An ESES-related epileptic encephalopathy prognosis evaluation model and its training and usage methods

Abstract

The present invention belongs to the technical field of auxiliary diagnosis of epileptic encephalopathy, and specifically relates to an automatic quantification and prognosis assessment model for ESES-related epileptic encephalopathy and its training and usage methods. The encephalopathy automatic quantification and prognosis assessment model of the present invention includes a brain network feature extraction module, an electroencephalogram (EEG) depth feature extraction module, and an identification and diagnosis module; the brain network feature extraction module includes a calculation module, a graph convolutional neural network, and an attention mechanism network, which obtain brain network graph theory features and atlas features, and fuse the two to obtain information on the state transition between data segments; the EEG depth feature extraction module is used to extract EEG depth features; the identification and diagnosis module is used to fuse the brain network features extracted by the EEG depth feature extraction module and the EEG depth features extracted by the EEG depth feature extraction module, and perform the identification of ESES characteristic waves and the assessment of the prognosis level. The present invention not only realizes the automatic quantification of ESES, but also realizes the prognosis assessment of ESES-related epileptic encephalopathy, thereby greatly improving the diagnosis and treatment efficiency of ESES.

Description

Technical Field

[0001] The present invention belongs to the technical field of auxiliary diagnosis of epileptic encephalopathy, and particularly relates to a prognosis evaluation model for epilepsy encephalopathy related to ESES and its training and use methods. Background Art

[0002] Electrical Status Epilepticus during Sleep (ESES) is an epileptic symptom with persistent spike and slow waves during slow-wave sleep. Its characteristic is epileptiform activity in a strongly activated state during sleep, which consists of a series of continuous abnormal spikes and slow waves in the electroencephalogram. It has a relatively high frequency during the seizure period of pediatric epilepsy, but generally shows idiopathic, cryptogenic, and symptomatic, and the abnormal discharge mostly occurs during the sleep state, so it is easily overlooked, difficult to be clinically detected, and delays treatment and management. Although the clinical symptoms and electroencephalogram of most children tend to improve around puberty, the neuropsychological damage caused by ESES is irreversible. Therefore, early diagnosis and prognosis evaluation of epilepsy encephalopathy related to ESES are crucial.

[0003] Clinically, the slow-wave sleep spike and slow wave index (SWI) based on long-term electroencephalogram analysis is an important criterion for diagnosing and treating epilepsy syndromes related to ESES, and is also an important indicator for evaluating the prognosis of children. The general basis for clinical diagnosis is to quantify the epileptic activity related to ESES and estimate the SWI index by electroencephalogram recording and combining with manual identification of epileptiform waves. However, the electroencephalogram recording time related to ESES often lasts for several hours to dozens of hours, and manual analysis is time-consuming and laborious, and it takes a long time (usually more than 1 hour) from the completion of the electroencephalogram examination to obtaining the results; manual estimation is not accurate enough, and there are differences in the analysis results among different doctors. These factors are not conducive to the timely diagnosis and effective treatment of children. At present, most studies often identify and quantify ESES by implementing spike detection methods, extract morphological features, time domain or frequency domain features from the signals, and then use template matching methods or machine learning methods for classification. Its limitation is that the feature extraction step depends on prior knowledge, and requires a large amount of feature engineering and complex parameter tuning. In addition, most methods only implement spike detection and ignore slow wave detection, and there are difficulties in accurately identifying the characteristic waveform of ESES. In addition to the identification of the ESES characteristic waveform, there is currently no research on the prognosis evaluation of epilepsy encephalopathy related to ESES.

[0004] In summary, the interpretation of ESES epileptiform waves has difficulties such as high professional requirements and large differences in manual interpretation. The methods proposed in existing research have drawbacks such as ignoring overlapping slow waves and inaccurate identification. Therefore, the research on the automatic quantification of ESES and the prognosis evaluation system for ESES-related epileptic encephalopathy can significantly improve the diagnosis and treatment efficiency of ESES. Summary of the Invention

[0005] Aiming at the deficiencies of the above-mentioned existing technologies, the purpose of the present invention is to propose a prognosis evaluation model for ESES-related epileptic encephalopathy that integrates electroencephalogram depth features and brain network features, as well as its training and use methods.

[0006] The prognosis evaluation model for ESES-related epileptic encephalopathy provided by the present invention that integrates electroencephalogram depth features and brain network features includes: a brain network feature extraction module, an electroencephalogram depth feature extraction module, and an identification and diagnosis module; among them:

[0007] The brain network feature extraction module is used to extract brain network features, including the graph theory features of the brain network calculated by the calculation module and the atlas features extracted by the graph convolutional neural network. The two are fused to extract the state transition information between data segments through the attention mechanism network; among them:

[0008] The calculation module calculates the graph theory features of the brain network based on the originally acquired electroencephalogram signals. The graph theory features of the brain network include node degree, node betweenness centrality, node strength, clustering coefficient, weighted characteristic path length, weighted global efficiency, and local efficiency, etc.

[0009] The graph convolutional neural network is used to extract the atlas features of the brain network, that is, to extract the connectivity between the nodes of the graph signal; at the same time, the original electroencephalogram signal features of each node of the graph signal are extracted as the node information of the graph signal for graph convolution operations.

[0010] Furthermore, the graph convolutional neural network specifically includes a graph signal generation module, which includes a node feature extraction layer and a topological structure extraction layer to generate a graph signal; this graph convolutional neural network includes 1 graph convolutional layer and 1 pooling layer; the graph convolutional layer is used to aggregate the information between levels, and the pooling layer is used to aggregate the information of each node within the layer. Finally, the atlas features are output through a readout layer.

[0011] The attention mechanism network is used to extract the effective information of the conversion between the features of each data segment, such as the effective information of the conversion of the brain network features between the appearance and disappearance of ESES-like waves.

[0012] Further, the attention mechanism network specifically includes a feature fusion module and an attention mechanism module; the feature fusion module is used to fuse the brain network graph theory features and the atlas features obtained by graph convolution; the attention mechanism module is composed of an encoder and a decoder, and is used to map the sequence of data segment features into the type space of auxiliary diagnosis, and finally complete classification through a fully connected layer and an activation function.

[0013] The electroencephalogram (EEG) depth feature extraction module is used to extract EEG depth features and is composed of a dual-scale convolutional neural network and a recurrent neural network; where:

[0014] The dual-scale convolutional neural network is used to extract the morphological features of EEG signals;

[0015] Further, the dual-scale convolutional neural network specifically includes two groups of parallel feature extraction modules, each group consisting of 3 convolutional layers, 2 pooling layers, a batch normalization layer, and a non-linear activation layer. The convolutional layer is used to reduce the dimension and extract features, the pooling layer is used to further reduce the dimension and reduce the computational amount, the batch normalization layer is used to prevent the vanishing gradient, and the non-linear activation layer is used to map the feature transformation to a high dimension and extract the depth information. The first convolutional layer of the dual-scale convolutional neural network uses different strides and convolutional kernels to extract and fuse the morphological features of different frequencies from the EEG signals, thereby increasing the robustness of the model.

[0016] The recurrent neural network is used to extract the temporal features of EEG signals;

[0017] Further, the recurrent neural network specifically includes a long short-term memory network. Compared with ordinary recurrent neural networks, it adds a forget gate, an input gate, an output gate, and a memory unit. By controlling the three gates, the unit state in the temporal learning is changed and the information is passed to the next moment. While learning the temporal signal features of memory, the problem of long-term dependence, that is, the problem of the gradient disappearing after multi-stage propagation, is solved.

[0018] The recognition and diagnosis module is used to fuse the brain network features extracted by the EEG depth feature extraction module and the EEG depth features extracted by the EEG depth feature extraction module, and perform the recognition, automatic quantification of ESES characteristic waves, and prognosis evaluation of related epileptic encephalopathies.

[0019] Further, the recognition and diagnosis module is composed of a convolutional neural network, specifically including 3 convolutional layers, 2 pooling layers, and 2 non-linear activation layers.

[0020] For the above-mentioned prognosis evaluation model of ESES-related epileptic encephalopathy that constructs the fusion of EEG depth features and brain network features of the present invention, the training is carried out according to the following specific steps:

[0021] (1) Obtain the original electroencephalogram (EEG) signals of N children in the sleeping state, and perform ESES feature annotation on the EEG signals of each subject; the ESES feature annotation results of the original EEG signals are used as the gold standard for model performance verification; generally, it is required that N is greater than or equal to 20;

[0022] (2) Cut the original EEG signals obtained in step 1 and their associated annotations, and perform data balancing processing to meet the model data specifications and requirements, so that various training data samples are relatively balanced; specifically, cut the EEG signals into 30s segments, and whether each segment of the signal contains ESES feature waves is used as the classification basis, and the downsampling method is used to make the number of majority-class samples and minority-class samples equivalent;

[0023] (3) Resample and filter the standardized EEG signals obtained in step 2 to filter out background noise and various artifact signals, etc. The specific steps are as follows:

[0024] (3.1) Resample the annotated EEG signals to the sampling rate required by the model;

[0025] (3.2) Use a Butterworth band-pass filter to filter out background noise;

[0026] (3.3) Use a finite impulse response (FIR) filter to remove artifacts caused by the subject's sweat, movement, and electrode interference, etc.;

[0027] (4) Train an encephalopathy prognosis evaluation model using the preprocessed training data; the input of the model is the preprocessed EEG signal, and the output is whether it contains ESES feature waves, the start and end positions of the ESES feature waveform, and the prognosis level of the related epileptic encephalopathy.

[0028] First, transform the EEG signal into the form of a graph signal. The EEG signal is input into the EEG deep feature extraction module, and the graph signal form of the EEG signal is input into the brain network feature extraction module. The features obtained by the two modules are concatenated and merged together, and then input into the recognition and diagnosis module. The output of this module is divided into two parts. Among them, the classification problem is whether it contains ESES feature waves, and the regression problem is the start and end positions of the feature waves. Two loss functions, cross-entropy loss and mean squared error loss, are superimposed for training. After multiple rounds of iteration, a trained encephalopathy prognosis evaluation model is obtained.

[0029] The usage method of the trained encephalopathy prognosis evaluation model is as follows:

[0030] (1) Perform the same preprocessing on the newly collected original EEG signal as the training data to obtain the graph signal form required by the model, and input it into the brain network feature extraction module. Input the EEG signal into the EEG deep feature extraction module;

[0031] (2)Fuse the obtained brain network features and EEG depth features;

[0032] (3)Input the obtained fused features into the recognition and diagnosis module to obtain the recognition result, and perform automatic quantification and prognosis assessment.

[0033] In the present invention, the annotation includes: the starting position of the ESES characteristic wave and the ending position of the characteristic wave.

[0034] In the present invention, the filter is a Butterworth band-pass filter and a finite impulse response (FIR) filter.

[0035] In the present invention, the resampling is to perform upsampling or downsampling on the original signal sequence to achieve a specific sampling frequency.

[0036] The features and beneficial effects of the present invention are as follows:

[0037] By extracting the topological structure and features of the original EEG signals, the present invention can greatly utilize various types of information contained in the EEG signals and significantly reduce the data volume and computational complexity in the subsequent processing while maintaining high robustness.

[0038] By extracting the morphological features and temporal features of the original EEG signals, the present invention can obtain the spatio-temporal connection of the morphological changes before and after the EEG signals and can reduce the data dimension and computational complexity.

[0039] The present invention extracts the brain network from the EEG signals, utilizes both the overall features of the changes in the original EEG signals and the local features of each node, and also combines the connectivity between each node and the information on the forward and backward transitions between different data segments. At the same time, fusing the brain network features and EEG depth features makes the signal features more extensive, ensures the robustness of the auxiliary diagnosis model, and is more conducive to the interpretation of EEG signals and the progress of cognitive research. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic flowchart of a method for constructing an automatic quantification and prognosis assessment model for ESES-related epileptic encephalopathy based on fused brain network features and EEG depth features. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0042] The auxiliary diagnosis method for epileptic encephalopathy related to electrical status epilepticus during sleep in children proposed by the present invention has the overall process as Figure 1 shown, and includes the following steps:

[0043] 1. Obtain at least 20 original electroencephalogram (EEG) signals during the sleep process of children. The signal duration of each child is at least about 2 hours (at least one sleep cycle is required), and an expert annotates the symptoms of the EEG signal of each subject to obtain the training data with annotations. In this embodiment, the original EEG signals collected by the medical EEG acquisition device do not contain the position labels of characteristic waves. To enable the proposed model to learn the characteristics, an expert needs to manually mark the start and end positions of the corresponding characteristic waves in the signals. The marked data will be used to train the proposed deep learning model.

[0044] 2. Preprocess the obtained data with annotations, mainly including: data cutting, filtering, resampling, data balancing, etc.

[0045] (1) Data cutting: Cut the original EEG signal to generate standardized data segments;

[0046] (2) Filtering: Use a Butterworth band-pass filter to remove the large background noise in the original EEG signal to obtain a relatively pure EEG signal; use a FIR filter to remove artifacts;

[0047] (3) Resampling: Resample the EEG data to 100 Hz.

[0048] 3. Use the preprocessed training data to train the auxiliary diagnosis model.

[0049] 4. Use the test set samples to test the performance of the trained model. The model performance is defined as the accuracy compared with the manual annotation.

[0050] Compared with the existing ESES auxiliary diagnosis methods, the method proposed by the present invention can improve the utilization rate of the information contained in the EEG signal, greatly reduce the data volume and computational complexity, and maintain the robustness of the model used among different populations on the premise of effectively realizing the auxiliary diagnosis. The present invention adopts the deep learning model method, which can greatly reduce the time and workload of manual identification.

Claims

1. An ESES-related epileptic encephalopathy prognosis evaluation model, characterized in that, Including: A brain network feature extraction module, an EEG depth feature extraction module, and an identification and diagnosis module; wherein: The brain network feature extraction module includes a calculation module, a graph convolutional neural network, and an attention mechanism network, wherein: The calculation module calculates graph theory features of the brain network based on the originally acquired EEG signals, and the graph theory features of the brain network include node degree, node betweenness centrality, node strength, clustering coefficient, weighted characteristic path length, weighted global efficiency, and local efficiency; The graph convolutional neural network is used to extract atlas features of the brain network, that is, to extract the connectivity between nodes of the graph signal; at the same time, it extracts the original EEG signal features of each node of the graph signal as node information of the graph signal for graph convolution operations; The attention mechanism network is used to fuse the graph theory features and atlas features of the brain network, and extract the information of state transitions between data segments, including the effective information of brain network feature transitions between the appearance and disappearance of ESES characteristic waves; The EEG depth feature extraction module is used to extract EEG depth features, which consists of a dual-scale convolutional neural network and a recurrent neural network; wherein: The dual-scale convolutional neural network is used to extract the morphological features of EEG signals; The recurrent neural network is used to extract the temporal features of EEG signals; The identification and diagnosis module is used to fuse the brain network features extracted by the brain network feature extraction module and the EEG depth features extracted by the EEG depth feature extraction module, and perform automatic quantification of ESES characteristic waves and prognosis evaluation of related epileptic encephalopathies; In the brain network feature extraction module: The graph convolutional neural network specifically includes a graph signal generation module, which contains a node feature extraction layer and a topological structure extraction layer to generate a graph signal; this graph convolutional neural network includes 1 graph convolutional layer and 1 pooling layer; the graph convolutional layer is used to aggregate information between levels, and the pooling layer is used to aggregate information of each node within the layer. Finally, the atlas features are output through a readout layer; The attention mechanism network specifically includes a feature fusion module and an attention mechanism module; the feature fusion module is used to fuse the graph theory features of the brain network and the atlas features obtained by graph convolution; the attention mechanism module consists of an encoder and a decoder, and is used to map the sequence of data segment features into the type space of auxiliary diagnosis and output.

2. The prognosis evaluation model for ESES-related epileptic encephalopathy according to claim 1, wherein In the EEG depth feature extraction module: The dual-scale convolutional neural network specifically includes two groups of parallel feature extraction modules, each group consisting of 3 convolutional layers, 2 pooling layers, a batch normalization layer, and a non-linear activation layer; the convolutional layer is used to reduce the dimension and extract features, the pooling layer is used to further reduce the dimension and reduce the computational amount, the batch normalization layer is used to prevent gradient disappearance, and the non-linear activation layer is used to map the feature transformation to a high dimension to extract depth information; the first convolutional layer of the dual-scale convolutional neural network uses different strides and convolutional kernels to extract morphological features of different frequencies from the EEG signals and fuse them, thereby increasing the robustness of the model; The recurrent neural network specifically includes a long short-term memory network. Compared with the recurrent neural network, a forget gate, an input gate, an output gate and a memory unit are added. By controlling the three gates, the unit state in sequential learning is changed and information is passed to the next moment, while solving the problem of long-term dependence, that is, the problem that the gradient disappears after multi-stage propagation.

3. The ESES-related epileptic encephalopathy prognosis evaluation model according to claim 2, wherein The recognition and diagnosis module is specifically composed of a convolutional neural network, specifically including 3 convolutional layers, 2 pooling layers, and 2 non-linear activation layers.

4. A training method for the prognosis evaluation model of ESES-related epileptic encephalopathy according to any one of claims 1-3, characterized in that The specific steps are as follows: (1) Obtain the original electroencephalogram (EEG) signals of N children during sleep, and perform ESES feature annotation on the EEG signals of each subject; The ESES feature annotation results of the original EEG signals are used as the gold standard for model performance verification; (2) Cut the original EEG signals obtained in step 1 and their associated feature annotations, and perform data balancing processing to meet the model data specifications and requirements, so that various training data samples are relatively balanced. Specifically, cut the EEG signals into 30s segments, and use whether each segment of the signal contains ESES feature waves as the classification basis, and use downsampling to make the number of samples in the majority class and the minority class equivalent; (3) Resample and filter the EEG signals processed in step 2, and filter out background noise and various artifact signals. The specific steps are as follows: (3.1) Resample the EEG signals with feature annotations to the sampling rate required by the model; (3.2) Use a Butterworth band-pass filter to filter out background noise; (3.3) Use a finite impulse response (FIR) filter to remove artifacts caused by the subject's sweat, movement, and electrode interference; (4) Train an ESES-related epileptic encephalopathy prognosis evaluation model using the preprocessed training data. The input of the model is the preprocessed EEG signal, and the output is whether it contains ESES feature waves, the start and end positions of the ESES feature waveform, and the prognosis grade of the related epileptic encephalopathy; First, transform the EEG signal into a graph signal form. The EEG signal is input into the EEG deep feature extraction module, and the graph signal form of the EEG signal is input into the brain network feature extraction module. The features obtained by the two modules are concatenated and merged together, and then input into the recognition and diagnosis module. The output of this module is divided into two parts. Among them, the classification problem is whether it contains ESES feature waves, and the regression problem is the start and end positions of the feature waves. Two loss functions, cross-entropy loss and mean squared error loss, are superimposed for training. After multiple rounds of iteration, a trained encephalopathy prognosis evaluation model is obtained.

5. A method for using an ESES-related epileptic encephalopathy prognosis evaluation model trained by claim 4, the specific steps are: (1) Perform the same preprocessing on the collected original EEG signals as the training data to obtain the graph signal form required by the model, and input it into the brain network feature extraction module, and input the EEG signal into the EEG deep feature extraction module; (2) Fuse the obtained brain network features and EEG deep features; (3) Input the obtained fusion features into the recognition and diagnosis module to obtain the recognition results, and conduct automatic quantification and prognostic evaluation.

Images