Regularization autism classification algorithm based on mutual information

Abstract

The invention relates to the technical field of mental detection, and discloses a mutual information-based regularization autism classification algorithm, which comprises the following steps of: firstly, extracting different frequency bands of theta (4-8Hz), # imgabs0 # beta (13-30Hz) and gamma (above-30Hz) from autism electroencephalogram data through a Butterworth filter, and secondly, calculating the power spectral density of a preprocessed signal through a Welch period method; according to the method, the important features with robustness can be adaptively selected by calculating the symmetric uncertainty between the features and the category labels, the important features with robustness can be adaptively selected according to the correlation between the features and the target category labels, noise or irrelevant features are eliminated, and the method not only can improve the classification precision, but also can improve the classification efficiency. And the feature dimension can be reduced, and the calculation complexity is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of mental detection, and specifically to a regularization autism classification algorithm based on mutual information. Background Art

[0002] In the process of diagnosing and treating mental disorders, the analysis of electroencephalogram (EEG) signals has always played an important role. Especially for the identification of diseases such as autism and depression, by analyzing the EEG signals of patients, direct information on brain nerve activities can be obtained non-invasively, providing an important basis for clinical diagnosis. However, due to the characteristics of EEG signals such as non-linearity, non-stationarity, and low signal-to-noise ratio, it is extremely difficult to extract valuable features for disease diagnosis from them.

[0003] Traditional methods usually rely on manual feature selection, which not only requires professional knowledge, but also is time-consuming and laborious, and is easily affected by subjective factors. In addition, due to the high-dimensionality of EEG signals, that is, the number of samples is much smaller than the number of features, this makes many traditional machine learning methods face great challenges in processing such data.

[0004] In recent years, the development of sparse regularization technology has provided new ideas for solving the above problems. By integrating the feature selection process into model training, not only can the most valuable features for the task be automatically selected from high-dimensional data, but also the overfitting problem can be effectively avoided. However, when existing sparse regularization methods such as Lasso process unbalanced data sets, they tend to select high-frequency features. Therefore, how to fairly measure the correlation between features and thus select the most valuable features for the classification task remains an unsolved problem.

[0005] Therefore, developing a new method that can adaptively select the most valuable features for autism classification and can process unbalanced data sets is of great significance for improving the diagnostic accuracy of autism. Summary of the Invention

[0006] In view of the above deficiencies in the prior art, the present invention provides a regularization autism classification algorithm based on mutual information.

[0007] The present invention provides the following technical solutions: A regularization autism classification algorithm based on mutual information, comprising the following steps:

[0008] S1. First, extract different frequency bands of θ (4 - 8 Hz), (8 - 13 Hz), β (13 - 30 Hz) and γ (above 30 Hz) from the autism EEG data through a Butterworth filter, and then calculate the power spectral density of the preprocessed signal by the Welch periodogram method;

[0009] S2. Measure features through symmetric uncertainty:

[0010] Given two discrete variables X = (x1, x2,..., x n ) and Y = (y1, y2,..., y n ), the symmetric uncertainty between them is:

[0011]

[0012] where H(X) and H(Y) represent the information entropy of variables X and Y respectively, and I(X, Y) is the mutual information between variables X and Y. The symmetric uncertainty normalizes the value to between [0, 1];

[0013] S3. Adopt the SU-LASSO embedded feature selection method in the adaptive LASSO method to identify highly discriminative features from the constructed feature set;

[0014] Its objective function is:

[0015] Here, X = (x1, x2,..., x n ) ∈ R n×p represents the data matrix constructed from the frequency domain features extracted from the power spectral density, where x i = (x 1i , x 2i ,..., x ni ) T ∈ R n×1 represents the features extracted from each EEG signal sample, y = (y1, y2,..., y n ) T ∈ R n is the response variable, β = (β1, β2,..., β n ) T ∈ R P is the unknown coefficient vector of the model, λ is the penalty parameter used to adjust the penalty degree, s i is the symmetric uncertainty between the i-th feature and the class label in the constructed feature space, and r is the adjustment parameter;

[0016] S4. Select an SVM classifier for autism recognition from EEG data, map the samples in the input space to a high-dimensional feature space through a non-linear transformation, and then find the optimal classification surface in the feature space that can linearly separate the samples.

[0017] Preferably, when there is a strong correlation between the feature and the class label, s i tends to 1, then a smaller penalty is imposed on the coefficient associated with the feature; when there is a weak correlation between the feature and the class label, s iIf it approaches 0, a greater penalty will be imposed on the model coefficient corresponding to the feature.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] By calculating the symmetric uncertainty between the feature and the class label, important features with robustness can be adaptively selected. According to the correlation between the feature and the target class label, important features with robustness are adaptively selected, and noise or irrelevant features are removed. This method can not only improve the classification accuracy but also reduce the feature dimension and computational complexity. Description of the Drawings

[0020] Figure 1 It is a schematic structural diagram of the present invention;

[0021] Figure 2 It is a schematic diagram of the distribution of 128 electrodes in Experiment 1 of the present invention;

[0022] Figure 3 It is a schematic diagram of the positions of the effective channels in the experiment of the present invention;

[0023] Figure 4 It is a comparison diagram of the classification results on the MODMA and EDRA data sets of the present invention. Detailed Embodiments

[0024] In order to make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present disclosure. To keep the following description of the embodiments of the present disclosure clear and concise, detailed descriptions of known functions and known components are omitted to avoid unnecessarily obscuring the concepts of the present invention.

[0025] Please refer to Figure 1 , the regularization autism classification algorithm based on mutual information, includes the following steps:

[0026] S1. First, extract different frequency bands of θ (4 - 8 Hz), (8 - 13 Hz), β (13 - 30 Hz), and γ (above 30 Hz) from the autism EEG data through a Butterworth filter. Secondly, calculate the power spectral density of the preprocessed signal by the Welch periodogram method;

[0027] S2. Measure the feature by symmetric uncertainty:

[0028] Given two discrete variables X = (x1, x2,..., x n ) and Y = (y1, y2,..., y n ), the symmetric uncertainty between them is:

[0029]

[0030] Among them, H(X) and H(Y) represent the information entropy of variable X and variable Y respectively, and I(X, Y) is the mutual information between variable X and Y. The symmetric uncertainty normalizes the value to between [0, 1]; when the value is 0, it means that variable X and Y are not related, and vice versa. In addition, measuring the importance of features through symmetric uncertainty can effectively reduce the impact of noise on the recognition model.

[0031] S3. Adopt the SU-LASSO embedded feature selection method in the adaptive LASSO method to identify highly discriminative features from the constructed feature set;

[0032] Its objective function is:

[0033] Here, X = (x1, x2,..., x n ) ∈ R n×p represents the data matrix constructed from the frequency domain features extracted from the power spectral density, where x i = (x 1i , x 2i ,..., x ni ) T ∈ R n×1 represents the features extracted from each EEG signal sample, y = (y1, y2,..., y n ) T ∈ R n is the response variable, β = (β1, β2,..., β n ) T ∈ R P is the unknown coefficient vector of the model, λ is the penalty parameter used to adjust the penalty degree, s i is the symmetric uncertainty between the i-th feature and the class label in the constructed feature space, and r is the adjustment parameter; when there is a strong correlation between the feature and the class label, s i tends to 1, then a smaller penalty is imposed on the coefficient associated with the feature; when there is a weak correlation between the feature and the class label, s i tends to 0, then a larger penalty is imposed on the model coefficient corresponding to the feature. This SU-LASSO method can adaptively select robust and important features according to the correlation between the features and the target class labels, and eliminate noise or irrelevant features.

[0034] S4. Select an SVM classifier for autism recognition of EEG data, map the samples in the input space to a high-dimensional feature space through a non-linear transformation, and then find the optimal classification surface in the feature space that can linearly separate the samples. The selection criterion for the optimal classification surface is that it can accurately divide the two types of samples into the maximum interval, so as to realize the classification of data samples.

[0035] Experiment

[0036] To verify the effectiveness of the algorithm in this project, a multi-modal open dataset MODMA for mental disorder analysis provided by the General Perception and Intelligent Systems Laboratory of a certain university was selected. This dataset records the electroencephalogram (EEG) signals of 53 subjects carefully selected by professional doctors in the hospital in the resting state and the stimulated state, including 24 patients with depression and 29 normal control groups. The exclusion criteria for participants include individuals with mental disorders or brain organ damage, severe physical diseases, or those with severe suicidal tendencies among patients with severe depression. The exclusion criteria for the control group include participants with a personal or family history of mental disorders. In addition, individuals who have abused or been dependent on alcohol or psychotropic drugs in the past year, pregnant, lactating women, or women taking contraceptives are excluded from the study. These exclusion criteria ensure that the participants are healthy and have not been exposed to any substances that may affect their brain function. This criterion also helps to reduce any potential confounding factors that may affect the results and increase the internal validity of the study. Before the experiment, all subjects were evaluated through the Patient Health Questionnaire-9 (PHQ-9). The data was specifically collected by 128 leads, as Figure 2 shown. The sampling frequency was 250 Hz, and the duration of each data segment was about 5 minutes.

[0037] To verify the wide applicability of the method proposed in this project, a second dataset EDRA was also selected for this experiment. This dataset was collected by Henan University and contains the EEG records of 50 subjects, among which 26 subjects belong to the high-score group (i.e., the high-risk depression group), and 24 subjects belong to the low-score group (i.e., the low-risk control group). During the experiment, all subjects completed the evaluation of the depression scale (PROMIS) and questionnaires. The EDRA dataset was collected from 62 electrodes placed above the right eye using an international standard EEG cap, and the positions of the effective channels are as Figure 3 shown, and the sampling frequency was 500 Hz.

[0038] In this experiment, the original brain wave data was processed by feature extraction and feature selection based on SU-LASSO to generate a compact and highly informative feature set. Subsequently, the selected features were input into an SVM classifier for binary classification. Here, the five-fold cross-validation method was used to verify the performance of the proposed framework. Figure 4 The average precision comparison results between the proposed method and similar methods are given. From Figure 4It can be seen that on the MODMA dataset, the proposed method achieved an accuracy of 98.33%, while on the EDRA dataset, it reached an accuracy of 96.26%. Additionally, through comparative analysis with advanced algorithms in the field, such as ShallowConvNet, EEGNet, and DeepConvNet, our method obtained the best classification accuracy. Specifically, the average accuracy of our method on the MODMA and EDRA datasets was 4.16% and 1.09% higher than that of the sub-optimal method DeepConvNet, respectively. In summary, the disease detection algorithm applied in this system has a fast recognition speed and a recognition rate superior to other algorithms in the same field, and has broad application prospects.

[0039] The above embodiments are only exemplary embodiments of the present invention and are not used to limit the present invention. The protection scope of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements within the essence and protection scope of the present invention, and such modifications or equivalent replacements should also be regarded as falling within the protection scope of the present invention.

Claims

1. A regularization autism classification algorithm based on mutual information, characterized in that Including the following steps: S1. First, extract different frequency bands of θ (4 - 8 Hz), (8 - 13 Hz), β (13 - 30 Hz), and γ (above - 30 Hz) from the EEG data of autism using a Butterworth filter. Secondly, calculate the power spectral density of the pre - processed signal by the Welch periodogram method; S2. Measure features through symmetric uncertainty: Given two discrete variables X = (x1, x2, …, x n ) and Y = (y1, y2, …, y n ), the symmetric uncertainty between them is: Among them, H(X) and H(Y) represent the information entropy of variable X and variable Y respectively, and I(X, Y) is the mutual information between variable X and Y. The symmetric uncertainty normalizes the value to between [0, 1]; S3. Adopt the SU-LASSO embedded feature selection method in the adaptive LASSO method to identify highly discriminative features from the constructed feature set; Its objective function is as follows: Here, X=(x1, x2, …, x n ) ∈ R n×p represents the data matrix constructed from the frequency-domain features extracted from the power spectral density, where x i =(x 1i , x 2i , …, x ni ) T ∈ R n×1 represents the features extracted from each EEG signal sample, y=(y1, y2, …, y n ) T ∈ R n is the response variable, β=(β1, β2, …, β n ) T ∈ R P is the vector of unknown coefficients of the model, λ is the penalty parameter used to adjust the degree of penalty, si is the symmetric uncertainty between the i-th feature and the class label in the constructed feature space, and r is the adjustment parameter; S4. Select an SVM classifier for autism recognition from EEG data, map the samples in the input space to a high-dimensional feature space through a non-linear transformation, and then find the optimal classification surface in the feature space that can linearly separate the samples.

2. The regularization autism classification algorithm based on mutual information according to claim 1, characterized in that: When there is a strong correlation between the feature and the class label, s i tends to 1, and a smaller penalty is imposed on the coefficient associated with the feature; When there is a weak correlation between the feature and the class label, s i tends to 0, and a large penalty will be imposed on the model coefficient corresponding to the feature.

Images