EEG report generation method combining hybrid neural network and analysis experience algorithm

Abstract

The invention discloses an EEG report generation method combining a hybrid neural network and an analysis experience algorithm. According to the method, through multi-scale time sequence deep convolution, symmetric channel convolution, cross convolution, whole-brain feature convolution, sub-window self-attention fusion, self-attention window fusion and a hybrid neural network of a classification network, a window highly associated with a classification result is selected, and a pathological judgment result is obtained. A window is described in combination with a mode of gradually enlarging a description area from waves to a channel to an area in an EEG analysis experience algorithm system, abnormal waves are found, and background activities are described; and finally, connecting the basic information, the window EEG signal graph, background activity description and abnormal wave description generated by the EEG analysis experience algorithm system and a classification result provided by the hybrid neural network in series to form an EEG report. A scheme for generating the EEG report by combining a deep learning method with generalization characteristics and an EEG analysis experience algorithm with experience knowledge is provided.

Description

Technical Field

[0001] The present invention belongs to the field of electroencephalogram information technology, and particularly relates to an EEG report generation method combining a hybrid neural network and an analysis experience algorithm. Background Art

[0002] EEG (electroencephalogram) report generation is one of the core steps in medical examination and diagnosis. As a physiological signal reflecting brain electrical activity, with its high time resolution and the property of presenting different waveform characteristics due to different brain regions, it can effectively display the physiological state of the subject's brain. However, currently, the generation of EEG reports faces problems such as difficult interpretation of EEG signals, weak anti-interference ability of EEG signals, and large differences in EEG signals among individuals.

[0003] Currently, the classification of EEG mainly relies on manual inspection. Among automated methods, there are mainly two categories: traditional machine learning and deep learning. Traditional machine learning methods rely on manually designed features and shallow classification methods (such as support vector machines, fuzzy search trees, random forests, etc.). However, this feature design is relatively difficult and there are also challenges in generalization. Deep learning methods use parameter-learnable methods to flexibly adjust parameters to adapt to capturing certain specific patterns in EEG, forming internal features, and classifying based on these features, thus having better generalization. For example, a convolutional neural network (CNN) extracts features in time and space and can capture specific patterns within and between channels. The interpretation of EEG in the manual method is to observe one window at a time and obtain the result by synthesizing all window information, and the interpretation results of EEG are structured and organized as the final EEG report. There is relatively little work on the generation of EEG reports, and the description of different EEG phenotypes is also affected by individual differences. Therefore, there is only an empirical standard but no "gold" standard in clinical practice. Fortunately, the literature [Liu Xiaoyan, "Clinical Electroencephalogram Training Course", 1st edition, Beijing, China: People's Medical Publishing House, November 2011, Chapters 6 to 8, pages 45 - 98] has quantitatively standardized the basic terms and phenotypes in EEG reports, which provides a standardized expression method and direction for the description of phenotypes in EEG and the generation of EEG reports.

[0004] Numerous studies are dedicated to solving the interference problem in EEG signals. For example, in the literature [Zhao Xin, Wu Jianxing, Wang Kun, et al. Review of Artifact Removal Algorithms for Electroencephalogram Signals [J / OL]. Signal Processing, 1 - 26], the applications of various techniques including regression methods, filtering methods, blind source separation, signal decomposition, and deep learning in removing EEG artifacts are discussed in detail. Artifact removal is not only a hot research field currently but also has broad development prospects, which is directly related to the accuracy of EEG data interpretation.

[0005] To address the challenge of significant inter-individual differences in EEG, enhancing the generalization ability of automated methods has become one of the key solutions. Transfer Learning, as a mature technology, has been widely applied in this area. For example, the literature [Qiu Yihui, Jiang Qiong, Wei Lingling, et al. Cross-subject EEG fatigue driving detection based on transfer learning [J]. Acta Scientiarum Naturalium Universitatis Nanchangensis, 2023, 47(04): 397-402.] proposed a method combining model transfer learning and improved Euclidean space alignment to enhance the classification performance of the model under the condition of a small amount of target domain data. In addition, in the face of the problem of generalization, strategies such as Domain Adaptation are also regarded as effective solutions.

[0006] Among the above methods, deep learning methods have the advantage of strong generalization ability but lack interpretability; empirical methods have the advantage of interpretable clinical implications but have insufficient generalization ability. The process from EEG interpretation to EEG report generation requires both rich clinical experience and high objectivity to handle the extremely diverse individual EEG characteristics, which is the core concern in the EEG report generation task. Summary of the Invention

[0007] In view of the above, the present invention provides an EEG report generation method combining a hybrid neural network and an analytical empirical algorithm, which can significantly improve the classification accuracy of EEG pathological classification across subjects.

[0008] An EEG report generation method combining a hybrid neural network and an analytical empirical algorithm includes the following steps: (1) Obtain an EEG dataset, where each EEG data includes an EEG signal collected from a subject with a duration of 20 - 30 minutes and its pathological label (normal or abnormal); (2) After preprocessing the EEG signal, divide the EEG dataset into a training set and a test set; (3) Construct a hybrid neural network model that fuses window features and self-attention, which includes: A window feature extraction network, with windows as the basic unit, extracts features from the data within each window of the EEG signal through multiple convolutional networks and self-attention networks; A self-attention window feature fusion network, which uses the self-attention mechanism to dynamically allocate attention weights to each window, and then fuses the features of each window to form a feature representation of the entire EEG signal; A classification network, which outputs the predicted probability that the EEG signal belongs to pathological normality based on the feature representation of the EEG signal; (4) Use the EEG data of the training set to train the above hybrid neural network model to obtain an EEG pathological classification model; (5) Construct an EEG analysis empirical algorithm system, which includes: A wave analysis module, which is used to analyze each turning point in the EEG signal, and dynamically match and generate single-wave features that conform to visual laws for each channel based on the empirical characteristics of forming brain waves; A channel analysis module, which is used to analyze the single-wave features of each channel and generate description data about the brain wave activities of a large-scale same frequency band; A region analysis module, which further analyzes the background activities between channels and on brain regions according to the description data of brain wave activities, and generates a description text about the background activities; (6) Input the EEG data of the test set into the EEG pathological classification model and the EEG analysis empirical algorithm system respectively, and combine the prediction results of the model about pathological normality or abnormality and the information calculated by the algorithm system to form an EEG report.

[0009] Further, the process of preprocessing the EEG signal in step (2) is as follows: First, select 21 channels specified by the international 10-20 standard lead system from the EEG signal, then perform band-pass filtering on the EEG signal at 0.5~70 Hz, then reduce the data sampling rate to 100 Hz, remove the data of the first minute of the EEG signal, extract a total of 20 minutes of data starting from the second minute of the EEG signal and multiply the data by 10 6 Thereby converting the data unit from volts to microvolts, and setting the data points less than -800 microvolts and greater than 800 microvolts in the EEG signal to -800 microvolts and 800 microvolts respectively.

[0010] Further, the window feature extraction network includes: A multi-scale temporal depth convolution module, which further subdivides the data in the EEG signal window into multiple sub-windows, is used to capture the patterns of different time courses that appear on each channel of the sub-window data, and can determine the weights of different time course patterns on the channel according to the characteristics of each channel; A symmetric channel convolution module, which is used to perform contrast fusion on the outputs of the multi-scale temporal depth convolution module on the channels symmetric about the left and right hemispheres; A cross-cross convolution module, which is used to perform convolution fusion on the output of the symmetric channel convolution module in the spatial domain along the horizontal and vertical directions to learn the local spatial information between channels; A whole-brain feature convolution module, which is used to perform convolution on the output of the cross-cross convolution module in the whole-brain space to obtain global spatial information; The sub-window self-attention fusion module is used to pass the output of the whole-brain feature convolution module through a self-attention layer, assign attention weights to each sub-window, and thus fuse to obtain window features.

[0011] Further, the multi-scale temporal depth convolution module performs one-dimensional depth convolution on the sub-window data of the EEG signal with 4 groups of different convolution kernel sizes in parallel, and then splices the outputs of these 4 groups of convolutions and recompresses them into 1 group of outputs through one-dimensional depth convolution; the symmetric channel convolution module stacks the outputs of the multi-scale temporal depth convolution module according to the left-right brain symmetric position features and then passes them through a two-dimensional depth convolution. Then, the output of this convolution is spliced with the features of the midline channel, and after passing through the activation function ELU (Exponential Linear Unit) and batch normalization processing, it is output; the cross-cross convolution module arranges the outputs of the symmetric channel convolution module two-dimensionally according to the brain region spatial positions, fills 0 at the four corners to form a 5-row 3-column matrix, and then extracts the local spatial fusion features of the rows and columns through two-dimensional depth convolution in the horizontal and vertical directions respectively. Then, the local spatial fusion features of the rows and columns are spliced with the output of the symmetric channel convolution module in the channel dimension and then output; the whole-brain feature convolution module passes the output of the cross-cross convolution module through one-dimensional convolution, activation function ELU, batch normalization, average pooling, one-dimensional convolution, activation function ELU, batch normalization, max pooling, and flattening operations in sequence to obtain sub-window features; the sub-window self-attention fusion module passes the output of the whole-brain feature convolution module through layer normalization L1, multi-head self-attention mechanism layer, layer normalization L2, and MLP (Multi-Layer Perceptron) in sequence to obtain window features, where the input of layer normalization L1 and the output of the multi-head self-attention mechanism layer are stacked as the input of layer normalization L2, and the input of layer normalization L2 and the output of the MLP are stacked to obtain window features.

[0012] Further, the self-attention window feature fusion network is sequentially composed of layer normalization L3, multi-head self-attention mechanism layer, layer normalization L4, and MLP from input to output, where the input of layer normalization L3 and the output of the multi-head self-attention mechanism layer are stacked as the input of layer normalization L4, and the input of layer normalization L4 and the output of the MLP are stacked as the final output of the self-attention window feature fusion network; the classification network uses a single-layer linear mapping layer without the activation function Softmax (Softmax).

[0013] Further, for the window data of any channel of the EEG signal, the wave analysis module converts it into discrete turning points (corresponding to possible wave peaks and wave valleys), and then selects and combines these turning points according to the shape conditions of the formed waves to generate a single-wave matching result in the form of a triple i s , i m ,i e , where i s , i m , i e respectively represent the starting point, midpoint (peak or trough) and ending point of a single wave. Finally, according to the triple i s , i m , i e , the single-wave characteristics including frequency, amplitude, morphology and duration are calculated through the wave attribute algorithm.

[0014] Furthermore, the channel analysis module divides the single waves of each channel into corresponding frequency bands according to the single-wave characteristics. For waves in the same frequency band that appear continuously over a certain scale and in time, they are regarded as one electroencephalogram activity. The descriptive data of such activities are statistically analyzed, including the start and end times, frequency range, amplitude range, and whether there are abnormal waveforms including sharp waves, spike waves and slow waves.

[0015] Furthermore, the region analysis module further analyzes the α rhythm in the posterior head region, the μ rhythm in the central region and the θ rhythm in the fronto-midline region according to the descriptive data of the electroencephalogram activity. For the α rhythm, the frequency range, amplitude range, symmetry and α index of the channels in the corresponding region of this rhythm are calculated as quantitative characteristics; for the μ rhythm, the frequency range, amplitude range and symmetry of the channels in the corresponding region of this rhythm are calculated as quantitative characteristics; for the θ rhythm, the frequency range and amplitude range of the channels in the corresponding region of this rhythm are calculated as quantitative characteristics; and then the quantitative characteristics of these three rhythms are used to generate a descriptive text about the background activity according to a fixed template and descriptive terms.

[0016] A computer device includes a memory and a processor. The memory stores a computer program, and the processor is configured to execute the computer program to implement the EEG report generation method combining the above-mentioned hybrid neural network and analysis experience algorithm.

[0017] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the EEG report generation method combining the above-mentioned hybrid neural network and analysis experience algorithm.

[0018] Based on the above technical solution, the present invention innovatively solves two major challenges of EEG pathological classification and EEG report generation by combining a hybrid neural network model with window feature-self-attention fusion and an EEG analysis empirical algorithm system: First, the window feature-self-attention fusion pattern performs information fusion of EEG at different time and space granularities, learns to assign different attention weights related to the classification task to windows, and completes the learning of EEG brain wave pattern features and selects representative windows in the EEG record; Second, the EEG analysis empirical algorithm visually increases the analysis granularity step by step from the matching of single waveforms to large-scale wave activities and then to regional background activities, and then renders a fixed template through features such as frequency, amplitude, symmetry, and exponent to complete the automatic generation of the core EEG findings description part in the EEG report. The present invention not only significantly improves the classification accuracy of EEG pathological classification across subjects, but also provides a solution for EEG pathological classification and further automatic report generation by combining the respective advantages of the EEG analysis empirical algorithm and the end-to-end hybrid neural network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flowchart of the EEG report generation method combining the hybrid neural network and the analysis empirical algorithm of the present invention.

[0020] Figure 2 It is a schematic structural diagram of the hybrid neural network model of the present invention.

[0021] Figure 3 It is a schematic structural diagram of the EEG analysis empirical algorithm system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to describe the present invention more specifically, the technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments.

[0023] As Figure 1 shown, the EEG report generation method combining the hybrid neural network and the analysis empirical algorithm of the present invention includes the following steps: (1) Obtain a conventional electroencephalogram (EEG) examination dataset, where the duration of each EEG collected in this dataset is 20 to 30 minutes, and at the same time record the label indicating whether the EEG record collected this time is pathologically normal or abnormal.

[0024] The dataset used in this embodiment is TUAB (an electroencephalogram dataset created from Temple University Hospital), version v3.0.1. There are a total of 2,993 EEG records in this dataset. Most of the recording durations are about 20 minutes, and all recording durations are not less than 15 minutes. The sampling rate is not less than 250 Hz, and the recording electrodes all include all the electrodes required by the international 10-20 standard lead system. There are 2,993 EEG records collected from 2,383 subjects in the dataset. The ages of the subjects are concentrated between 20 and 80 years old. Among them, 1,103 subjects are male and 1,280 subjects are female. The dataset has annotations for EEG records rather than for subjects. The annotations are divided into normal and abnormal. The former indicates that the EEG record is non-pathological, and the latter indicates that the EEG record is pathological; among the 2,993 EEG records, 1,521 are normal and 1,472 are abnormal.

[0025] (2) Perform conventional preprocessing on the EEG signals in the dataset to process the original EEG signals into a form that the neural network is good at processing.

[0026] In this embodiment, for the TUAB dataset, first select 21 channels of the international 10-20 standard lead system. These channels are included in all EEG records. Then use a band-pass filter of 0.5 - 70 Hz, which can remove high-frequency interference signals. At the same time, this is also a filtering method summarized from clinical experience. Then reduce the data sampling rate to 100 Hz to reduce the computational load. Furthermore, extract a total of 20 minutes of data from each EEG record starting from the 2nd minute to the 21st minute. The insufficient part is filled with 0s at the end. This can remove the significant artifacts in the first minute caused by reasons such as the acquisition device not being worn tightly at the beginning of the acquisition. Subsequently, multiply the data by 10 6 Thereby converting the data unit from volts to microvolts, which can reduce the precision overhead of the model for recording data. Finally, set the data points less than -800 microvolts and greater than 800 microvolts in the data to -800 microvolts and 800 microvolts respectively. This is to limit the input data within a certain range and at the same time counteract the large-range fluctuating artifacts caused by poor electrode contact, etc.

[0027] Next, divide the training set and the test set: Follow the publicly disclosed division method of TUAB. Put 2,717 EEG records out of the 2,993 data into the training set, and the other 276 EEG records into the test set. Among them, 1,371 labels in the training set are normal and 1,346 labels are abnormal. 150 labels in the test set are normal and 126 labels are abnormal. There is no situation where multiple EEG records collected from the same subject appear in the training set and the test set respectively.

[0028] (3)Construct a hybrid neural network model that fuses window features and self-attention, as Figure 2 shown. Its architecture includes a multi-scale temporal depth convolution module, a symmetric channel convolution module, a cross convolution module, a whole-brain feature convolution module, and a sub-window self-attention fusion module, where: The multi-scale temporal depth convolution module uses causal convolutions with different kernel sizes in four parallel branches to extract features of different scales in sub-windows, and then uses a convolutional network to fuse the features obtained from the four branches into a comprehensive feature. The specific expression is as follows:

[0029] Where: X represents the input signal, represents the i th convolution operation of the multi-scale temporal depth convolution module, Concat represents channel dimension concatenation, MergeConv represents the multi-scale feature fusion convolution operation.

[0030] The symmetric channel convolution module emphasizes comparing the features of symmetric channels in the left and right hemispheres of the brain. First, it divides H 1 to form left-right and midline channel features, then stacks and convolves the left and right brain features to form a set of fused features to obtain contrast features such as inter-hemispheric symmetry and synchronization, and then concatenates them with the midline channel features. The specific expression is as follows:

[0031]

[0032] Where: H 1-l , H 1-r , H 1-m respectively represent the features obtained by dividing H 1 according to brain regions, Split is the division operation, Stack represents stacking the corresponding channels in the left and right brain regions, Conv represents two-dimensional convolution; The cross convolution module pads 0 on both sides of the Fp1-Fp2 symmetric channel fusion feature and the O1-O2 symmetric channel fusion to arrange the channel features in a two-dimensional form to represent the relative position relationship of channels on the head. Then, convolutions are performed along the horizontal and vertical directions respectively to extract horizontal and vertical features, and then the horizontal features, vertical features, and H 2 are concatenated. The specific expression is as follows:

[0033]

[0034] Among them: Padding Indicates padding 0 to H 2, Reshape Indicates converting H 2 into a two-dimensional shape, Indicates performing depth convolution on rows, ColConv Indicates performing depth convolution on columns; The whole-brain feature convolution module fuses features using convolution, activation, normalization, and pooling. The specific expression is as follows:

[0035]

[0036] Among them: Conv Indicates the convolution operation, E Indicates the ELU activation function, BN Indicates batch normalization, Avgpool Indicates average pooling, Maxpool Indicates max pooling.

[0037] The sub-window self-attention fusion module concatenates the features of multiple sub-windows, then dynamically allocates attention weights between sub-windows based on the multi-head self-attention mechanism, and fuses the features of sub-windows into the features of the window. Among them, rotational position encoding is introduced to retain the temporal dependence features of bioelectric signals. The specific expression is as follows:

[0038]

[0039]

[0040] Among them: Attention Indicates the multi-head self-attention mechanism, Q , K and V respectively indicate the query, key, and value matrices, d K Is the dimension of the key, i Indicates the serial number of the sub-windows divided by each window. A 10-second window can be divided into 10 1-second sub-windows, Stack Indicates stacking the sub-window features in the channel dimension.

[0041] The self-attention window fusion network concatenates the features of multiple windows, then dynamically allocates attention weights between windows based on the multi-head self-attention mechanism, and fuses the features of windows into the features of the EEG record. Among them, rotational position encoding is introduced to retain the temporal dependence features of bioelectric signals. The specific expression is as follows:

[0042]

[0043] Wherein: i represents the serial number of the window divided for each EEG record. A 20-minute EEG signal can be divided into 120 windows of 10 seconds. Stack represents stacking the window features in the channel dimension.

[0044] The classification network uses a single-layer linear mapping layer to project the feature dimension onto the number of classifications to obtain the classification result. The specific expression is as follows:

[0045] Wherein: Linear represents the linear mapping layer.

[0046] (4) Use the collected routine electroencephalogram examination data to train the above window feature-self-attention fusion model to obtain an EEG pathological classification model.

[0047] Randomly select 5 random number seeds, fix the 5 random number seeds respectively to conduct 5 independent experiments to reduce accidental errors. Send the data processed in step (2) to the hybrid neural network model of window feature-self-attention fusion constructed in step (3). The window size is selected as 10 seconds without overlap, which is a commonly used EEG interpretation window size in clinics. In order to focus on the wave patterns within a short time, the sub-window size is selected as 1 second without overlap. The cross-entropy loss function is used as the loss function during the training process. There is a LogSoftmax (logarithm of softmax) operation in the cross-entropy loss function, which can convert the output of the last layer of the network model into a probability vector. The specific expression is as follows:

[0048] Wherein: Z y is the correct category y is the unnormalized score, represents the sum of the exponential scores of all categories. Here there are only two categories: normal and abnormal. Mini-batch training is used during the training. The loss within a batch is the mean of the losses of each sample. The specific expression is as follows:

[0049] Wherein: L batch represents the batch loss, N represents the number of samples within the batch, L n represents the nCross-entropy loss of a sample.

[0050] Adam optimizer is used for training, and the cosine annealing learning rate adjustment strategy is adopted, which helps to update the parameters stably during the training process. The method of the present invention is compared with the existing methods CWA_T (Channelized Autoencoder-Transformer Neural Network), Wavelet_LSTM (Wavelet Transform-Long Short-Term Memory Neural Network), EEGNet (Electroencephalogram Neural Network), EEGConformer (Electroencephalogram Convolutional Transformer Neural Network), and Deep4Net (Deep Four-Layer Convolutional Neural Network) in terms of accuracy, sensitivity, and specificity. It can be seen that the hybrid neural network of the method of the present invention has a significant improvement compared with other methods. As shown in Table 1, the data in the table are in the form of percentage values ± percentage standard deviation.

[0051] Table 1

[0052] (5) Construct an EEG analysis empirical algorithm system, such as Figure 3 shown, its structure includes: The core of the wave analysis module is the single-wave matching algorithm. The analysis object is a single wave on the EEG recording channel. First, the turning point acquisition algorithm is used to screen out some possible starting points, middle points, or ending points of the wave; then the single-wave matching algorithm is used, and the greedy strategy is applied. According to the empirical settings such as the ratio of the rising branch length to the falling branch length of the single wave being not less than 0.5 and not greater than 2, the single wave is converted into non-overlapping , , triples, which represent the starting point, middle point, and ending point of the wave respectively. Finally, the wave duration, frequency, amplitude, waveform characteristics, etc. are calculated according to the wave attribute calculation algorithm and the triples. The specific expressions are as follows:

[0053]

[0054]

[0055]

[0056] Among them: f represents the sampling rate, kertosis represents the kurtosis, Judge represents mapping the kurtosis value to the label of blunt or sharp, Duration i , Frequency i , Amplitude i , WaveForm i respectively represent thei The time course, frequency, amplitude, and waveform of a wave.

[0057] The core of the channel analysis module is the single-channel wave activity analysis algorithm, which analyzes the frequency characteristics of the waves matched on a single channel. According to δ (0.5~4Hz), θ (4~8Hz), α (8~14Hz), β (14~30Hz), γ (30Hz+), each wave is divided into the corresponding frequency band. Then, the waves in the same frequency band that continuously appear for a certain period of time are regarded as a wave activity. The frequency range, amplitude range of this wave activity are statistically analyzed, and whether there are abnormal waveforms such as sharp waves (time course less than 200 ms , amplitude greater than 100 μV ), spike waves (time course less than 70 ms , amplitude greater than 100 μV ), and slow waves (frequency less than 4Hz) are formed to obtain the descriptive data on the large-scale frequency band electroencephalogram activities. The specific expressions for statistically analyzing the wave activity information are as follows, and each formula has the constraint of i ∈[1,2,…, m :

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] Among them: avg Indicates calculating the average value, max , min respectively indicate calculating the maximum and minimum values. Vote represents making a vote based on the waveform characteristics of all waves, and the waveform characteristic of the entire activity takes the main waveform. m Indicates the number of single waves in the current channel. actName Indicates the frequency band name of the current wave activity. SharpWave , SpikeWave ,SlowWave The parameters used are all set based on clinical experience. act Indicates the activity name. Desp Indicates the description of the generated wave activity and the discovery of abnormal waves.

[0067] The core of the regional analysis module is the regional rhythm activity analysis algorithm. Based on the description data of the wave activities of each EEG channel, it further analyzes the background activities between channels and in brain regions to form a description of the background activities. It mainly analyzes the α rhythm in the posterior head region (channels O1, O2, F3, F4, T5, T6), the μ rhythm in the central region (channels C3, C4, Cz), and the θ rhythm in the frontal midline (channels FPz, Fz, Cz) of the background activities. For the posterior head α rhythm, calculate the frequency range, amplitude range, symmetry, and α index; for the central region μ rhythm, calculate the frequency range, amplitude range, symmetry; for the frontal midline θ rhythm, calculate the frequency range and amplitude range. The maximum and minimum values in the frequency range and amplitude range respectively come from the average values of the corresponding values of the channel wave activities. Symmetry compares both frequency and amplitude. If the amplitude difference is less than 30% and the frequency difference is less than 0.5 Hz, it is considered symmetric; if the amplitude difference is greater than 30% and less than 50%, and the frequency difference is greater than 0.5 Hz and less than 1 Hz, it is considered mildly asymmetric; if the amplitude difference is greater than 50% and the frequency difference is greater than 1 Hz, it is considered significantly asymmetric; the α index refers to the proportion of time occupied by α activity. Finally, a description of the regional background activity is formed based on these characteristics, and the specific expression is as follows:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] Where: n Indicates the quantity of a certain wave activity in the current channel. SymJudge Is the logic for empirical judgment of symmetry. In ch 1 and ch 2 respectively represent two channels with symmetric spatial positions. FreRange And AmpRange Indicate the frequency and amplitude range of the rhythm. rhythm Indicates the rhythm name. Desp Indicates the generated rhythm description.

[0074] (6) Combine the outputs of the hybrid neural network and the empirical algorithm to form an EEG report.

[0075] Connect the following five parts in sequence to form an EEG report: Basic information, including basic information of the subjects and EEG acquisition parameters, is read from the raw data; Representative windows showing the original EEG signal maps of windows with higher correlation with EEG classification results, obtained from high attention windows in the hybrid neural network; Background activities and abnormal waves, which consist of abnormal wave discovery and background activity description obtained from empirical algorithms; Conclusion, the hybrid neural network can classify the pathological results of EEG recordings, i.e., non-pathological EEG or pathological EEG.

[0076] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art to the present invention based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A method for generating an EEG report by combining a hybrid neural network and an analytical empirical algorithm, characterized in that, The steps are as follows: (1) Obtain an electroencephalogram (EEG) dataset, where each EEG data includes an EEG signal collected from a subject with a duration of 20 - 30 minutes and its pathological label; (2) After preprocessing the EEG signal, divide the EEG dataset into a training set and a test set; (3) Construct a hybrid neural network model that fuses window features and self - attention, which includes: A window feature extraction network, with windows as basic units, and uses various convolutional networks and self - attention networks to extract features from the data within each window of the EEG signal; A self - attention window feature fusion network, which uses the self - attention mechanism to dynamically assign attention weights to each window, and then fuses the features of each window to form a feature representation of the entire EEG signal; A classification network, which outputs the prediction probability that the EEG signal belongs to pathological normality according to the feature representation of the EEG signal; (4) Use the EEG data in the training set to train the above - mentioned hybrid neural network model to obtain an EEG pathological classification model; (5) Construct an EEG analysis empirical algorithm system, which includes: A wave analysis module, which is used to analyze each turning point in the EEG signal and dynamically match and generate single - wave features that conform to visual laws for each channel based on the empirical characteristics of forming brain waves; A channel analysis module, which is used to analyze the single - wave features of each channel and generate descriptive data about the brain wave activities of a large - scale same - frequency band; A region analysis module, which further analyzes the background activities between channels and on brain regions according to the descriptive data of brain wave activities and generates a descriptive text about the background activities; (6) Input the EEG data in the test set into the EEG pathological classification model and the EEG analysis empirical algorithm system respectively, and combine the prediction results of the model about pathological normality or abnormality and the information calculated by the algorithm system to form an EEG report.

2. The EEG report generation method combining a hybrid neural network and an analytical experience algorithm according to claim 1, characterized in that: The process of preprocessing the EEG signal in step (2) is as follows: First, select 21 channels specified by the international 10-20 standard lead system from the EEG signal. Then, perform band-pass filtering on the EEG signal with a frequency range of 0.5~70 Hz. Next, reduce the data sampling rate to 100 Hz, remove the data of the first minute of the EEG signal, extract a total of 20 minutes of data starting from the second minute of the EEG signal, and multiply the data by 10. 6 Thereby, convert the data unit from volts to microvolts, and set the data points in the EEG signal less than -800 microvolts and greater than 800 microvolts to -800 microvolts and 800 microvolts respectively.

3. A method for generating an EEG report by combining a hybrid neural network and an analysis experience algorithm according to claim 1, characterized in that: The window feature extraction network includes: A multi - scale temporal depth convolution module, which further divides the data within the EEG signal window into multiple sub - windows, and is used to capture the patterns of different time courses that appear on each channel of the sub - window data; A symmetric channel convolution module, which is used to perform contrast fusion on the output of the multi - scale temporal depth convolution module on the channels that are symmetric in the left and right hemispheres; A cross - cross convolution module, which is used to perform convolution fusion on the output of the symmetric channel convolution module in the horizontal and vertical directions in space to learn the local spatial information between channels; A whole - brain feature convolution module, which is used to perform convolution on the output of the cross - cross convolution module in the whole - brain space to obtain global spatial information; A sub - window self - attention fusion module, which is used to pass the output of the whole - brain feature convolution module through a self - attention layer, and assign attention weights to each sub - window to fuse and obtain window features.

4. A method for generating an EEG report by combining a hybrid neural network and an analytical empirical algorithm according to claim 3, characterized in that: The multi-scale temporal depth convolution module performs one-dimensional depth convolution on the sub-window data of the EEG signal with 4 groups of parallel different convolution kernel sizes, and then splices the outputs of these 4 groups of convolutions and recompresses them into 1 group of outputs through one-dimensional depth convolution; the symmetric channel convolution module stacks the outputs of the multi-scale temporal depth convolution module according to the left-right brain symmetric position features and then performs a two-dimensional depth convolution, and then splices the output of this convolution with the features of the midline channel, and then outputs after passing through the activation function ELU and batch normalization; the cross-cross convolution module arranges the outputs of the symmetric channel convolution module two-dimensionally according to the brain region spatial positions, fills 0 at the four corners to form a 5-row and 3-column matrix, and then extracts the local spatial fusion features of the rows and columns through two-dimensional depth convolution in the horizontal and vertical directions respectively, and then splices the local spatial fusion features of the rows and columns with the outputs of the symmetric channel convolution module in the channel dimension and outputs; the whole-brain feature convolution module sequentially passes the outputs of the cross-cross convolution module through one-dimensional convolution, activation function ELU, batch normalization, average pooling, one-dimensional convolution, activation function ELU, batch normalization, max pooling, and flattening operations to obtain sub-window features; the sub-window self-attention fusion module sequentially passes the outputs of the whole-brain feature convolution module through layer normalization L1, multi-head self-attention mechanism layer, layer normalization L2, and MLP to obtain window features, where the input of layer normalization L1 and the output of the multi-head self-attention mechanism layer are stacked as the input of layer normalization L2, and the input of layer normalization L2 and the output of MLP are stacked to obtain window features.

5. A method for generating an EEG report by combining a hybrid neural network and an analysis experience algorithm according to claim 1, characterized in that: The self-attention window feature fusion network is sequentially composed of layer normalization L3, multi-head self-attention mechanism layer, layer normalization L4, and MLP from input to output, where the input of layer normalization L3 and the output of the multi-head self-attention mechanism layer are stacked as the input of layer normalization L4, and the input of layer normalization L4 and the output of MLP are stacked as the final output of the self-attention window feature fusion network; the classification network uses a single-layer linear mapping layer without the activation function Softmax.

6. A method for generating an EEG report by combining a hybrid neural network and an analytical empirical algorithm according to claim 1, characterized in that: For the window data of any channel of the EEG signal, the wave analysis module converts it into discrete turning points, and then selects and combines these turning points according to the shape conditions of the formed waves to generate a single-wave matching result in the form of a triple. i s , i m , i e , where i s 、 i m 、 i e represent the starting point, midpoint, and ending point of the single wave respectively. Finally, according to the triple i s , i m , i e , the single-wave characteristics including frequency, amplitude, morphology, and duration are calculated through the wave attribute algorithm.

7. A method for generating an EEG report by combining a hybrid neural network and an analytical empirical algorithm according to claim 1, characterized in that: The channel analysis module divides the single waves of each channel into corresponding frequency bands according to the single-wave features, and regards the waves of the same frequency band that have a certain scale and continuously appear in time as one brain wave activity, and the statistical description data of such activities include start and end times, frequency range, amplitude range, and whether there are abnormal waveforms including sharp waves, spike waves, and slow waves.

8. A method for generating an EEG report by combining a hybrid neural network and an analytical experience algorithm according to claim 1, characterized in that: The region analysis module further analyzes the α rhythm in the posterior head region, the μ rhythm in the central region, and the θ rhythm in the fronto-midline region according to the description data of the brain wave activity. For the α rhythm, calculate the frequency range, amplitude range, symmetry, and α index of the channels in the corresponding region of this rhythm as quantitative features; For the μ rhythm, calculate the frequency range, amplitude range, and symmetry of the channels in the corresponding region of this rhythm as quantitative features; for the θ rhythm, calculate the frequency range and amplitude range of the channels in the corresponding region of this rhythm as quantitative features; Furthermore, generate descriptive texts about the background activity according to the fixed template and descriptive terms for the quantitative features of these three rhythms.

9. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, and characterized in that: The processor is used to execute the computer program to implement an EEG report generation method combining a hybrid neural network and an analytical empirical algorithm as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements an EEG report generation method combining a hybrid neural network and an analytical empirical algorithm as described in any one of claims 1 to 8.

Images