An Epilepsy Grading Prediction Model Based on an Improved CNN Fusion Model

Abstract

This invention belongs to the field of computer application technology and discloses an epilepsy grading prediction model based on an improved CNN fusion model. The model includes an EEG signal segmentation module, a synchronous compressed wavelet transform module, a data processing module, a model fusion module, and a prediction probability processing module. The EEG signal segmentation module is used to segment the EEG signal into different time periods. The synchronous compressed wavelet transform module processes the EEG signal using synchronous compressed wavelet transform. The data processing module includes an MLF-CNN model and a semi-dilated convolutional network model for data processing. The model fusion module fuses the MLF-CNN model and the semi-dilated convolutional network model. The prediction probability processing module uses a self-offset post-processing method to process the prediction probability. The self-offset fusion model of this invention has a significant performance improvement compared to individual models and can be used on clinical datasets, exhibiting generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of computer application technology, and in particular relates to an epilepsy grading prediction model based on an improved CNN fusion model. Background Technology

[0002] In recent years, the number of people suffering from epilepsy has been increasing. Due to the randomness of epileptic seizures, this places a significant psychological burden on patients and poses a potential risk of harm. Therefore, designing a highly sensitive seizure prediction algorithm is of great importance. Among these, the epilepsy grading prediction model is an important application of deep learning and data mining techniques in the field of biomedical big data research. It can effectively predict epileptic seizures based on the resting scalp electroencephalogram (EEG) signals of epilepsy patients and promptly notify medical staff or patients to take appropriate intervention measures. However, due to the randomness of epileptic seizures, improving the predictive ability of the model remains a challenging task. Summary of the Invention

[0003] The purpose of this invention is to provide an epilepsy grading prediction model based on an improved CNN fusion model to solve the above-mentioned technical problems.

[0004] To address the aforementioned technical problems, the present invention provides a specific technical solution for an epilepsy grading prediction model based on an improved CNN fusion model, as follows:

[0005] An epilepsy grading prediction model based on an improved CNN fusion model includes an EEG signal segmentation module, a synchronous compressed wavelet transform module, a data processing module, a model fusion module, and a prediction probability processing module. The EEG signal segmentation module is used to segment the EEG signal into different time periods. The synchronous compressed wavelet transform module processes the EEG signal using synchronous compressed wavelet transform. The data processing module includes an MLF-CNN model and a semi-dilated convolutional network model for data processing. The model fusion module is used to fuse the MLF-CNN model and the semi-dilated convolutional network model. The prediction probability processing module uses a self-offset post-processing method to process the prediction probability.

[0006] Furthermore, the EEG signal segmentation module defines the pre-seizure period as 15 minutes before the epileptic seizure, and the interictal period data is taken from the period between four hours after the last epileptic seizure and four hours before the current seizure.

[0007] Furthermore, the synchronous compressed wavelet transform module processes the EEG signal using synchronous compressed wavelet transform, extracts the 0-50Hz frequency band from the signal, and maps it to the frequency domain using synchronous compressed wavelet transform to generate a time-frequency graph, which is used as the input to the prediction probability processing module.

[0008] Furthermore, the MLF-CNN model reduces the convolutional layers in the fourth and fifth blocks of VGG16 to limit the receptive field of the top layer. Then, it uses 1×1 convolutional kernels to process the results of each convolutional layer into vectors of the same dimension. The vectors obtained from each block are then merged and deconvolutioned to enlarge the feature size to obtain five feature maps of the same size. Finally, the five enlarged feature maps are concatenated and fused, and dimensionality reduction and softmax classification are performed. These five feature maps come from five different modules and contain feature information of different dimensions.

[0009] Furthermore, the MLF-CNN model uses synchronous compressed wavelet transform to map EEG data from the CHB-MIT dataset to the frequency domain.

[0010] Furthermore, the semi-dilated convolutional network model is used to input the time-frequency map obtained by the synchronous compressed wavelet transform module into the semi-dilated convolutional network model. The data passes through a network structure consisting of 6 sequential layers: 3 convolutional layers and 3 pooling layers. The convolutional layers consist of 5 parallel semi-dilated convolutional layers with dilation rates of 1, 2, 4, 8, and 16, respectively. Each convolutional layer is followed by a max-pooling layer with a pooling stride of 2. The number of convolutional kernels in the three convolutional layers are 64, 128, and 256, respectively. Finally, the fully connected layer module learns the extracted features to achieve prediction and classification. Further, the model fusion module uses post-model fusion, employing an MLF-CNN network in parallel with the constructed semi-dilated convolutional network to form a novel prediction model. This model is converted into a time-frequency map by the synchronous compressed wavelet transform and used as input to the new model. The two branch networks simultaneously extract feature information from the time-frequency map and generate prediction probabilities through a softmax layer.

[0011] Furthermore, the prediction probability processing module uses the prediction probabilities generated by the two models in the data processing module as input to the self-offset post-processing method. The prediction probability processing module includes a module that executes the following program steps: Let P... n p is the predicted probability vector generated by the nth branch network. ni Let P be the probability vector. n The i-th probability value in the model, where i ≤ k, and k is the number of categories the model classifies.

[0012] P n =[p n1 ,p n2 ,…,p nk ], 1≤i≤k (1)

[0013] At the same time, let the maximum value in the probability vector of each branch network be denoted as . have

[0014] p nmax =maxpni ,1≤i≤k (2)

[0015] A self-offset post-processing method is used to calculate a weighted vector based on the probability vector of each branch model.

[0016] w = [w1, ..., w n ], where w i The calculation method is as follows:

[0017]

[0018] Let P be the probability output matrix composed of the probability vectors of n branch models, then the final classification prediction q after fusing multiple branch models is:

[0019] q=w×P (4)

[0020] By using the self-offset post-processing method, the probability vectors of each branch model are weighted, the most confident branch model is highlighted, and the final prediction vector is generated.

[0021] The epilepsy grading prediction model based on an improved CNN fusion model of the present invention has the following advantages: The self-offset fusion model used in this invention achieves better overall classification results for synchronous compressed wavelet transform time-frequency maps than the model using MLF-CNN. Similar to the performance comparison with MLF-CNN, the semi-dilated convolution model also shows a certain gap in performance compared to the self-offset fusion model. Therefore, the self-offset fusion model offers a significant performance improvement compared to individual models. Furthermore, it can be used on clinical datasets and possesses generalization ability. Attached Figure Description

[0022] Figure 1 A detailed network structure diagram of MLF-CNN.

[0023] Figure 2 This is a diagram of the structure of a semi-dilated convolutional network model.

[0024] Figure 3 This is a structural diagram of an epilepsy grading prediction model based on an improved CNN fusion model. Detailed Implementation

[0025] To better understand the purpose, structure, and function of this invention, the following detailed description of an epilepsy grading prediction model based on an improved CNN fusion model is provided in conjunction with the accompanying drawings.

[0026] The present invention provides an epilepsy grading prediction model based on an improved CNN fusion model, comprising an EEG signal segmentation module, a synchronous compressed wavelet transform module, a data processing module, a model fusion module, and a prediction probability processing module.

[0027] The EEG signal segmentation module is used to divide the EEG signal into different periods. The preic phase is 15 minutes before the epileptic seizure, and the interictal phase data is taken from four hours after the last epileptic seizure to four hours before the current seizure, in order to reduce the interference between data.

[0028] The synchronous compressed wavelet transform module processes EEG signals using synchronous compressed wavelet transform, then extracts a 0-50Hz frequency band from the signal and maps it to the frequency domain using synchronous compressed wavelet transform to generate a time-frequency plot. This time-frequency plot is used as input to the prediction probability processing module. Synchronous compressed wavelet transform is a time-frequency domain rearrangement algorithm developed based on wavelet transform. This algorithm can better describe the characteristics of the time spectrum than continuous wavelet transform and can clearly represent sudden energy discharges.

[0029] The data processing module includes the MLF-CNN model and the semi-dilated convolutional network model.

[0030] like Figure 1 As shown, the MLF-CNN model reduces the number of convolutional layers in the fourth and fifth blocks of VGG16 to limit the receptive field of the top layer. Then, it uses 1×1 convolutional kernels to process the results of each convolutional layer into vectors of the same dimension. The vectors obtained from each block are then merged and deconvolved to amplify the feature size, resulting in five feature maps of the same size. Finally, the five amplified feature maps are concatenated and fused, and dimensionality reduction and softmax classification are performed. These five feature maps come from five different modules and contain feature information of different dimensions. This is our MLF-CNN, a novel seizure prediction model that utilizes synchronous compressed wavelet transform to map EEG data from the CHB-MIT dataset to the frequency domain. In this algorithm, synchronous compressed wavelet transform can well represent the energy transformation before a seizure in the time-frequency plane. MLF-CNN automatically extracts multi-level features from the time-frequency map, achieving better classification of epilepsy data and effectively predicting seizures.

[0031] The MLF-CNN model takes the time-frequency map obtained from the synchronous compressed wavelet transform module as input and processes it into vectors of the same dimension using 1×1 convolutional kernels. Then, the vectors from each block are merged and deconvolutioned to enlarge the feature size, resulting in five feature maps of the same size. Finally, the five enlarged feature maps are concatenated and fused, and dimensionality reduction and softmax classification are applied to obtain the desired prediction probability.

[0032] Dilated convolution, also known as attenuated convolution, was initially proposed to address semantic segmentation tasks. Later, with the development of deep learning, it was widely applied to image segmentation and classification. Based on dilated convolution, we propose a novel convolution strategy called semi-dilated convolution. It can grow exponentially over a one-dimensional receptive domain. For example... Figure 2 As shown, in this invention, we construct a semi-dilated convolutional network model that can exponentially expand the receptive domain along the time axis in the time-frequency graph while maintaining a regular receptive domain along the frequency axis. The network structure consists of six sequential layers: three convolutional layers and three pooling layers. The convolutional layers consist of five parallel semi-dilated convolutional layers with dilation rates of 1, 2, 4, 8, and 16. Each convolutional layer is followed by a max-pooling layer with a pooling stride of 2. The number of convolutional kernels in the three convolutional layers are 64, 128, and 256, respectively. Finally, the fully connected layer module learns the extracted features to achieve prediction and classification.

[0033] A semi-dilated convolutional network model is used to input the time-frequency map obtained from the synchronous compressed wavelet transform module. The data passes through a network structure consisting of six sequential layers: three convolutional layers and three pooling layers. The convolutional layers consist of five parallel semi-dilated convolutional layers with dilation rates of 1, 2, 4, 8, and 16. Each convolutional layer is followed by a max-pooling layer with a stride of 2. The number of kernels in the three convolutional layers are 64, 128, and 256, respectively. Finally, the fully connected layer learns the extracted features to achieve prediction and classification, ultimately yielding the desired prediction probability.

[0034] Model fusion is divided into two types: pre-model fusion and post-model fusion. Pre-model fusion involves fusing and concatenating the processed feature data before model training. Post-model fusion, on the other hand, first uses multiple models to extract features, then fuses and concatenates all the extracted features to generate new features for further processing. The model fusion module of this invention uses post-model fusion. We use an MLF-CNN network and a constructed semi-dilated convolutional network in parallel to form a novel prediction model, which is then converted into a time-frequency image using synchronous compressed wavelet transform as input to the new model. The two branch networks simultaneously extract feature information from the time-frequency image and generate prediction probabilities through a softmax layer.

[0035] The prediction probability processing module uses a self-shift post-processing method to process the prediction probabilities. This module takes the prediction probabilities generated by the two models in the data processing module as input to our proposed self-shift post-processing method. Let P... n p is the predicted probability vector generated by the nth branch network. ni Let P be the probability vector. n The i-th probability value in the model, where i ≤ k, and k is the number of categories in the model classification.

[0036] Right now

[0037] P n =[p n1 ,p n2 ,…,p nk ], 1≤i≤k (1)

[0038] At the same time, let the maximum value in the probability vector of each branch network be denoted as . have

[0039] p nmax =maxp ni ,1≤i≤k (2)

[0040] Since the maximum value in the probability vector of each branch model can essentially represent the classification result, and the impact of each model on the final result should be considered during fusion, we propose a self-offset post-processing method. This method can calculate a weighted vector w = [w1, ..., w] based on the probability vector of each branch model. n ], where w i The calculation method is as follows:

[0041]

[0042] Let P be the probability output matrix composed of the probability vectors of n branch models, then the final classification prediction q after fusing multiple branch models is:

[0043] q=w×P (4)

[0044] By employing a self-biased post-processing method, the probability vectors of each branch model are weighted, highlighting the most confident branch model and generating the final prediction vector. The fusion of the two models helps extract different feature maps from interictal and preictal EEG data, improving model performance and generalization ability.

[0045] like Figure 3 As shown, the fused model, combined with the self-offset post-processing method, constitutes an automatic weighted fusion self-offset fusion model, namely, an epilepsy grading prediction model based on an improved CNN fusion model.

[0046] Experimental procedure:

[0047] 1. Binary seizure prediction experiment

[0048] The epilepsy EEG dataset used in this experiment was the CHB-MIT dataset, publicly available from Boston Children's Hospital in the United States. This dataset is currently the most widely used public dataset in the field of epilepsy, containing EEG recordings from 24 cases of refractory epilepsy. Signals from all cases were sampled at a rate of 256 samples per second with a resolution of 16 bits. In most cases, the files contain 23 EEG signals, with a few files containing 24 or 26 signals. These recordings were recorded using the international 10-20 standard EEG electrode placement and naming system. The dataset also provides the start and end times of each seizure for each patient.

[0049] To measure the accuracy of the experimental results, we use accuracy (ACC), sensitivity (SEN), and specificity (SPE) as performance metrics.

[0050] 1) First, we compared the self-offset fusion model with the single MLF-CNN model. The data in the table shows that multiple patients achieved 100% accuracy, sensitivity, and specificity. Furthermore, the self-offset fusion model generally outperformed the MLF-CNN model in classifying synchronous compressed wavelet transform time-frequency maps. For example, in MLF-CNN, the accuracy, sensitivity, and specificity of chb05 patients were 84.58%, 86.19%, and 83.11%, respectively, which improved to 98.50%, 100.00%, and 96.94% in the self-offset fusion model; chb06 also saw improvements from 80.74%, 78.75%, and 83.04% in accuracy, sensitivity, and specificity to 94.68%, 92.48%, and 96.98%; chb10 also achieved significant improvements, from 79.10%, 86.56%, and 74.17% to 97.13%, 96.28%, and 99.81%. Furthermore, chb14, chb15, and chb21 also showed considerable improvements. The results are shown in Table 1.

[0051] Table 1 Comparison of the performance of SWT time-frequency plots under MLF-CNN and self-offset fusion models.

[0052]

[0053]

[0054] 2) To further investigate the performance of the self-offset fusion model, we conducted a separate experiment using a semi-dilated convolutional network branch. The purpose of this experiment was to verify that the self-offset fusion model is more effective than the single model. The results of this experiment are shown in Table 2. Similar to the performance comparison with MLF-CNN, the performance of the semi-dilated convolutional model also shows a certain gap compared to the self-offset fusion model. For example, in the self-offset fusion model, the sensitivity of chb03 is 98.80%, while that of the semi-dilated convolutional network is only 71.84%. Similarly, the accuracy, sensitivity, and specificity of chb05 are 85.91%, 71.36%, and 88.72%, respectively, which are significantly lower than the 98.50%, 100.00%, and 96.94% of the self-offset fusion model. The performance in chb10 patients was also insufficient. Compared to the accuracy, sensitivity, and specificity of 97.13%, 96.28%, and 97.97% achieved by the self-offset fusion model, the accuracy, sensitivity, and specificity of the semi-dilated convolutional network were 89.58%, 78.81%, and 89.00%, respectively. Furthermore, the self-offset fusion model showed improved performance in other patients.

[0055] Table 2 Comparison of SWT time-frequency plot performance under the semi-dilated convolution model and the self-offset fusion model.

[0056]

[0057]

[0058] 3) To more directly demonstrate the performance improvement of the self-offset fusion model, we compared the average performance of MLF-CNN, the semi-dilated convolutional model, and the self-offset fusion model on the CHB-MIT dataset, as shown in Table 3. As mentioned earlier, among accuracy, sensitivity, and specificity, sensitivity is the most important indicator, reflecting the probability of not missing diagnoses. The table shows that the sensitivity of MLF-CNN and the semi-dilated convolutional network are 93.90% and 93.76%, respectively, with almost no difference. Accuracy and specificity are 92.74% and 93.83%, and 91.82% and 94.77%, respectively. In the self-offset fusion model, accuracy, sensitivity, and specificity are 97.50%, 97.76%, and 98.19%, respectively. Compared to MLF-CNN, these represent improvements of 4.76%, 3.86%, and 6.37%, respectively. Compared to the semi-dilated convolutional model, these represent improvements of 3.67%, 3.86%, and 3.42%, respectively. The above comparison shows that the self-offset fusion model has a significant performance improvement compared to the individual branch models.

[0059] Table 3 Comparison of the performance of each branch model and the self-offset fusion model.

[0060] Predictive Model ACC SEN SPE MLF-CNN 92.74 93.90 91.82 Semi-dilated convolutional networks 93.83 93.76 94.77 Self-offset fusion model 97.50 97.76 98.1

[0061] 2. Multi-level prediction experiment

[0062] The ability to accurately predict epileptic seizures is of great significance for patients with epilepsy. To further improve accuracy and reduce patient stress and anxiety without hindering timely intervention by healthcare professionals, we further divide the preictal phase into three stages: Pre-I, Pre-II, and Pre-III. These correspond to the first five minutes, the second five minutes, and the last five minutes of the original preictal phase, respectively.

[0063] In the three pre-seizure phases, we only use PreⅠ and PreⅢ. Predicting PreⅠ is used to send timely warnings to medical staff so that intervention can be taken; predicting PreⅢ allows patients to go to a safe place before a seizure to avoid accidents during the seizure. PreⅠ, PreⅢ, and the interictal period constitute the three-class classification problem in this grading prediction. However, for EEG signals, the closer the categories are, the more similar the characteristics of these EEG signals tend to be, making classification more difficult.

[0064] We tested the proposed self-offset fusion model on the aforementioned epileptic seizure classification prediction. To verify the effectiveness of the self-offset fusion model on this three-class classification problem, experiments were also conducted using MLF-CNN and semi-dilated convolutional models. All experiments followed the same setup as previous experiments, using the same preprocessing and synchronous compressed wavelet transform to convert the raw EEG data into time-frequency maps, which were then used as input to the prediction models. The maximum number of iterations during training for all models was uniformly set to 50, using the cross-entropy loss function and SGD optimizer, with an initial learning rate of 0.01.

[0065] Table 4 shows a comparison of the performance of MLF-CNN and the self-offset fusion model in predicting seizure grading. Currently, there are few studies on multi-class tasks for epilepsy detection and seizure prediction, and most use the accuracy of each class as the evaluation metric. Therefore, the data in the table represent the accuracy of each seizure stage under different models. As can be seen from the table, although the accuracy for the interictal period remains at a high level, the accuracy of PerⅠ and PerⅢ fluctuates significantly. Overall, the accuracy of multi-level prediction is significantly lower than that of binary seizure prediction. In general, only a few patients, such as chb04 and chb12, experience classification confusion in the preictal stage; most patients can still achieve multi-level seizure prediction using the self-offset fusion model.

[0066] Table 4 Comparison of SWT time-frequency plot performance under MLF-CNN and self-offset fusion models.

[0067]

[0068]

[0069] Table 5 compares the performance of the semi-dilated convolutional model and the self-offset fusion model in epilepsy patients. Similar to the results in Table 4, the self-offset fusion model significantly outperforms the semi-dilated convolutional model. The semi-dilated convolutional model is also affected by the relatively small differences in EEG characteristics among similar categories, and there are instances of inaccurate classification in the pre-seizure phase in some patients.

[0070] Table 4-5 Comparison of the performance of 1s SWT time-frequency plots under the semi-dilated convolution model and the self-offset fusion model

[0071]

[0072]

[0073] In summary, the self-offset fusion model is more effective than MLF-CNN and semi-dilated convolution models in achieving multi-level prediction of epileptic seizures.

[0074] Therefore, this invention offers the following advantages: it performs well in both binary seizure prediction experiments and tri-class seizure multi-level prediction experiments and clinical dataset validation experiments. Experimental results show that the self-offset fusion model significantly improves performance compared to MLF-CNN and semi-dilated convolutional models, and it can be used on clinical datasets, demonstrating generalization ability.

[0075] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An epilepsy grading prediction model based on an improved CNN fusion model, characterized in that, It includes a brainwave signal segmentation module, a synchronous compressed wavelet transform module, a data processing module, a model fusion module, and a prediction probability processing module; The EEG signal segmentation module is used to segment the EEG signal into different periods; the synchronous compressed wavelet transform module processes the EEG signal using synchronous compressed wavelet transform; the data processing module includes an MLF-CNN model and a semi-dilated convolutional network model for data processing; the model fusion module is used to fuse the MLF-CNN model and the semi-dilated convolutional network model; the prediction probability processing module uses a self-offset postprocessing method to process the prediction probability; the EEG signal segmentation module defines the preictal period as 15 minutes before the epileptic seizure, and the interictal period data is taken from four hours after the last epileptic seizure to four hours before the current seizure; the MLF-CNN model reduces the convolutional layers in the fourth and fifth blocks of VGG16 to limit the top receptive field, and then uses 1×1 The convolutional kernel processes the results of each convolutional layer into vectors of the same dimension. Then, the vectors obtained from each block are merged, and deconvolution is performed to amplify the feature size, resulting in five feature maps of the same size. Finally, the five amplified feature maps are concatenated and fused, and dimensionality reduction and softmax classification are applied. These five feature maps come from five different modules and contain feature information of different dimensions. The MLF-CNN model uses synchronous compressed wavelet transform to map EEG data from the CHB-MIT dataset to the frequency domain. The semi-dilated convolutional network model is used to input the time-frequency map obtained from the synchronous compressed wavelet transform module into the semi-dilated convolutional network model. The data passes through a network structure consisting of six sequential layers: three convolutional layers and three pooling layers. The convolutional layers consist of five parallel semi-dilated convolutional layers with dilation rates of 1, 2, 4, 8, and 16. Each convolutional layer is followed by a max-pooling layer with a stride of 2. The number of convolutional kernels in the three convolutional layers are 64 and 128, respectively. And 256, finally the fully connected layer module learns the extracted features to achieve prediction and classification; the model fusion module uses model post-fusion, using an MLF-CNN network and a constructed semi-dilated convolution to form a new prediction model, which is converted into a time-frequency map by synchronous compressed wavelet transform as the input of the new model. The two branch networks simultaneously extract feature information from the time-frequency map and generate prediction probabilities through a softmax layer; the prediction probability processing module takes the prediction probabilities generated by the two models in the data processing module as the input of the self-offset post-processing method. The prediction probability processing module includes a module that executes the following program steps: Let It is the predicted probability vector generated by the nth branch network. Probability vector The i-th probability value in the model, where i ⩽k, and k is the number of categories in the model classification. （1） At the same time, let the maximum value in the probability vector of each branch network be denoted as . ,have （2） A self-offset post-processing method is used to calculate a weighted vector based on the probability vector of each branch model. ,in The calculation method is as follows: （3） Let P be the probability output matrix composed of the probability vectors of n branch models, then the final classification prediction q after fusing multiple branch models is: q = w×P （4） By using the self-offset post-processing method, the probability vectors of each branch model are weighted, the most confident branch model is highlighted, and the final prediction vector is generated.

2. The epilepsy grading prediction model according to claim 1, characterized in that, The synchronous compressed wavelet transform module processes the EEG signal using synchronous compressed wavelet transform, extracts the 0-50Hz frequency band from the signal, and maps it to the frequency domain using synchronous compressed wavelet transform to generate a time-frequency graph, which is used as the input to the prediction probability processing module.

Images