It involves assigning a value of -1 or 1 to a category i, and creating a binary classifier y for each category i. i When the value is -1, it represents the combination of other categories excluding category i, y i When the value is 1, it represents category i; the feature vector z = (z1, z2, ..., z... r ) is the input feature vector.

[0053] The specific process of training the Support Vector Machine (SVM) machine learning model is as follows:

[0054] Step 1: Input a certain number of selected neonatal EEG signal samples into the data preprocessing module, and then sequentially pass through the feature extraction module, the preliminary feature screening module, and the optimal feature subset establishment module. Finally, use all the independent samples processed by the optimal feature subset establishment module as the training set.

[0055] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set;

[0056] Step 3: Collect 2062 independent samples with replacement from the labeled training set obtained in Step 2. This random sampling with replacement generates multiple sample sets, which are then input into the model classification and decision module for training. During training, monitor the performance of the SVM machine learning model on an independent validation set and use cross-validation to avoid overfitting. Set a convergence threshold; once the model's loss function changes less than this threshold or the model's performance on the validation set no longer improves, the iteration stops, resulting in the trained Support Vector Machine (SVM) machine learning model.

[0057] The text data of neonatal EEG signals received by the data preprocessing module refers to 6 hours of continuous EEG signals from full-term newborns. The data is exported from the proprietary format of the NicoletOne machine, which is an open European data file format, and is securely stored for offline analysis. The EEGLAB toolkit in MATLAB is used to convert the acquired raw neonatal EEG data in European data file format into text data of neonatal EEG signals, and each text data of neonatal EEG signals is regarded as a neonatal EEG signal sample.

[0058] Samples with missing values ​​>66% in the neonatal EEG signal samples will be excluded from further analysis. Neonatal EEG signal samples with missing values ​​<66% will be processed by missing value imputation, where missing value imputation will be performed using the mode imputation method. Then, each neonatal EEG signal sample after imputation will be segmented into 5-minute slice windows, and these slice data will be selected as independent samples.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. The neonatal brain injury grading system based on the fusion algorithm of elastic net regression and crow optimization described in this invention employs a novel feature selection method. It integrates an elastic net regression-based feature selection algorithm with a crow optimization algorithm to select the optimal feature subset from neonatal EEG signal samples. This fusion algorithm reduces the computational burden by using the elastic net regression feature selection algorithm and initially filters features irrelevant to the label results. It combines the advantage of the crow optimization algorithm, which considers the correlation between features, thereby improving the search efficiency of the optimal feature subset and the accuracy of neonatal brain injury classification. This fusion feature selection algorithm has been successfully applied to the grading and diagnosis of neonatal brain injury.

[0061] 2. The neonatal brain injury grading system based on the elastic net and crow optimization fusion algorithm described in this invention, after determining the optimal feature subset using a new feature selection method, outputs sparse feature vectors. Combined with the support vector machine (SVM) machine learning model, it is applied to the grading and diagnosis of neonatal brain injury, screening out features that are strongly correlated with the grading results of neonatal brain injury. This improves the current accuracy of predicting and classifying neonatal brain injury, enables rapid grading and diagnosis of the degree of neonatal brain injury, and provides an effective and usable auxiliary tool for the clinical diagnosis of neonatal brain injury. Detailed Implementation

[0062] The neonatal brain injury grading diagnosis method described in this invention applies a fusion feature selection algorithm to predict neonatal brain injury grading. Based on the high-dimensionality and characteristic relationships of neonatal EEG data, it employs a feature selection algorithm based on elastic network regression and a crow optimization algorithm, appropriately fusing the two algorithms to more effectively select the optimal feature subset. Finally, a support vector machine (SVM) machine learning model is used to classify and predict neonatal EEG signal samples, resulting in a more accurate and comprehensive prediction accuracy for the neonatal brain injury grading diagnosis system described in this invention. Simultaneously, it establishes feature variables strongly correlated with neonatal brain injury grading.

[0063] Example 1

[0064] A neonatal brain injury grading system based on the fusion algorithm of elastic net and crow optimization includes a data preprocessing module, a feature extraction module, a preliminary feature screening module, an optimal feature subset establishment module, and a model grading decision module. The data preprocessing module is used to receive text data of neonatal EEG signals and perform artifact removal, missing value imputation, and data slicing on the data to obtain preprocessed independent samples.

[0065] The feature extraction module uses quantitative electroencephalography (qEEG) to obtain quantitative information from the independent sample data processed by the data preprocessing module. For each independent sample, qEEG-related features and MNE-related EEG signal features such as amplitude, spectral density, connectivity correlation, and pulse train interval are extracted and standardized.

[0066] The preliminary feature screening module uses the Elastic Network Regression-based Feature Selection Algorithm (ENRA) to initially select features related to the neonatal brain injury grading results from the independent sample feature data processed by the feature extraction module.

[0067] The purpose of using the Elastic Network Regression Feature Selection (ENRA) algorithm is to help filter out features that are not relevant to the label results. This is because the CSA optimization algorithm used in the next module has a high computational burden and high time complexity. Filtering out features that are not relevant to the label results in advance can improve the search efficiency of the CSA optimization algorithm. This step mainly considers the computational time cost.

[0068] The optimal feature subset establishment module uses the Raven optimization algorithm to further select the features selected by the initial feature screening module, and establishes the optimal feature subset that is strongly correlated with the neonatal brain injury results.

[0069] The model grading decision module uses a support vector machine (SVM) machine learning model to predict the grading of neonatal brain injury based on the best features established by the best feature subset establishment module, and outputs the prediction results.

[0070] The text data of neonatal EEG signals received by the data preprocessing module refers to 6 hours of continuous EEG signals collected from full-term newborns in the hospital using a NicoletOne machine (after attaching 12 electrodes to the newborn and collecting the raw signals, the raw signal data is exported from the proprietary format of the NicoletOne machine to the open European Data File (EDF) format and securely stored for experimental use); and the EEGLAB toolkit in MATLAB is used to convert the acquired raw neonatal EEG data in European Data File (EDF) format into text data of neonatal EEG signals; the text data of each neonatal EEG signal includes 12 time series data, each time series corresponding to one electrode channel, recorded in digital form, and each number represents the potential difference (i.e., the amplitude of brain waves) captured by the electrode at a specific time point.

[0071] Each converted neonatal EEG signal text data is considered a neonatal EEG signal sample for subsequent data processing and feature extraction. Samples with missing values ​​>66% (missing values ​​refer to situations where no valid EEG activity data was recorded at a specific time point or time period, possibly due to various reasons such as poor electrode contact or signal interference) are excluded from further analysis. Neonatal EEG signal samples with missing values ​​<66% are processed using missing value imputation, specifically mode imputation (using the data that appears most frequently in the feature sample of that dimension for imputation; in this embodiment, the mode of the missing data in each column (feature) is calculated, and the missing values ​​are imputed using the data that appears most frequently in that column). Each imputed neonatal EEG signal sample is then segmented into 5-minute slice windows (without aliasing), and each sliced ​​neonatal EEG signal sample is considered an independent sample. The purpose is to ensure that more neonatal data can be included, rather than data from fewer neonatals with longer observation periods.

[0072] The feature extraction module extracts features related to the neonatal brain injury grading results using quantitative electroencephalography (EEG) analysis, and the specific process is as follows:

[0073] Step 1: Quantitative Electroencephalography (qEEG) Analysis. Import the 5-minute independent sample data processed by the data preprocessing module. Use the NEURAL_py_EEG package in Python for band decomposition, setting the bands to four: δ (0.5-4Hz), θ (4-7Hz), α (7-13Hz), and β (13-30Hz). Then, set the time window to 64 seconds per window with a 50% overlap between windows. After setting, perform decomposition and feature value calculation. The software automatically segments the independent samples under each band according to the specified time window and overlap rate, obtaining short-duration window samples for each band. Simultaneously, the software calculates the feature values ​​of each channel (channel refers to the physical channel used during signal acquisition; the experiment used 12 channels, meaning 12 electrode patches were applied to each newborn during data acquisition) within each short-duration window sample, estimating the features for each channel of each short-duration window sample.

[0074] For the feature values ​​of multiple short-duration window samples calculated in each channel of the short-duration window samples in each frequency band, the median is used to summarize the feature values ​​of each channel. That is, the median is taken as the feature value of the corresponding channel. For the feature values ​​summarized from all channels in each frequency band, the median of the feature values ​​summarized from all channels is taken as the feature value of the same feature in each frequency band, which is recorded as the multiple first features obtained. The number of first features extracted is 102, of which there are 30 features in Table 1. Among them, after calculating the feature values ​​of 6 features (spectral_edge_frequency, FD, IBI_length_max, IBI_length_median, IBI_burst_prc, IBI_burst_number) in 4 frequency bands, they will be automatically summarized into a feature value of 0.5-30 Hz.

[0075] Therefore, for the remaining 24 features, each feature will have 4 feature values ​​calculated in 4 frequency bands, and 6 features will have one feature value calculated in 0.5-30 Hz. The final total is 24×4+6×1=102.

[0076] The relevant characteristics of qEEG are shown in Table 1:

[0077] Table 1. Quantitative Electroencephalography (qEEG) Characteristics

[0078]

[0079]

[0080] Then, using the mne package in Python, the 5-minute independent sample data processed by the data preprocessing module is imported. Using the mne.features module, the selected_funcs parameter is set to the specific features selected. The mne package will automatically process each independent sample and extract a series of time-domain, frequency-domain, and complexity features according to the specified function. These are denoted as the multiple second features obtained. The number of second features extracted is 187.

[0081] Table 2 shows the relevant EEG characteristics of mne:

[0082] Table 2. EEG-related characteristics of mne

[0083]

[0084]

[0085] For each independent sample, a total of 289 (187+102) EEG signal features were obtained; after all features were extracted, the feature data existed in the form of a number array, which could be further analyzed and processed directly in Python;

[0086] The main purpose of feature extraction is to find unique patterns from input neonatal EEG signal samples to achieve high classification rates in machine learning. Feature extraction transforms the raw EEG signal into a feature vector x = (x1, x2, ..., x...). n This involves highlighting key data related to neonatal brain injury outcomes by removing noise;

[0087] Step 2: Each independent sample contains 289-dimensional feature samples (i.e., the features extracted in Step 1); the Min-Max standardization method is used to standardize each dimension of the feature samples in each independent sample, and the formula for Min-Max standardization is as follows:

[0088]

[0089] Where: x′ represents the standardized one-dimensional feature value, x represents the original feature value in the independent samples, min(x) is the minimum value of the feature value in all independent samples, and max(x) is the maximum value of the feature value in all independent samples.

[0090] The preliminary feature screening module employs an Elastic Network Regression-based Feature Selection Algorithm (ENRA) to initially select features relevant to the neonatal brain injury grading results. The specific details are as follows:

[0091] Step 1: Reconstruct the n-dimensional EEG feature data from the independent samples extracted by the feature extraction module into a feature data matrix X (feature vector x = (x1, x2, ..., x...)). n The size of n is 2^89, which makes it suitable for feature selection algorithms.

[0092]

[0093] Where: x m,n The value representing the nth feature on the mth independent sample is called the feature value.

[0094] Step two: The feature data matrix X is pre-screened using the Elastic Net Regression module to obtain the largest feature sample combination after screening by the Elastic Net Regression module, which is then reconstructed into a new feature data matrix Y (feature vector y = (y1, y2, ..., y...)). s ), eigenvector y sThe selected feature is represented by 1 ≤ s ≤ n; its goal is to filter out obviously irrelevant feature data, reduce the space for the next feature search, and improve search speed and efficiency.

[0095]

[0096] Where: y m,s This represents the value of the s-th feature on the m-th independent sample after filtering.

[0097] The elastic mesh regression module uses the elastic mesh regression loss function to filter features, and the process of obtaining the elastic mesh regression loss function is as follows:

[0098] Step 1: Select a certain number of newborn EEG signal samples and input them into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data processed by the feature extraction module into a feature data matrix X. Use the feature data matrix X of all independent samples as the training set.

[0099] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set;

[0100] Step 3: Input the labeled training set into the Elastic Regression module for training. The Elastic Regression loss function of the Elastic Regression module is calculated using the following formula:

[0101]

[0102] Where: M is the number of all independent samples, y i It is the actual label value of each independent sample after feature extraction. λ1 is the predicted label value for each independent sample extracted from the features, i.e., normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, respectively; λ1 is the regularization parameter, a is the mixture parameter, p is the number of features, and β is the number of features. j It is the regression coefficient of the j-th feature;

[0103] The goal of Elastic Net Regression is to minimize the loss function described above, while simultaneously achieving feature selection and reducing multicollinearity by adjusting λ1 and α. When α = 1, the model is equivalent to Lasso regression, tending to compress some coefficients to zero, overcoming the instability of Lasso in the presence of highly correlated features, thus achieving feature selection. When α = 0, the model is equivalent to Ridge regression, tending to distribute coefficients evenly, which helps reduce multicollinearity and preserves useful feature information.

[0104] Step four: Repeat step three a specified number of times to obtain the trained Elastic Regression Matrix (SVM) module. The SVM module is not a prediction model; it assigns feature importance coefficients to the 289-dimensional features in the feature data matrix X. These features have actual labels in the training set. Through continuous training, the regression coefficients β of some unimportant features will become 0. By continuously training, these features with coefficients of 0 can be eliminated. When new unlabeled sample data is input, the feature combinations in these sample data that are the same as the feature combinations retained by the SVM module in the training set are directly retained as the new feature data matrix Y, which is then used for prediction in the subsequently trained SVM prediction model.

[0105] The optimal feature subset determination module uses the crow optimization algorithm to select the optimal feature subset that is strongly correlated with the neonatal brain injury results. The specific details are as follows:

[0106] Within the framework of a population-based crow optimization algorithm, consider a population of 10 crows, where each crow represents a new feature combination. First, from the largest feature sample combination Y output by the initial feature screening module, a subset of feature sample combinations to be searched is randomly selected. Each crow has a different subset of feature sample combinations. In each iteration, each crow maintains its own best feature combination during the iteration process, while searching other crows and updating its own best feature combination.

[0107] The crow optimization algorithm uses a specific fitness function to determine the quality of each crow's search result. By traversing and searching all feature samples in the maximum feature sample combination Y, it finally determines the subset of feature samples that are the same as the feature sample set left during training as the strongly correlated feature combination for neonatal brain injury grading. The result is then reconstructed into a new feature data matrix Z (feature vector z = (z1, z2, ..., z...)). r ), eigenvector z r The selected feature is 1 ≤ r ≤ s;

[0108]

[0109] Where: z m,r This represents the value of the r-th feature on the m-th independent sample after filtering by the optimal feature subset determination module.

[0110] The training process of the optimal feature subset determination module is as follows:

[0111] Step 1: Select a certain number of newborn EEG signal samples and input them into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data after processing by the feature extraction module into a feature data matrix X. Use the feature data matrix Y of all independent samples after preliminary feature screening as the training set.

[0112] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set (each row in matrix X represents an independent sample with the same label).

[0113] Step 3: Calculate the fitness function `fitness()` using the Raven optimization algorithm on the labeled training set as follows:

[0114]

[0115]

[0116] Where: Acc is the classification accuracy calculated using the nearest neighbor (KNN) classifier, A is a weighting factor with values ​​in [0, 1], and L... t It is the number of feature subsets selected by each crow from the maximum feature sample combination Y, 1≤t≤s, L s It is the number of features in the maximum feature sample combination Y output by the feature preliminary screening module;

[0117] Step four: Repeat step three a specified number of times to obtain the trained best feature subset establishment module. During the training process, a subset of feature combinations will be selected from the feature data matrix Y of all independent samples. These feature combinations are fixed z. New input samples can directly retain the same feature combination Z as this combination.

[0118] In this defined fitness function, the K-fold method is used to divide the dataset into training and test sets, where k is the median parameter of the K-fold method, set to 10: all independent samples are divided into 10 parts. In each cross-validation process, the model uses 9 parts as the training set and the remaining part as the test set, and this process is repeated 10 times. During the search process, each crow calculates the accuracy Acc of the selected feature subset under the KNN classifier, as well as the number of features L contained in the selected feature subset. t For the fitness function fitness(), the higher the accuracy Acc, the more features L are selected. t The smaller the value, the better the search results (i.e., the combination of feature subsets maintained by the crow);

[0119] The new feature data matrix Z (feature vector z = (z1, z2, ..., z) is used to generate the new feature data matrix Z (feature vector z = (z1, z2, ..., z)). r The data is input into the model grading and decision module for brain injury label prediction. The model grading and decision module uses a support vector machine (SVM) machine learning model to classify each independent sample, and then votes on the classification results of all independent samples. The category with the most votes is designated as the final classification result of the newborn's EEG signal sample.

[0120] The specific process of classification by the model's hierarchical decision module is as follows:

[0121] For an input sample feature vector z, its class is determined by the sign of the decision function: that is, if w i ·z+b i If the value is greater than 0, then the independent sample belongs to category i; otherwise, it belongs to other categories.

[0122] The decision rule for the entire multi-class classification problem is to select the class with the largest decision function value as the predicted class: for a new input independent sample, four decision functions f are calculated by four binary classifiers (normal, mildly abnormal, moderately abnormal, and abnormal). i (z) values: 3, -1, 5, 0, where the category corresponding to the largest decision function value will be output as the prediction result for the new input independent sample.

[0123] For example, if the maximum decision function value of a new input independent sample is 5, corresponding to the classifier for the "moderately abnormal" category, the new input independent sample will be predicted as the "moderately abnormal" category.

[0124] There are four categories (normal, mildly abnormal, moderately abnormal, and abnormal), requiring four classifiers. Each classifier's task is to distinguish between one category and all other combinations, for example:

[0125] Normal / (mild + moderate + abnormal) binary classifier decision function f i (z) value: 3;

[0126] The decision function f of the mild / (normal + moderate + abnormal) binary classifier i (z) value: -1;

[0127] The decision function f of the moderate / (normal + mild + abnormal) binary classifier i (z) value: 5;

[0128] The decision function f of the anomaly / (mild + moderate + anomaly) binary classifier i (z) value: 0;

[0129] The new input independent sample will be processed by four binary classifiers to obtain a decision function value. The largest decision function value is 5, which corresponds to the "moderately abnormal" classifier, i.e., moderate / (normal + mild + abnormal). This means that for this new input sample, the "moderately abnormal" classifier is more confident than any other classifier in believing that the new input sample belongs to the category it represents.

[0130] The classification plane was repeatedly divided into 10 times for each independent sample. The classification results were obtained 10 times. Then, the 10 classification results for each independent sample were voted on, and the category with the most votes was designated as the final classification result of the newborn's EEG signal sample.

[0131] The model classification and decision module adopts the Support Vector Machine (SVM) machine learning model, and its structure is as follows:

[0132] f i (z)=w i ·z+b i

[0133]

[0134] Where: w i It is a weight vector associated with class i, C is a regularization parameter used to balance the weights that minimize the loss and maximize the margin, y j In a neonatal brain injury classification task, the corresponding label values ​​for a specific category j (normal, mildly abnormal, moderately abnormal, and abnormal) are: 0, 1, 2, 3, and f. i (z) is the decision function for each category i, which calculates the score for predicting whether a sample belongs to category i based on the input feature vector z. The feature vector z = (z1, z2, ..., z...) r ) is the input feature vector, b i It is the bias term of the corresponding classifier;

[0135] The Support Vector Machine (SVM) machine learning model performs classification using multiple classification planes. Each classification plane employs a one-to-one strategy, treating each class as a separate category and distinguishing it from all other classes. All other classes are then merged into a single class. The constraints of the classification planes are as follows:

[0136] y i (w i ·z+b i )≥1

[0137] Where: y i It assigns a value of -1 or 1 to a certain category i (one of four categories: normal, mildly abnormal, moderately abnormal, and abnormal), and creates a binary classifier y for each category i. iWhen the value is -1, it represents the combination of other categories excluding category i, y i When the value is 1, it represents category i; the feature vector z = (z1, z2, ..., z... r ) is the input feature vector.

[0138] The specific process of training the Support Vector Machine (SVM) machine learning model is as follows:

[0139] Step 1: Input a certain number of selected neonatal EEG signal samples into the data preprocessing module, and then sequentially pass through the feature extraction module, the preliminary feature screening module, and the optimal feature subset establishment module. Finally, use all the independent samples processed by the optimal feature subset establishment module as the training set.

[0140] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3 (each row in matrix Z represents an independent sample with the same label), thus obtaining the labeled training set;

[0141] Step 3: In the labeled training set obtained in Step 2, 2062 independent samples with replacement are collected (i.e., 2062 rows in matrix Z, each row representing an independent sample). This random sampling with replacement generates multiple sample sets, which are then input into the Support Vector Machine (SVM) machine learning model in the model classification and decision module for training. During training, the performance of the SVM machine learning model on an independent validation set is monitored, and cross-validation is used to avoid overfitting. A convergence threshold is set. Once the change in the model's loss function is less than the threshold or the model's performance on the validation set no longer improves, the iteration stops, and the trained SVM machine learning model is obtained.

[0142] Example 2

[0143] A neonatal brain injury grading system based on a fusion algorithm of elastic nets and crow optimization specifically includes a data preprocessing module, a feature extraction module, a preliminary feature screening module, an optimal feature subset determination module, and a model grading decision module, wherein:

[0144] The data preprocessing module refers to cleaning the text data of neonatal EEG signals, specifically including artifact removal and missing value filling. At the same time, the cleaned neonatal EEG signal text data is sliced, resulting in 2602 neonatal EEG signal texts, which are then treated as independent samples.

[0145] The feature extraction module uses quantitative electroencephalography (qEEG) analysis to obtain neonatal brain injury feature information. The specific process is as follows: each neonatal EEG sample is decomposed into four frequency bands, namely, δ (0.5-4Hz), θ (4-7Hz), α (7-13Hz) and β (13-30Hz). Then, the amplitude, spectral density, connectivity correlation and pulse interval correlation features are extracted and standardized.

[0146] The initial feature selection module uses an Elastic Network Regression-based Feature Selection Algorithm (ENRA). The specific process of this algorithm is as follows: First, the feature data of each independent sample is reconstructed into a feature data matrix (feature vector x = (x1, x2, ..., x...)) at the input layer. n Then, the feature data matrix is ​​sent to Elastic Net Regression for preliminary pre-screening. The goal is to filter out obviously irrelevant feature data and improve the speed and efficiency of the next search step.

[0147] The optimal feature subset determination module uses the crow optimization algorithm, which approximates each feature subset as the location where a crow hides food. A specific search crow searches for each feature. Each crow has the ability to store and recall food locations (i.e., solutions for different feature subsets). A fitness function is used to determine the quality of each feature subset. Each search crow continuously updates and returns the search results. Finally, the feature subset with the largest fitness function value is considered the optimal feature subset.

[0148] The model grading and decision module uses a support vector machine (SVM) machine learning model. We feed the best feature subset data of newborns in the test set into the trained SVM machine learning model to predict the grading of newborn brain injury and output the final prediction result.

[0149] The data preprocessing module includes data cleaning and data slicing. Data cleaning includes artifact removal and missing value imputation. Artifact removal refers to the removal of physiological, power supply, and motion artifacts from neonatal EEG data before further analysis. Missing value imputation addresses the issue that some data may be missing due to instability or other problems with neonatal data entry. Samples with more than 66% missing values ​​in their neonatal EEG signal samples will be excluded from further analysis. Missing value imputation uses the most frequent data from the samples with that feature dimension.

[0150] The term "data slicing" refers to the fact that the original neonatal EEG data collected was 6 hours long. In order to include more neonatal data, rather than data from fewer neonatals with longer observation periods, each neonatal EEG signal sample was divided into 5-minute slice windows (without aliasing). These 5-minute slices were selected as independent samples.

[0151] The feature extraction module extracts features related to the neonatal brain injury grading results using quantitative electroencephalography (EEG) analysis. The specific process is as follows:

[0152] Step 1: The neonatal EEG signal samples output from the data preprocessing module are decomposed into four frequency bands: δ (0.5-4Hz), θ (4-7Hz), α (7-13Hz), and β (13-30Hz). Features are calculated within short duration windows, and these windows are shifted over time. The window length is 64 seconds with 50% overlap. Features are estimated for each channel of each short duration window sample. For the feature values ​​of multiple short duration window samples calculated in each channel of each frequency band, the median is used to summarize the feature values ​​of each channel. For the feature values ​​summarized for all channels in each frequency band, the median of all channels is taken as the feature value of the same feature in each frequency band.

[0153] Step 2: The text data of each slice result is treated as an independent sample, and each independent sample contains 289-dimensional feature samples; the data standardization process is performed using the Min-Max standardization method to standardize each feature sample. The formula for the Min-Max standardization method is as follows:

[0154]

[0155] Where: x′ represents the standardized one-dimensional feature value, x represents the original feature value in the independent samples, min(x) is the minimum value of the feature value in all independent samples, and max(x) is the maximum value of the feature value in all independent samples.

[0156] The preliminary feature selection module employs an Elastic Network Regression-based Feature Selection Algorithm (ENRA), the details of which are as follows:

[0157] First, the EEG feature data extracted from the feature extraction module is reconstructed into a feature data matrix at the input layer to adapt it for the feature selection algorithm. Then, the feature data matrix is ​​fed into Elastic Net Regression for preliminary pre-screening, where the loss function of Elastic Net Regression is calculated using the following formula:

[0158]

[0159] Where: M is the number of all independent samples, y i It is the actual label value of each independent sample after feature extraction. λ1 is the predicted label value for each independent sample, α is the regularization parameter, p is the mixture parameter, and β is the number of features. j It is the regression coefficient of the j-th feature;

[0160] The optimal feature subset determination module uses the crow optimization algorithm to select the optimal feature subset that is strongly correlated with the neonatal brain injury results. The specific details are as follows:

[0161] This algorithm approximates each feature subset as the location where a crow hides food. A specific search crow searches for each feature, using a fitness function to determine the quality of each feature subset. Each search crow continuously updates and returns search results. By traversing all feature samples in the largest feature sample combination, the algorithm ultimately determines the feature sample subset with the largest fitness function value. This subset is then used as the strongly correlated feature combination for neonatal brain injury grading. The fitness function is calculated using the following formula:

[0162]

[0163]

[0164] Where: Acc is the classification accuracy calculated using the nearest neighbor (KNN) classifier, A is a weighting factor with values ​​in [0, 1], and L... t L is the length of the feature subset selected by each crow. s It is the length of the maximum feature sample combination output by the preliminary feature screening module;

[0165] The model classification and decision module uses a Support Vector Machine (SVM) machine learning model, and the specific training process is as follows:

[0166] Step 1: Input the selected neonatal EEG signal samples into the data preprocessing module, and divide the independent samples after processing by the optimal feature subset establishment module into the training set;

[0167] Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3.

[0168] Step 3: Collect 2062 independent samples with replacement from the training set labeled in Step 2, and classify these 2062 independent samples through multiple classification planes into four categories: normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3.

[0169] Step 4: Repeat Step 3 10 times. Each time, the independent sample will be divided into classification planes to obtain 10 classification results. Then, vote on the 10 classification results of each independent sample and designate the category with the most votes as the final classification result of the newborn's EEG signal sample.

[0170] Training stops when the Support Vector Machine (SVM) machine learning model achieves a classification accuracy of 95% on the labeled training set, resulting in a well-trained SVM machine learning model. The specific structure of the SVM machine learning model is as follows:

[0171] f i (z)=w i ·z+b i

[0172]

[0173] Where: w i It is a weight vector associated with class i, C is a regularization parameter used to balance the weights that minimize the loss and maximize the margin, y j In a neonatal brain injury classification task, the corresponding label values ​​0, 1, 2, and 3 for a specific category j (normal, mildly abnormal, moderately abnormal, and abnormal) are f. i (z) is the decision function for each category i, which calculates the score for predicting whether a sample belongs to category i based on the input feature vector z. The feature vector z = (z1, z2, ..., z...) r ) is the input feature vector, b i It is the bias term of the corresponding classifier.

[0174] In step three, the classification plane employs a pairwise strategy, where each class is treated as a separate class and distinguished from all other classes (which are merged into one class). The constraints of the classification plane refer to the following part of the Support Vector Machine (SVM) machine learning model:

[0175] y i (w i ·z+b i )≥1

[0176] Where: y i The class i (one of four categories: normal, mildly abnormal, moderately abnormal, and abnormal) is assigned a value of -1 or 1. A binary classifier is created for each class i, where -1 represents the combination of other classes excluding class i. i These are the weight vectors of the corresponding classifier, and the feature vector z = (z1, z2, ..., z2). r ) is the input feature vector, b iThis is the bias term of the corresponding classifier; for an input sample feature vector z, the class is determined by the sign of the decision function: that is, if w i ·z+b i If the value is greater than 0, then the independent sample belongs to category i; otherwise, it belongs to other categories. The decision rule for the entire multi-class classification problem is to select the category with the largest decision function value as the predicted category.

[0177] This invention validates the generalization ability and reliability of the proposed method on real clinical case datasets. The fusion-based feature selection algorithm has been successfully applied to neonatal brain injury classification, providing strong support for the graded diagnosis of neonatal brain injury.

[0178] This invention studies a support vector machine (SVM) machine learning model where the difference between misclassified samples and their true categories on the test set data does not exceed one level, thus improving classification accuracy while reducing false detection rate.

Claims

1. A neonatal brain injury grading system based on a fusion algorithm of elastic nets and crow optimization, characterized in that, It includes a data preprocessing module, a feature extraction module, a preliminary feature screening module, an elastic network regression module, an optimal feature subset establishment module, and a model classification decision module. The data preprocessing module is used to receive the text data of the newborn's raw EEG signal and process the data by removing artifacts, filling missing values, and slicing the data into independent samples. The feature extraction module uses quantitative electroencephalography (QEEG) analysis to obtain quantitative information from the independent sample data after data preprocessing. For each independent sample, it extracts qEEG-related features and MNE-related features of amplitude, spectral density, connectivity correlation, and pulse train interval, and performs standardization processing. The specific content is as follows: Step 1: The quantitative EEG analysis method uses the NEURAL_py_EEG package in Python to decompose each independent sample output from the data preprocessing module into four frequency bands: δ (0.5-4Hz), θ (4-7Hz), α (7-13Hz), and β (13-30Hz). Features are calculated within a short duration window, and this window is shifted over time. The window length is 64 seconds with 50% overlap. Features are estimated for each channel of each short duration window sample. For the feature values ​​of multiple short duration window samples calculated in each channel under each frequency band, the median is used to summarize the feature values ​​of each channel. For the feature values ​​summarized for all channels under each frequency band, the median of all channels is taken as the feature value of the same feature in each frequency band, and is recorded as the multiple first features obtained. Next, using the mne package in Python, the 5-minute independent sample data processed by the data preprocessing module is imported. Using the mne.features module, the selected_funcs parameter is set to the specific features selected. The mne package will automatically process each independent sample and extract a series of time-domain, frequency-domain, and complexity features according to the specified function, which are denoted as the multiple second features obtained. Step 2: Each independent sample contains multi-dimensional feature samples; the data standardization process uses the Min-Max standardization method to standardize each feature sample. The formula for the Min-Max standardization method is as follows: in: This represents the standardized one-dimensional eigenvalue. Represents the original feature values ​​in the independent samples. It is the minimum value of this feature among all independent samples. It represents the maximum value of this feature across all independent samples; The preliminary feature screening module uses a feature selection algorithm based on elastic network regression to initially select features related to the neonatal brain injury grading results from the independent sample data after the feature extraction module. The specific content is as follows: Step 1: Reconstruct the n-dimensional EEG feature data from the independent samples extracted by the feature extraction module into a feature data matrix X and feature vectors. This adapts it to the processing of feature selection algorithms. X in: Representing the On the independent sample The numerical value of a feature, i.e., the feature value; Step two: The feature data matrix X is pre-screened using the elastic network regression module to obtain the largest feature sample combination after screening by the elastic network regression module, which is then reconstructed into a new feature data matrix Y and feature vectors. Indicates the selected feature. ; Y in: Representing the selected first On the independent sample The numerical values ​​of each feature; The Elastic Net Regression module uses the Elastic Net Regression loss function to filter features, as follows: Step 1: Input a certain number of selected neonatal EEG signal samples into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data processed by the feature extraction module into a feature data matrix X. Use the feature data matrix X of all independent samples as the training set. Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set; Step 3: Input the labeled training set into the Elastic Regression module for training. The Elastic Regression loss function of the Elastic Regression module is calculated using the following formula: Where M is the number of all independent samples. It is the actual label value for each individual sample. These are the predicted label values ​​for each independent sample, namely normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, respectively. It is a regularization parameter. It is a mixed parameter. It is the number of features. It is the regression coefficient of the j-th feature; Step four: Repeat step three a specified number of times to obtain the trained elastic network regression module; the elastic network regression module sets feature importance coefficients for the multidimensional features in the feature data matrix X, and through continuous training, the regression coefficients β of some unimportant features will become 0. By continuously training, these features with coefficients of 0 can be eliminated; when new unlabeled sample data is input, the feature combinations in these sample data that are the same as the feature combinations retained by the elastic network regression module on the training set are directly retained as the new feature data matrix Y; The optimal feature subset determination module uses the Raven optimization algorithm to further select features from the initial feature screening module, establishing the optimal feature subset that is strongly correlated with neonatal brain injury outcomes. The specific details are as follows: Within the framework of a population-based crow optimization algorithm, consider a population of 10 crows, where each crow represents a new combination of features; The crow optimization algorithm uses a specific fitness function to determine the quality of each crow's search result. It iterates through all feature samples in the maximum feature sample combination Y, ultimately identifying a subset of feature samples identical to the feature sample set left during training as the strongly correlated feature combination for neonatal brain injury grading. The result is then reconstructed into a new feature data matrix Z and feature vectors. Indicates the selected feature. ; Z in: Represents the first feature subset after being filtered by the optimal feature subset establishment module. On the independent sample The numerical values ​​of each feature; The training process of the optimal feature subset determination module is as follows: Step 1: Select a certain number of newborn EEG signal samples and input them into the data preprocessing module. Input the independent samples processed by the data preprocessing module into the feature extraction module. Reconstruct the n-dimensional EEG feature data after processing by the feature extraction module into a feature data matrix X. Use the feature data matrix Y of all independent samples after preliminary feature screening as the training set. Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set; Step 3: Calculate the fitness function of the labeled training set using the Raven optimization algorithm as follows: : Where: Acc is the classification accuracy calculated using the nearest neighbor classifier, and A is a weighting factor with values ​​in [0, 1]. It is the number of features that each crow selects from the combination of the largest feature samples Y. , It is the number of features in the maximum feature sample combination Y output by the feature preliminary screening module; Step four, repeat step three a specified number of times to obtain the trained best feature subset establishment module. During the training process, a subset of feature combinations will be selected from the feature data matrix Y of all independent samples. These feature combinations are fixed z. New input samples can directly retain the same feature combination Z as this combination. The new feature data matrix Z is input into the model grading and decision module for brain injury label prediction. The model grading and decision module uses the support vector machine (SVM) machine learning model to classify each independent sample, and then votes on the classification results of all independent samples. The category with the most votes is designated as the final classification result of the newborn's EEG signal sample. The model grading decision module uses a support vector machine (SVM) machine learning model to predict the grading of neonatal brain injury based on the best features established by the best feature subset establishment module, and outputs the prediction results: For an input sample feature vector z, its class is determined by the sign of the decision function: that is, if If a sample belongs to class i, then the independent sample belongs to class i; otherwise, it belongs to other classes. It is the weight vector associated with category i. It is the bias term of the corresponding classifier; The class with the largest decision function value is selected as the predicted class: For a new input independent sample, four decision functions are calculated by four binary classifiers respectively. The value, where the category corresponding to the largest decision function value, will be output as the prediction result for the independent sample of the new input.

2. The neonatal brain injury grading system based on a fusion algorithm of elastic nets and crow optimization as described in claim 1, characterized in that, The model classification and decision module adopts the Support Vector Machine (SVM) machine learning model, and its structure is as follows: in: It is a regularization parameter used to balance the weights of minimizing the loss and maximizing the margin. In a neonatal brain injury classification task, the corresponding label values ​​for a specific category j are: 0, 1, 2, and 3. It is the decision function for each category i, which calculates a score for predicting whether a sample belongs to category i based on the input feature vector z. It is the input feature vector; The Support Vector Machine (SVM) machine learning model performs classification using multiple classification planes. Each classification plane employs a one-to-one strategy, treating each class as a separate category and distinguishing it from all other classes. All other classes are then merged into a single class. The constraints of the classification planes are as follows: in: This involves assigning a value of -1 or 1 to a category i, and creating a separate binary classifier for each category i. When the value is -1, it represents the combination of other categories excluding category i. When the value is 1, it represents category i; feature vector z It is the input feature vector.

3. A neonatal brain injury grading system based on a fusion algorithm of elastic nets and crow optimization as described in claim 2, characterized in that, The specific process of training the Support Vector Machine (SVM) machine learning model is as follows: Step 1: Input a certain number of selected neonatal EEG signal samples into the data preprocessing module, and then sequentially pass through the feature extraction module, the preliminary feature screening module, and the optimal feature subset establishment module. Finally, use all the independent samples processed by the optimal feature subset establishment module as the training set. Step 2: Manually label the brain injury level of each independent sample in the training set as normal, mildly abnormal, moderately abnormal, and abnormal, with corresponding label values ​​of 0, 1, 2, and 3, to obtain the labeled training set; Step 3: Collect 2062 independent samples with replacement from the labeled training set obtained in Step 2. This random sampling with replacement generates multiple sample sets, which are then input into the model classification and decision module for training. During training, monitor the performance of the SVM machine learning model on an independent validation set and use cross-validation to avoid overfitting. Set a convergence threshold; once the model's loss function changes less than this threshold or the model's performance on the validation set no longer improves, the iteration stops, resulting in the trained Support Vector Machine (SVM) machine learning model.

4. A neonatal brain injury grading system based on a fusion algorithm of elastic nets and crow optimization as described in claim 1, characterized in that, The text data of neonatal EEG signals received by the data preprocessing module refers to 6 hours of continuous EEG signals from full-term newborns. The data is exported from the proprietary format of the NicoletOne machine, which is an open European data file format, and is securely stored for offline analysis. The EEGLAB toolkit in MATLAB is used to convert the acquired raw neonatal EEG data in European data file format into text data of neonatal EEG signals, and each text data of neonatal EEG signals is regarded as a neonatal EEG signal sample. Samples with missing values ​​>66% in the neonatal EEG signal samples will be excluded from further analysis. Neonatal EEG signal samples with missing values ​​<66% will be processed by missing value imputation, where missing value imputation will be performed using the mode imputation method. Then, each neonatal EEG signal sample after imputation will be segmented into 5-minute slice windows, and these slice data will be selected as independent samples.

Images