Epilepsy prediction system based on mixed feature extraction

Abstract

The invention relates to an epilepsy prediction system based on mixed feature extraction. The epilepsy prediction system comprises a data preprocessing module, a data denoising module, a mixed extraction module and a model training module, the method has the advantages that 1) high-efficiency denoising: data denoising is performed by using a convolution auto-encoder, so that the quality of EEG signals is remarkably improved, and particularly, physiological information can be effectively kept in a complex noise environment; 2) comprehensive feature extraction: combining experience extraction and an automatic feature extraction method to comprehensively capture time-frequency and time-sequence features in the EEG signals, significantly improving depth and breadth of feature extraction, and ensuring richness of data dimensions; and 3) efficient feature screening: performing feature screening by adopting a feature correlation matrix, so that redundant features can be efficiently eliminated, a feature set is optimized, and the classification precision and the classification efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of epilepsy prediction, and more specifically, to an epilepsy prediction system based on hybrid feature extraction. Background Art

[0002] Epilepsy is a common chronic non-communicable brain disease, characterized by sudden abnormal discharges of brain neurons, which lead to transient dysfunctions. Epileptic seizures may have various symptom manifestations, including loss of consciousness, limb convulsions, sensory abnormalities, etc., and may even endanger the patient's life in severe cases. According to the statistics of the World Health Organization (WHO), as of 2024, there are approximately 50 million epilepsy patients globally, and nearly 30% of them are ineffective in the existing drug treatments and belong to intractable epilepsy.

[0003] The suddenness and unpredictability of epileptic seizures cause patients to face huge physical and psychological pressures for a long time. At the same time, the effectiveness of existing treatment methods for intractable epilepsy still has limitations. Therefore, it is particularly important to develop an efficient and accurate epilepsy prediction method. Electroencephalogram (EEG) is an important tool for epilepsy prediction, which can directly reflect the electrical activities of brain neurons. Traditional epilepsy prediction methods mainly rely on doctors' visual analysis of EEG signals, but this method is time-consuming and easily affected by subjective factors. In recent years, with the development of artificial intelligence and signal processing technologies, epilepsy prediction algorithms based on machine learning and deep learning have gradually become a research hotspot.

[0004] Currently, the EEG-based epilepsy prediction technology faces many challenges, and the complexity of EEG signals is one of the main bottlenecks. EEG signals are non-linear and non-stationary, and are easily affected by noise interference (such as electromyogram noise, eye movement artifacts), resulting in difficult feature extraction. Existing denoising methods usually rely on traditional filtering techniques (such as band-pass filtering, wavelet threshold denoising), and these methods have limited effects in dealing with complex noise and are prone to losing useful physiological information. In terms of feature extraction, existing technologies mainly rely on manually designed features (such as time-frequency features, non-linear dynamics features) or features automatically extracted by deep learning models. However, manually designing features requires rich domain knowledge and is difficult to capture complex patterns in EEG signals; while a pure deep learning model (such as a convolutional neural network CNN) can automatically extract features, but is sensitive to noise and is prone to overfitting on small sample data. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an epilepsy prediction system based on hybrid feature extraction.

[0006] To solve the above technical problems, the technical solution of the present invention is: an epilepsy prediction system based on hybrid feature extraction, including a data preprocessing module, a data denoising module, a hybrid extraction module, and a model training module;

[0007] The data preprocessing module is used to preprocess the sample data set to obtain corresponding sample segment data;

[0008] The data denoising module includes a preset convolutional encoder and a convolutional decoder. The data denoising module processes the sample data segment through the convolutional encoder to obtain encoded low-dimensional data, and processes the encoded low-dimensional data through the convolutional decoder to obtain decoded reconstructed data;

[0009] The hybrid extraction module includes a wavelet decomposition unit, a time series extraction unit, and a data fusion unit. The wavelet decomposition unit extracts the time-frequency features in the decoded reconstructed data through the DB4 wavelet decomposition algorithm, and generates corresponding statistical features according to a preset time-frequency statistical table. The time series extraction unit extracts the time series features in the decoded reconstructed data through the LSTM network. The data fusion unit fuses the time-frequency features and the time series features to obtain a data comprehensive feature subset;

[0010] The model training module is configured with a support vector machine, and trains the support vector machine through the comprehensive feature subset to obtain a corresponding epilepsy prediction classifier.

[0011] Further: The data denoising module further includes a mapping function training unit. The mapping function training unit includes constructing a basic mapping function, and generating an encoding mapping function and a decoding mapping function according to the basic mapping function. The encoding mapping function and the decoding mapping function respectively include a number of coordination parameters. The mapping function training unit is configured with a loss function and a parameter adjustment mapping sub-strategy. The loss function is used to calculate the reconstruction error during each training. The parameter adjustment mapping sub-strategy is used to calculate and generate a parameter adjustment instruction set according to the loss function, and adjust the corresponding coordination parameters according to the parameter adjustment instruction set to reduce the reconstruction error.

[0012] Further: The parameter adjustment mapping sub-strategy includes a number of parameter adjustment trigger conditions and parameter adjustment mapping sub-functions corresponding to the parameter adjustment trigger conditions. The parameter adjustment mapping sub-strategy retrieves the corresponding parameter adjustment mapping sub-functions according to the parameter adjustment trigger conditions, and brings the currently obtained reconstruction error into the parameter adjustment mapping sub-functions to obtain corresponding parameter adjustment instruction values, and aggregates the parameter adjustment instruction values to generate a parameter adjustment instruction set.

[0013] Further: The hybrid extraction module further includes a data dimensionality reduction unit, which is configured with a dimensionality reduction sub-strategy. The dimensionality reduction sub-strategy includes matching a corresponding plurality of adaptation channels for each decoded and reconstructed data, calculating the matching value of each adaptation channel through a preset channel value algorithm, screening the adaptation channels with matching values higher than a preset matching threshold, and applying the adaptation channels to the decoded and reconstructed data.

[0014] Further: The data preprocessing module is configured with a sliding window sub-strategy. The length of the sliding window is set to 5 - 10 seconds. The sliding window sub-strategy is configured with a sliding overlap ratio, and the sliding overlap ratio is proportional to the mean value of the reconstruction error in the previous acquisition process.

[0015] Further: The time-frequency statistical generation table includes a number of statistical items, and each statistical item is used to generate a corresponding statistical feature. The statistical items include a mean item, a variance item, a skewness item, and a kurtosis item.

[0016] Further: It further includes a feature screening module, which is configured with a feature screening strategy. The feature screening strategy is used to eliminate redundant features from the comprehensive feature subset. The feature screening strategy is configured with a correlation calculation algorithm to calculate the correlation value of each data feature. When the correlation value is less than a preset reference correlation value, the data feature is determined to be a redundant feature.

[0017] The technical effects of the present invention are mainly reflected in the following aspects: 1) High-efficiency denoising: Using a convolutional autoencoder for data denoising significantly improves the quality of EEG signals, especially in a complex noise environment, and can effectively maintain physiological information. 2) Comprehensive feature extraction: Combining empirical extraction and automatic feature extraction methods can comprehensively capture the time-frequency and time-sequence features in EEG signals, significantly enhancing the depth and breadth of feature extraction and ensuring the richness of data dimensions. 3) High-efficiency feature screening: Using a feature correlation matrix for feature screening can efficiently eliminate redundant features, optimize the feature set, and improve the classification accuracy and classification efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 : Schematic diagram of the architecture principle of the epilepsy prediction system based on hybrid feature extraction of the present invention;

[0019] Figure 2 : Working flowchart of the epilepsy prediction system based on hybrid feature extraction of the present invention;

[0020] Figure 3 : Schematic diagram of the convolutional encoder structure of the epilepsy prediction system based on hybrid feature extraction of the present invention;

[0021] Figure 4 : Schematic diagram of the waveform of the original EEG signal;

[0022] Figure 5 :Waveform schematic diagram after denoising using a convolutional encoder.

[0023] Reference numerals: 100, data preprocessing module; 200, data denoising module; 300, hybrid extraction module; 400, feature screening module; 500, model training module. Detailed implementation manners

[0024] The following further details the detailed implementation manners of the present invention in conjunction with the accompanying drawings, so that the technical solutions of the present invention are easier to understand and master.

[0025] An epilepsy prediction system based on hybrid feature extraction, including a data preprocessing module 100, a data denoising module 200, a hybrid extraction module 300, a feature screening module 400, and a model training module 500;

[0026] The data preprocessing module 100 is used to preprocess the sample data set to obtain corresponding sample segment data; the data preprocessing module 100 is configured with a sliding window sub-strategy, the sliding window length is set to 5 - 10 seconds, the sliding window sub-strategy is configured with a sliding overlap ratio, and the sliding overlap ratio is proportional to the mean of the reconstruction errors in the previous acquisition process. For example, taking the publicly available epilepsy patient EEG signal data set CHB - MIT as the research data, defining 5 minutes before the seizure as the pre-seizure period, using a sliding window to divide the data into segments of 5-second length, and respectively extracting interictal and pre-seizure data, selecting the common electrode channels in the data to ensure data consistency, and a function can be set to calculate the average reconstruction error of the previous acquisition band. If the reconstruction error is large, the repetition ratio before and after the sliding window is increased. For example, if the sliding overlap ratio is selected as 20%, then the first acquisition is from 0 - 5 seconds, and the second acquisition is from 4 - 9 seconds, so as to increase the number of samples and avoid incomplete recognition of some features due to the sliding window.

[0027] The data denoising module 200 includes a preset convolutional encoder and a convolutional decoder. The data denoising module processes sample data segments through the convolutional encoder to obtain encoded low-dimensional data, and processes the encoded low-dimensional data through the convolutional decoder to obtain decoded reconstructed data. One of the innovations of this method is to use a Convolutional Autoencoder (CAE) for EEG signal denoising. Traditional denoising methods usually rely on techniques such as wavelet decomposition or filtering, but these methods have limited effects under complex noise and may lose key information. The CAE can automatically optimize the denoising process according to different noise conditions through an adaptive unsupervised learning process, maximizing the retention of physiological information in the EEG signal. During the denoising process, the CAE can provide stronger denoising capabilities for problems such as complex non-Gaussian noise and electromyogram noise. The CAE is an unsupervised learning model that maps the input signal to a low-dimensional representation through an encoder and decoder structure, and reconstructs the signal during the decoding process, removing noise components and retaining effective physiological information. Its goal is to minimize the error between the input signal and the reconstructed signal, usually optimized using the Mean Squared Error (MSE) as the loss function. The CAE is a variant of the autoencoder that uses a Convolutional Neural Network (CNN) to replace the traditional fully connected layers. This method can effectively suppress noise and improve signal quality when processing EEG signals, providing more accurate data for subsequent analysis. The specific working principle of the autoencoder is that the encoder maps the input data x to a low-dimensional representation z through a series of neural network layers:

[0028] z = f e (x) (1)

[0029] where f e is the mapping function of the encoder.

[0030] After that, the decoder reconstructs the low-dimensional representation z into the output data through a series of neural network layers

[0031]

[0032] where f d is the mapping function of the decoder.

[0033] The data denoising module further includes a mapping function training unit, and the mapping function training unit includes constructing a basic mapping function and generating an encoding mapping function f e and a decoding mapping function f d, the encoding mapping function and the decoding mapping function respectively include a number of coordination parameters. The mapping function training unit is configured with a loss function and a parameter adjustment mapping sub-strategy. The loss function is used to calculate the reconstruction error during each training, and the parameter adjustment mapping sub-strategy is used to generate a parameter adjustment instruction set according to the loss function and adjust the corresponding coordination parameters according to the parameter adjustment instruction set to reduce the reconstruction error. The goal of the autoencoder is to minimize the difference between the input data x and the reconstructed data therebetween. Usually, the mean square error (MSE) is used as the loss function:

[0034]

[0035] Optimize the parameters of the encoder and decoder through the backpropagation algorithm to minimize the reconstruction error.

[0036] The parameter adjustment mapping sub-strategy includes a number of parameter adjustment trigger conditions and parameter adjustment mapping sub-functions corresponding to the parameter adjustment trigger conditions. The parameter adjustment mapping sub-strategy retrieves the corresponding parameter adjustment mapping sub-functions according to the parameter adjustment trigger conditions, and brings the currently obtained reconstruction error into the parameter adjustment mapping sub-functions to obtain the corresponding parameter adjustment instruction values, and aggregates the parameter adjustment instruction values to generate a parameter adjustment instruction set. Specifically, referring to Figure 3 as shown, the coordination parameters can be configured in the first convolutional layer, the second convolutional layer, the third convolutional layer, the upsampling layer and the downsampling layer. According to different parameter adjustment trigger conditions, different parameter adjustment mapping sub-functions are generated to reduce the reconstruction error after adjustment. The parameter adjustment trigger conditions can be set to different waveform feature recognition results of the corresponding output data, and different parameter adjustment sub-functions are called according to different waveform features. So that the entire convolutional unit is continuously optimized, while ensuring useful data, a strong filtering and denoising effect is achieved.

[0037] Referring to Figure 2As shown, the hybrid extraction module 300 includes a wavelet decomposition unit, a timing extraction unit, and a data fusion unit. The wavelet decomposition unit extracts the time-frequency features in the decoded and reconstructed data through the DB4 wavelet decomposition algorithm, and generates corresponding statistical features according to a preset time-frequency statistical table. The time-frequency statistical generation table includes several statistical items, and each statistical item is used to generate a corresponding statistical feature. The statistical items include a mean item, a variance item, a skewness item, and a kurtosis item, that is, the specific mean, variance, skewness, and kurtosis values of the corresponding video features are calculated according to the corresponding statistical items. In the feature extraction stage, the present invention combines the deep learning characteristics of wavelet decomposition and the LSTM long short-term memory network to form a powerful time-frequency and timing feature capture mechanism. Wavelet decomposition is used to extract the time-frequency features of EEG signals and can well capture the dynamic changes before epileptic seizures; while the LSTM long short-term memory network can automatically extract the time-dependent features in EEG signals and capture complex non-linear patterns. Combining these two methods makes the feature extraction more comprehensive and can better cope with the complexity and non-stationarity in EEG signals. When applying, first use principal component analysis (PCA) to reduce the dimension of the electrode channel number to reduce the data dimension and data complexity. Then, manually extract features (extract time-frequency features through DB4 wavelet decomposition and calculate statistical features) and automatically extract features (extract timing features through LSTM) from the data after dimension reduction to comprehensively capture the time-frequency and timing information in EEG signals. Finally, fuse the manually extracted features and the automatically extracted features to construct a comprehensive feature set.

[0038] The timing extraction unit extracts the timing features in the decoded and reconstructed data through the LSTM network, and the data fusion unit fuses the time-frequency features and the timing features to obtain a data comprehensive feature subset;

[0039] The hybrid extraction module 300 further includes a data dimension reduction unit, and the data dimension reduction unit is configured with a dimension reduction sub-strategy. The dimension reduction sub-strategy includes matching a corresponding plurality of adaptation channels for each decoded and reconstructed data, calculating the matching value of each adaptation channel through a preset channel value algorithm, screening the adaptation channels with a matching value higher than a preset matching threshold, and applying the adaptation channels to the decoded and reconstructed data.

[0040] Feature screening module 400, the feature screening module is configured with a feature screening strategy, the feature screening strategy is used to eliminate redundant features from the comprehensive feature subset, the feature screening strategy is configured with a correlation calculation algorithm for calculating the correlation value of each data feature, and when the correlation value is less than the preset reference correlation value, it is determined that the data feature is a redundant feature. By calculating the feature correlation matrix to eliminate redundant features and screening out effective features. This method is more efficient than traditional feature selection methods (such as analysis of variance, recursive feature elimination, etc.), especially in high-dimensional feature spaces, it can operate better, greatly improving the processing speed and automation level. Calculate the feature correlation matrix, screen out representative and high-value features by eliminating redundant features, and set the correlation threshold to 0.9. The formula for the Pearson correlation coefficient used to calculate feature correlation is as follows:

[0041]

[0042] where X i (k) and X j (k) respectively represent the kth observation value of feature X i and X j . and respectively represent the means of features X i and X j . X i and X j are two different data features respectively. The correlation value is calculated using a superposition algorithm to determine the data with a higher correlation value as the result.

[0043] The model training module 500 is configured with a support vector machine, and the support vector machine is trained through the comprehensive feature subset to obtain the corresponding epilepsy prediction classifier. Finally, the support vector machine (SVM) is used to classify the screened features to achieve epilepsy prediction. Compared with traditional classification methods, SVM has stronger anti-noise ability in high-dimensional spaces, can better handle small sample data, and improve the robustness of prediction.

[0044] In a specific embodiment, the CHB-MIT dataset (Children's Hospital Boston MIT Scalp EEG Database) is a widely used electroencephalogram (EEG) dataset collected in cooperation between Children's Hospital Boston and the Massachusetts Institute of Technology (MIT), which contains 24 medical records. To ensure data consistency, 22 electrode channels common to different records are selected, including 'P3-O1', 'FZ-CZ', 'P8-O2', 'F4-C4', 'F7-T7', 'F8-T8', 'P7-O1', 'FP1-F3', 'FP2-F4', 'FP1-F7', 'FT9-FT10', 'CZ-PZ', 'C3-P3', 'T7-P7', 'FT10-T8', 'P4-O2', 'T7-FT9', 'C4-P4', 'T8-P8', 'P7-T7', 'FP2-F8', 'F3-C3'. After segmentation, classification, and sorting, there are a total of 14,940 data, including 7,560 early-stage data and 7,380 interphase data.

[0045] Next, a denoising task is performed using a convolutional autoencoder. In a specific embodiment, the schematic diagram of the convolutional autoencoder structure used is as Figure 2 shown. In the encoder part, a convolutional layer with 64 filters is first used for convolution operations, then the spatial dimension is reduced through a max-pooling layer, and Dropout is used to prevent overfitting; then a convolutional layer with 128 filters is used, pooled again, and Dropout is used; next, two convolutional layers with 256 filters are used to further enhance the network's expressive ability. In the decoder part, the spatial dimension is increased through an upsampling layer, and Dropout is used to avoid overfitting; then a convolutional layer with 128 filters is used to adjust the feature map, continue upsampling to restore the size, and Dropout is used again; finally, a convolutional layer with 1 filter is used to generate a single-channel output image, and the size of the output image is adjusted through a cropping layer. The average mean square error MSE of the denoised data is 0.0004, and the average signal-to-noise ratio SNR is 15.2509, demonstrating a good denoising effect. The schematic diagrams of the waveforms of the original data and the data after denoising using the convolutional autoencoder are as Figure 3 、 4 shown.

[0046] Principal component analysis (PCA) is applied to the denoised data to reduce the number of electrode channels. The threshold for the variance explanation degree of the principal components is set to 0.8. All data is traversed to determine the maximum number of channels to be retained, and this number of channels is applied to all data to ensure data consistency and unity. Finally, after dimensionality reduction processing, the number of electrode channels for all data is unified to 13.

[0047] Then feature extraction is carried out. First, the time-frequency domain features are extracted from the preprocessed data using the DB4 wavelet transform, and four types of statistical features such as mean, variance, skewness, and kurtosis are manually calculated. Then, the preprocessed data is input into the LSTM model for automatic feature extraction. The LSTM layer contains 128 units, followed by a Dense layer with 64 units, and finally a Dense layer with 1 unit. Then, the manually extracted features are combined with the features automatically extracted by the LSTM to form the final feature set.

[0048] Finally, the support vector machine (SVM) is used to classify the interictal and preictal data. Through ten-fold cross-validation, the average accuracy of the model is 76.21%, and the accuracy on the test set is 77.18%. For the positive class (predicted to have a seizure), the prediction sensitivity of the model is 0.79, and for the negative class (predicted not to have a seizure), the prediction sensitivity is 0.76. For the precision, the prediction precision of the positive class is 0.77, and the prediction precision of the negative class is 0.78.

[0049] In summary, the present invention provides an epilepsy prediction system that combines manual and automatic feature extraction. By integrating traditional methods, machine learning, and deep learning technologies, a multi-stage EEG signal processing flow is proposed, and the prediction of epileptic seizures is successfully achieved. The experimental results show that the system performs excellently in data dimensionality reduction, feature extraction, feature screening, and classification, and has good prediction performance, providing an effective and reliable technical solution for epilepsy prediction.

[0050] Of course, the above are only typical examples of the present invention. In addition, the present invention can also have many other specific implementation manners. Any technical solutions formed by equivalent replacement or equivalent transformation fall within the scope of protection required by the present invention.

Claims

1. An epilepsy prediction system based on hybrid feature extraction, characterized by: It includes data preprocessing module, data denoising module, hybrid extraction module and model training module; The data preprocessing module is used to preprocess the sample data set to obtain corresponding sample segment data; The data denoising module includes a preset convolutional encoder and a convolutional decoder. The data denoising module processes the sample data segments through the convolutional encoder to obtain encoded low-dimensional data, and processes the encoded low-dimensional data through the convolutional decoder to obtain decoded reconstructed data; The hybrid extraction module includes a wavelet decomposition unit, a time series extraction unit and a data fusion unit. The wavelet decomposition unit extracts the time-frequency features in the decoded and reconstructed data through the DB4 wavelet decomposition algorithm, and generates corresponding statistical features according to a preset time-frequency statistical table. The time series extraction unit extracts the time series features in the decoded and reconstructed data through the LSTM network. The data fusion unit fuses the time-frequency features and the time series features to obtain a data comprehensive feature subset. The model training module is configured with a support vector machine, and the support vector machine is trained by comprehensive feature subsets to obtain a corresponding epilepsy prediction classifier.

2. The epilepsy prediction system based on hybrid feature extraction according to claim 1, characterized in that: The data denoising module also includes a mapping function training unit, which includes constructing a basic mapping function and generating an encoding mapping function and a decoding mapping function based on the basic mapping function. The encoding mapping function and the decoding mapping function respectively include a number of coordination parameters. The mapping function training unit is configured with a loss function and a parameter adjustment mapping sub-strategy. The loss function is used to calculate the reconstruction error during each training. The parameter adjustment mapping sub-strategy is used to generate a parameter adjustment instruction set based on the loss function calculation, and adjust the corresponding coordination parameters according to the parameter adjustment instruction set to reduce the reconstruction error.

3. The epilepsy prediction system based on hybrid feature extraction according to claim 2, characterized in that: The parameter mapping sub-strategy includes several parameter tuning trigger conditions and parameter tuning sub-functions corresponding to the parameter tuning trigger conditions. The parameter tuning sub-strategy calls the corresponding parameter tuning sub-function according to the parameter tuning trigger conditions, and brings the currently obtained reconstruction error into the parameter tuning sub-function to obtain the corresponding parameter tuning instruction value, and aggregates the parameter tuning instruction values ​​to generate a parameter tuning instruction set.

4. The epilepsy prediction system based on hybrid feature extraction according to claim 2, characterized in that: The hybrid extraction module also includes a data dimensionality reduction unit, which is configured with a dimensionality reduction sub-strategy. The dimensionality reduction sub-strategy includes matching corresponding multiple adaptation channels for each decoded and reconstructed data, calculating the matching value of each adaptation channel through a preset channel value algorithm, screening adaptation channels with matching values ​​higher than a preset matching threshold, and applying the adaptation channels to the decoded and reconstructed data.

5. The epilepsy prediction system based on hybrid feature extraction according to claim 4, characterized in that: The data preprocessing module is configured with a sliding window sub-strategy, the sliding window length is set to 5-10 seconds, and the sliding window sub-strategy is configured with a sliding overlap ratio, which is proportional to the mean of the reconstruction error in the previous acquisition process.

6. The epilepsy prediction system based on hybrid feature extraction according to claim 1, characterized in that: The time-frequency statistics generation table includes several statistical items, each of which is used to generate corresponding statistical features, and the statistical items include a mean item, a variance item, a skewness item, and a kurtosis item.

7. The epilepsy prediction system based on hybrid feature extraction according to claim 1, characterized in that: It also includes a feature screening module, which is configured with a feature screening strategy. The feature screening strategy is used to eliminate redundant features from the comprehensive feature subset. The feature screening strategy is configured with a correlation calculation algorithm for calculating the correlation value of each data feature. When the correlation value is less than a preset baseline correlation value, the data feature is determined to be a redundant feature.

Images