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

Abstract

The present invention discloses an EEG report generation method combining a hybrid neural network and an analysis experience algorithm. The method selects a window that is highly correlated with a classification result and obtains a pathological judgment result through a hybrid neural network of multi-scale time series deep convolution, symmetric channel convolution, cross convolution, whole-brain feature convolution, sub-window self-attention fusion, self-attention window fusion, and a classification network. The window is described in combination with a pattern of gradually expanding the description area from wave to channel to region in the EEG analysis experience algorithm system, abnormal waves are discovered and background activities are described. Finally, basic information, window EEG signal graph, background activity description generated by the EEG analysis experience algorithm system, abnormal wave description, and classification results provided by the hybrid neural network are connected in series to form an EEG report. A solution for generating an EEG report that combines 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 (EEG) information technology, and in particular relates to an EEG report generation method combining a hybrid neural network and an analytical experience algorithm. Background Art

[0002] Generating EEG (electroencephalogram) reports is a core step in medical examination and diagnosis. As a physiological signal reflecting the brain's electrical activity, it effectively displays the subject's brain's physiological state thanks to its high temporal resolution and waveform characteristics that vary depending on the brain region. However, the generation of EEG reports currently faces challenges such as difficulty interpreting EEG signals, weak anti-interference capabilities, and significant inter-individual variability.

[0003] Currently, EEG classification relies primarily on manual inspection, while automated methods primarily fall into two categories: traditional machine learning and deep learning. Traditional machine learning methods rely on hand-designed features and shallow classification methods (such as support vector machines, fuzzy search trees, and random forests). However, this feature design is difficult and presents generalization challenges. Deep learning methods utilize parameter-learnable methods, flexibly adjusting parameters to capture specific patterns in the EEG, forming internal features that are then used for classification, resulting in better generalization. For example, convolutional neural networks (CNNs) extract temporal and spatial features, capable of capturing specific patterns within and between channels. Manual EEG interpretation is achieved by observing a single window at a time, integrating information from all windows, and then organizing the EEG interpretation into a structured report. There is relatively little work related to the generation of EEG reports, and the description of different EEG phenotypes is also affected by individual differences. Therefore, in clinical practice, there are only empirical standards but no "gold" standard. Fortunately, the literature [Liu Xiaoyan, "Clinical EEG Training Course", 1st edition, Beijing, China: People's Medical Publishing House, November 2011, Chapters 6-8, pp. 45-98] has quantitatively standardized the basic terms and phenotypes in EEG reports. This provides a standardized expression method and direction for the description of EEG phenotypes and the generation of EEG reports.

[0004] Numerous studies have been devoted to addressing interference issues in EEG signals. For example, the article [Zhao Xin, Wu Jianxing, Wang Kun, et al. A Review of EEG Artifact Removal Algorithms [J / OL]. Signal Processing, 1-26] discusses in detail the application of various techniques for EEG artifact removal, including regression, filtering, blind source separation, signal decomposition, and deep learning. Artifact removal is not only a hot topic of current research but also holds broad development prospects, directly impacting the accuracy of EEG data interpretation.

[0005] In response to the challenge of large inter-individual EEG differences, improving the generalization ability of automated methods has become one of the key solutions. Transfer learning, as a mature technology, has been widely used in this regard. For example, the literature [Qiu Yihui, Jiang Qiong, Wei Lingling, et al. Cross-subject EEG fatigue driving detection based on transfer learning [J]. Journal of Nanchang University (Science Edition), 2023, 47(04): 397-402.] proposed a method that combines model transfer learning with 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 considered to be effective solutions.

[0006] Among the aforementioned methods, deep learning approaches offer the advantage of strong generalization capabilities but suffer from a lack of interpretability. Empirical approaches offer the advantage of interpreting clinical implications but suffer from insufficient generalization capabilities. The process from EEG interpretation to EEG report generation requires both extensive clinical experience and a high degree of objectivity to account for the vastly varying individual EEG characteristics. This is a 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 that combines a hybrid neural network and an analytical experience algorithm, which can significantly improve the classification accuracy of EEG pathology classification across subjects.

[0008] A method for generating an EEG report by combining a hybrid neural network and an analytical empirical algorithm comprises the following steps:

[0009] (1) Obtain an EEG dataset, where each EEG data point contains an EEG signal collected from a subject for 20 to 30 minutes and its pathological label (normal or abnormal);

[0010] (2) dividing the EEG data set into a training set and a test set after preprocessing the EEG signal;

[0011] (3) Constructing a hybrid neural network model of window feature-self-attention fusion, which includes:

[0012] The window feature extraction network uses the window as the basic unit and extracts features from the data in each window of the EEG signal through various convolutional networks and self-attention networks;

[0013] The self-attention window feature fusion network 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;

[0014] The classification network outputs the predicted probability of the EEG signal being pathological or normal based on the characteristic representation of the EEG signal;

[0015] (4) Using the EEG data in the training set to train the hybrid neural network model, an EEG pathology classification model is obtained;

[0016] (5) Construct an EEG analysis empirical algorithm system, which includes:

[0017] The wave analysis module is used to analyze the turning points in the EEG signal and dynamically match the characteristics of the brain waves to generate single wave features that conform to visual laws in each channel;

[0018] Channel analysis module, used to analyze the single wave characteristics of each channel and generate descriptive data about brain wave activities in the same frequency band on a large scale;

[0019] The regional analysis module further analyzes the background activities between channels and brain regions based on the description data of brain wave activities and generates a descriptive text about the background activities;

[0020] (6) The EEG data of the test set are input into the EEG pathological classification model and the EEG analysis empirical algorithm system respectively, and the prediction results of the model output on pathological normality or abnormality and the information calculated by the algorithm system are combined to form an EEG report.

[0021] Furthermore, the process of preprocessing the EEG signal in step (2) is as follows: first, 21 channels specified by the international 10-20 standard lead system are selected from the EEG signal, and then the EEG signal is band-pass filtered at 0.5-70 Hz, and then the data sampling rate is reduced to 100 Hz, the data of the first minute of the EEG signal is removed, and the data of the EEG signal starting from the second minute, totaling 20 minutes, is taken out and the data is multiplied by 10. 6 The data unit is thereby converted from volts to microvolts, and the data points in the EEG signal that are less than -800 microvolts and greater than 800 microvolts are set to -800 microvolts and 800 microvolts, respectively.

[0022] Furthermore, the window feature extraction network includes:

[0023] The multi-scale temporal deep convolution module further subdivides the data within the EEG signal window into multiple sub-windows to capture the different time-course patterns of the sub-window data on each channel. It can determine the weights of different time-course patterns on the channel based on the characteristics of each channel.

[0024] Symmetric channel convolution module, used to compare and fuse the output of the multi-scale temporal depth convolution module on channels that are symmetrical in the left and right hemispheres;

[0025] The cross convolution module is used to spatially fuse the output of the symmetric channel convolution module along the horizontal and vertical convolutions to learn the local spatial information between channels;

[0026] The whole-brain feature convolution module is used to convolve the output of the cross convolution module on the whole-brain space to obtain global spatial information;

[0027] The sub-window self-attention fusion module is used to pass the output of the whole-brain feature convolution module through the self-attention layer, assign attention weights to each sub-window, and thus fuse the window features.

[0028] Furthermore, the multi-scale temporal depth convolution module subjects the sub-window data of the EEG signal to four parallel one-dimensional depth convolutions with different convolution kernel sizes, and then concatenates the four convolution outputs and recompresses them into one set of outputs through one-dimensional depth convolution; the symmetric channel convolution module stacks the output of the multi-scale temporal depth convolution module according to the symmetrical position features of the left and right brains and then concatenates the output of the convolution with the features of the midline channel and then processes it through the activation function ELU (exponential linear unit) and batch normalization and outputs it; the cross convolution module arranges the output of the symmetric channel convolution module in two dimensions according to the spatial position of the brain area, fills the four corners with 0, forms a matrix of 5 rows and 3 columns, and then extracts the local space of the rows and columns through horizontal and vertical two-dimensional depth convolutions respectively. Fusion features, and then the local spatial fusion features of 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 sequentially obtains sub-window features after the output of the cross convolution module undergoes one-dimensional convolution, activation function ELU, batch normalization, average pooling, one-dimensional convolution, activation function ELU, batch normalization, maximum pooling, and flattening operations; the sub-window self-attention fusion module sequentially obtains window features after the output of the whole-brain feature convolution module undergoes layer normalization L1, multi-head self-attention mechanism layer, layer normalization L2, and MLP (multi-layer perceptron), wherein the input of layer normalization L1 and the output of the multi-head self-attention mechanism layer are superimposed as the input of layer normalization L2, and the input of layer normalization L2 and the output of MLP are superimposed to obtain window features.

[0029] Furthermore, the self-attention window feature fusion network is composed of layer normalization L3, multi-head self-attention mechanism layer, layer normalization L4, and MLP connection from input to output, wherein the input of layer normalization L3 and the output of multi-head self-attention mechanism layer are superimposed as the input of layer normalization L4, and the input of layer normalization L4 and the output of MLP are superimposed as the final output of the self-attention window feature fusion network; the classification network adopts a single-layer linear mapping layer without the activation function Softmax (softened maximum).

[0030] Furthermore, the wave analysis module converts the window data of any channel of the EEG signal into discrete turning points (corresponding to possible peaks and troughs), and then selects and combines these turning points according to the shape conditions of the wave to generate single wave matching results in the form of triples [ i s , i m , i e ],in i s 、 i m 、 i e Represent the starting point, midpoint (peak or trough) and end point of a single wave respectively. Finally, according to the triple [ i s , i m , i e ]The single wave characteristics including frequency, amplitude, shape and duration are calculated through wave attribute algorithm.

[0031] Furthermore, the channel analysis module divides the single waves of each channel into corresponding frequency bands according to the single wave characteristics. Waves of the same frequency band that are of a certain scale and appear continuously in time are regarded as one brain wave activity. The descriptive data of such activities are statistically analyzed, including the start and end time, frequency range, amplitude range, and whether there are abnormal waveforms including sharp waves, spike waves, and slow waves.

[0032] Furthermore, the regional analysis module further analyzes the alpha rhythm of the occipital region, the mu rhythm of the central region, and the theta rhythm of the frontal midline region based on the descriptive data of brain wave activity. For the alpha rhythm, the frequency range, amplitude range, symmetry, and alpha index of the regional channel corresponding to the rhythm are calculated as quantitative features; for the mu rhythm, the frequency range, amplitude range, and symmetry of the regional channel corresponding to the rhythm are calculated as quantitative features; for the theta rhythm, the frequency range and amplitude range of the regional channel corresponding to the rhythm are calculated as quantitative features; and then the quantitative features of these three rhythms are used to generate a descriptive text about the background activity according to a fixed template and descriptive terms.

[0033] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor is used to execute the computer program to implement the EEG report generation method combining the hybrid neural network and the analysis experience algorithm.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the EEG report generation method combining the hybrid neural network and the analysis experience algorithm.

[0035] Based on the above technical solution, the present invention innovatively solves the two major challenges of EEG pathology classification and EEG report generation by combining a hybrid neural network model of window feature-self-attention fusion and an EEG analysis empirical algorithm system: first, the window feature-self-attention fusion model performs EEG information fusion at different time and space granularities, learns to assign different attention weights related to the classification task to windows, completes the learning of EEG brain wave pattern features and selects representative windows in EEG records; second, the EEG analysis empirical algorithm visually improves the granularity of analysis from single waveform matching to large-scale wave activity to regional background activity, and then renders a fixed template through features such as frequency, amplitude, symmetry, and index to complete the automatic generation of the core EEG description part in the EEG report. The present invention not only significantly improves the classification accuracy of EEG pathology classification across subjects, but also provides a solution for EEG pathology 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

[0036] Figure 1 This is a flow chart of the EEG report generation method combining a hybrid neural network and an analytical empirical algorithm according to the present invention.

[0037] Figure 2 Schematic diagram of the structure of the hybrid neural network model in the present invention.

[0038] Figure 3 Schematic diagram of the structure of the EEG analysis empirical algorithm system in the present invention. DETAILED DESCRIPTION

[0039] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown, the EEG report generation method combining the hybrid neural network and the analysis experience algorithm of the present invention includes the following steps:

[0041] (1) Obtain a routine EEG examination data set, in which each EEG recorded lasts for 20 to 30 minutes, and records the label of whether the EEG recorded this time is pathologically normal or abnormal.

[0042] The dataset used in this embodiment is TUAB (an electroencephalogram dataset created from Temple University Hospital), version v3.0.1. The dataset contains 2993 EEG records, most of which are around 20 minutes long, with all records lasting no less than 15 minutes. The sampling rate is no less than 250 Hz, and the recording electrodes all include all electrodes required by the international 10-20 standard lead system. The dataset contains 2993 EEG records collected from 2383 subjects, whose ages range from 20 to 80 years old, of which 1103 subjects are male and 1280 subjects are female. The dataset has annotations for the EEG records rather than the 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. Of the 2993 EEG records, 1521 are normal and 1472 are abnormal.

[0043] (2) Perform conventional preprocessing on the EEG signals in the dataset to process the raw EEG signals into a form that the neural network is good at processing.

[0044] In this embodiment, for the TUAB data set, the 21 channels of the international 10-20 standard lead system are first selected. These channels are included in all EEG records. Then, a 0.5-70 Hz bandpass filter is used to remove high-frequency interference signals. This is also a filtering method summarized in clinical experience. The data sampling rate is then reduced to 100 Hz. This is to reduce the amount of calculation. Then, a total of 20 minutes of data from the 2nd minute to the 21st minute of each EEG record is taken. The insufficient part is padded with 0 at the end. This can remove the significant artifacts of the first minute caused by the fact that the acquisition device is not worn tightly at the beginning of the acquisition. The data is then multiplied by 10 6 This converts the data unit from volts to microvolts, which can reduce the accuracy overhead of the model recording data. Finally, the data points with a value less than -800 microvolts and the data points with a value greater than 800 microvolts are set to -800 microvolts and 800 microvolts respectively. This is done to limit the input data to a certain range and combat the artifacts of large-scale fluctuations caused by poor electrode contact.

[0045] Next, we divided the data into training and test sets: Following the publicly available partitioning method of TUAB, we placed 2717 EEG recordings from the 2993 data points into the training set and the remaining 276 EEG recordings into the test set. Of these, 1371 recordings in the training set were labeled normal and 1346 were labeled abnormal, while 150 recordings in the test set were labeled normal and 126 were labeled abnormal. No EEG recordings from the same subject appeared in both the training and test sets.

[0046] (3) Construct a hybrid neural network model of window feature-self-attention fusion, such as Figure 2 As shown in the figure, 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:

[0047] The multi-scale temporal deep convolution module uses four parallel branches of causal convolution with different kernel sizes to extract features of sub-windows at different scales, 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:

[0048]

[0049] in: X represents the input signal, Represents the first layer of the multi-scale temporal depth convolution module i convolution operations, Concat represents channel dimension splicing, MergeConv Represents the multi-scale feature fusion convolution operation.

[0050] The symmetric channel convolution module emphasizes the comparison of the features of the symmetric channels of the left and right hemispheres. First, the H 1 forms left and right and midline channel features, and 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 splices them with the midline channel features. The specific expression is as follows:

[0051]

[0052]

[0053] in: H 1-l 、 H 1-r 、 H 1-m Respectively express H 1 According to the characteristics of brain region segmentation, Split For the split operation, Stack It means stacking the left and right brain channels accordingly. Conv represents a two-dimensional convolution;

[0054] The cross convolution module fills 0 on both sides of the Fp1-Fp2 symmetric channel fusion feature and the O1-O2 symmetric channel fusion feature to arrange the channel features into a two-dimensional form to represent the relative position relationship between the channels in the head. Then, convolution is performed along the horizontal and vertical directions respectively to extract the horizontal and vertical features, and then the horizontal features, vertical features and H 2 splicing, the specific expression is as follows:

[0055]

[0056]

[0057] in: Padding Express H 2 do 0 filling, Reshape Indicates that H 2 into a two-dimensional shape, represents the depthwise convolution of the rows, ColConv Indicates depthwise convolution of columns;

[0058] The whole-brain feature convolution module uses convolution, activation, normalization, and pooling to fuse features. The specific expression is as follows:

[0059]

[0060]

[0061] in: Conv represents the convolution operation, E represents the ELU activation function, BN represents batch normalization, Avgpool represents average pooling, Maxpool Represents maximum pooling.

[0062] The sub-window self-attention fusion module concatenates the features of multiple sub-windows, and then dynamically assigns attention weights between sub-windows based on the multi-head self-attention mechanism, fusing the sub-window features into the window features. Rotational position encoding is introduced to preserve the temporal dependency characteristics of bioelectric signals. The specific expression is as follows:

[0063]

[0064]

[0065]

[0066] in: Attention represents the multi-head self-attention mechanism, Q 、 K and V denote query, key, and value matrices respectively, d K is the dimension of the key, i Indicates the sequence 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 of sub-window features in the channel dimension.

[0067] The self-attention window fusion network concatenates the features of multiple windows, and then dynamically assigns attention weights between windows based on the multi-head self-attention mechanism, fusing the features of the windows into the features of the EEG recording. Rotational position encoding is introduced to preserve the temporal dependency characteristics of the bioelectric signal. The specific expression is as follows:

[0068]

[0069]

[0070] in: i Indicates the serial number of the window divided into each EEG record. A 20-minute EEG signal can be divided into 120 10-second windows. Stack Indicates stacking of window features in the channel dimension.

[0071] The classification network uses a single-layer linear mapping layer to project the feature dimension onto the classification number to obtain the classification result. The specific expression is as follows:

[0072]

[0073] in: Linear Represents a linear mapping layer.

[0074] (4) The collected routine EEG examination data is used to train the above-mentioned window feature-self-attention fusion model to obtain an EEG pathological classification model.

[0075] Randomly select 5 random number seeds, fix 5 random number seeds respectively, and conduct 5 independent experiments to eliminate accidental errors. The data processed in step (2) is sent to the hybrid neural network model of window feature-self-attention fusion constructed in step (3). The window size is selected as a non-overlapping 10-second size, which is a commonly used EEG interpretation window size in clinical practice. In order to focus on the wave pattern in a short period of time, the sub-window is selected as a non-overlapping 1-second size. The training process uses the cross entropy loss function as the loss function. The cross entropy loss function contains the LogSoftmax (softened maximum logarithm) operation, which can convert the output of the last layer of the network model into a probability vector. The specific expression is as follows:

[0076]

[0077] in: Z y Is the correct category y Unnormalized score, It represents the sum of the index scores of all categories. Here, there are only two categories: normal and abnormal. Mini-batch training is used in training. The loss in a batch is the mean loss of each sample. The specific expression is as follows:

[0078]

[0079] in: L batch Indicates batch loss, N represents the number of samples in the batch, L n Indicates the n The cross entropy loss of samples.

[0080] The Adam optimizer is used for training, and the cosine annealing learning rate adjustment strategy is used to help 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 convolution 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 is significantly improved compared to other methods, as shown in Table 1, where the data are in the form of percentage values ​​± percentage standard deviation.

[0081] Table 1

[0082]

[0083] (5) Construct an EEG analysis experience algorithm system, such as Figure 3 As shown, its structure includes:

[0084] 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 the starting point, middle point or end point of some possible waves. Then, the single wave matching algorithm is used with a greedy strategy. According to the empirical settings such as the ratio of the length of the rising branch to the falling branch of a single wave is not less than 0.5 and not greater than 2, the single wave is converted into non-overlapping [ , , ] triples, representing the starting point, midpoint, and end point of the wave respectively. Finally, the wave attribute calculation algorithm and the triples are used to calculate the time course, frequency, amplitude, waveform characteristics, etc. of the wave. The specific expressions are as follows:

[0085]

[0086]

[0087]

[0088]

[0089] in:f represents the sampling rate, kertosis represents the kurtosis, Judge Indicates mapping the kurtosis value to a round or sharp label. Duration i 、 Frequency i 、 Amplitude i 、 WaveForm i Respectively represent i The time course, frequency, amplitude and waveform of a wave.

[0090] 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. δ (0.5~4Hz), θ (4~8Hz), α (8~14Hz), β (14~30Hz), γ (30Hz+) Divide each wave into the corresponding frequency band, and then regard the waves of the same frequency band that appear continuously for a certain period of time as one wave activity, and count the frequency range and amplitude range of this wave activity, and whether there is a sharp wave (time course less than 200 ms , the amplitude is greater than 100 μV ), spike wave (duration less than 70 ms , the amplitude is greater than 100 μV ) and slow waves (frequency less than 4Hz) and other abnormal waveforms, forming descriptive data on large-scale band brain wave activity. The specific expression of statistical wave activity information is as follows, each formula has i ∈[1,2,…, m ] constraints:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] in: avg Indicates the average value, max 、 min They represent finding the maximum and minimum values ​​respectively. Voting means voting based on the waveform characteristics of all waves. The waveform characteristics of the entire activity take the main waveform. m Indicates the number of single waves in the current channel. actName Indicates the name of the frequency band of the current wave activity, SharpWave 、 SpikeWave 、 SlowWave The parameters used are all based on clinical experience. act Indicates the name of the activity. Desp Represents the generated wave activity and description of abnormal wave findings.

[0101] The core of the regional analysis module is the regional rhythmic activity analysis algorithm. Based on the descriptive data of wave activity in each EEG channel, it further analyzes background activity between channels and across brain regions to form a description of background activity. The main analysis of background activity is the alpha rhythm in the occipital region (channels O1, O2, F3, F4, T5, and T6), the mu rhythm in the central region (channels C3, C4, and Cz), and the theta rhythm in the frontal midline (channels FPz, Fz, and Cz). For the occipital alpha rhythm, the frequency range, amplitude range, symmetry, and alpha index are calculated; for the central mu rhythm, the frequency range, amplitude range, and symmetry are calculated; and for the frontal midline theta rhythm, the frequency range and amplitude range are calculated. The maximum and minimum values ​​in the frequency and amplitude ranges are derived from the average values ​​of the corresponding values ​​of channel wave activity, respectively. Symmetry is compared in terms of frequency and amplitude. If the amplitude difference is less than 30% and the frequency difference is less than 0.5 Hz, it is considered symmetrical; 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 regional background activity is formed based on these characteristics. The specific expression is as follows:

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] in: nIndicates the number of a certain wave activity in the current channel, SymJudge For the logic of empirical judgment of symmetry, in ch 1 and ch 2 represent two channels with symmetrical spatial positions, FreRange and AmpRange Indicates the frequency and amplitude range of the rhythm, rhythm Indicates the rhythm name, Desp Represents the generated rhythm description.

[0109] (6) Combining the outputs of the hybrid neural network and the empirical algorithm to form an EEG report.

[0110] Concatenate the following five parts in sequence to form an EEG report:

[0111] Basic information, including the subject's basic information and EEG acquisition parameters, is read from the raw data;

[0112] 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;

[0113] Background activity and abnormal waves, which consist of abnormal wave discovery and background activity description obtained from empirical algorithms;

[0114] Conclusion, the hybrid neural network can classify EEG recordings into pathological categories, i.e., non-pathological EEG or pathological EEG.

[0115] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It is apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring creative effort. Therefore, the present invention is not limited to the above embodiments. Any improvements or modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for generating an EEG report by combining a hybrid neural network and an empirical analysis algorithm, characterized in that: The steps include: (1) Obtain an EEG dataset, where each EEG data contains an EEG signal collected from a subject for 20 to 30 minutes and its pathological label; (2) dividing the EEG data set into a training set and a test set after preprocessing the EEG signal; (3) Constructing a hybrid neural network model of window feature-self-attention fusion, which includes: The window feature extraction network uses the window as the basic unit and extracts features from the data in each window of the EEG signal through various convolutional networks and self-attention networks; The self-attention window feature fusion network 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; The classification network outputs the predicted probability of the EEG signal being pathological or normal based on the characteristic representation of the EEG signal; The window feature extraction network includes: The multi-scale temporal deep convolution module further subdivides the data within the EEG signal window into multiple sub-windows to capture the patterns of different time courses of the sub-window data on each channel; Symmetric channel convolution module, used to compare and fuse the output of the multi-scale temporal depth convolution module on channels that are symmetrical in the left and right hemispheres; The cross convolution module is used to spatially fuse the output of the symmetric channel convolution module along the horizontal and vertical convolutions to learn the local spatial information between channels; The whole-brain feature convolution module is used to convolve the output of the cross convolution module on 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 the self-attention layer, assign attention weights to each sub-window, and thus fuse the window features; (4) Using the EEG data in the training set to train the hybrid neural network model, an EEG pathology classification model is obtained; (5) Construct an EEG analysis empirical algorithm system, which includes: The wave analysis module is used to analyze the turning points in the EEG signal and dynamically match the characteristics of the brain waves to generate single wave features that conform to visual laws in each channel; Channel analysis module, used to analyze the single wave characteristics of each channel and generate descriptive data about brain wave activities in the same frequency band on a large scale; The regional analysis module further analyzes the background activities between channels and brain regions based on the description data of brain wave activities and generates a descriptive text about the background activities; (6) The EEG data of the test set are input into the EEG pathological classification model and the EEG analysis empirical algorithm system respectively, and the prediction results of the model output on pathological normality or abnormality and the information calculated by the algorithm system are combined to form an EEG report.

2. The EEG report generation method combining a hybrid neural network and an analytical empirical algorithm according to claim 1, characterized in that: The process of preprocessing the EEG signal in step (2) is as follows: first, 21 channels specified by the international 10-20 standard lead system are selected from the EEG signal, and then the EEG signal is band-pass filtered at 0.5-70 Hz. Then, the data sampling rate is reduced to 100 Hz, the data of the first minute of the EEG signal is removed, and the data of the EEG signal starting from the second minute, totaling 20 minutes, is taken and the data is multiplied by 10. 6 The data unit is thus converted from volts to microvolts, and the data points in the EEG signal that are less than -800 microvolts and greater than 800 microvolts are set to -800 microvolts and 800 microvolts, respectively.

3. The EEG report generation method combining a hybrid neural network and an empirical analysis algorithm according to claim 1, characterized in that: The multi-scale temporal depth convolution module performs four parallel one-dimensional depth convolutions of different convolution kernel sizes on the sub-window data of the EEG signal, and then splices the four convolution outputs and recompresses them into one set of outputs through one-dimensional depth convolution; the symmetric channel convolution module stacks the output of the multi-scale temporal depth convolution module according to the symmetrical position features of the left and right brains and then performs a two-dimensional depth convolution, and then splices the output of the convolution with the features of the midline channel and then processes it through the activation function ELU and batch normalization and outputs it; the cross convolution module arranges the output of the symmetric channel convolution module in two dimensions according to the spatial position of the brain area, fills the four corners with 0, forms a matrix of 5 rows and 3 columns, and then extracts the local spatial fusion features of the rows and columns through horizontal and vertical two-dimensional depth convolutions respectively. Features, and then the local spatial fusion features of 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 sequentially subjects the output of the cross convolution module to one-dimensional convolution, activation function ELU, batch normalization, average pooling, one-dimensional convolution, activation function ELU, batch normalization, maximum pooling, and flattening operations to obtain sub-window features; the sub-window self-attention fusion module sequentially subjects the output of the whole-brain feature convolution module to layer normalization L1, multi-head self-attention mechanism layer, layer normalization L2, and MLP to obtain window features, wherein the input of layer normalization L1 and the output of the multi-head self-attention mechanism layer are superimposed as the input of layer normalization L2, and the input of layer normalization L2 and the output of MLP are superimposed to obtain window features.

4. The EEG report generation method combining a hybrid neural network and an analytical empirical algorithm according to claim 1, characterized in that: The self-attention window feature fusion network is composed of layer normalization L3, multi-head self-attention mechanism layer, layer normalization L4, and MLP connection from input to output, wherein the input of layer normalization L3 and the output of multi-head self-attention mechanism layer are superimposed as the input of layer normalization L4, and the input of layer normalization L4 and the output of MLP are superimposed as the final output of the self-attention window feature fusion network; the classification network adopts a single-layer linear mapping layer without the activation function Softmax.

5. The EEG report generation method combining a hybrid neural network and an empirical analysis algorithm according to claim 1, characterized in that: The wave analysis module converts the window data of any channel of the EEG signal into discrete turning points, and then selects and combines these turning points according to the shape conditions of the wave to generate a single wave matching result in the form of a triplet [ i s , i m , i e ],in i s 、 i m 、 i e Represent the starting point, midpoint and end point of a single wave respectively, and finally according to the triplet [ i s , i m , i e ]The single wave characteristics including frequency, amplitude, shape and duration are calculated through wave attribute algorithm.

6. The EEG report generation method combining a hybrid neural network and an empirical analysis 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 characteristics. Waves of a certain scale and continuous occurrence in time in the same frequency band are regarded as one brain wave activity. The descriptive data of such activities are statistically analyzed, including the start and end time, frequency range, amplitude range, and whether there are abnormal waveforms including sharp waves, spike waves, and slow waves.

7. The EEG report generation method combining a hybrid neural network and an empirical analysis algorithm according to claim 1, characterized in that: The regional analysis module further analyzes the alpha rhythm in the occipital region, the mu rhythm in the central region, and the theta rhythm in the frontal midline region based on the descriptive data of brain wave activity. For the alpha rhythm, the frequency range, amplitude range, symmetry, and alpha index of the regional channel corresponding to the rhythm are calculated as quantitative features. For the μ rhythm, the frequency range, amplitude range, and symmetry of the regional channel corresponding to the rhythm are calculated as quantitative features; for the θ rhythm, the frequency range and amplitude range of the regional channel corresponding to the rhythm are calculated as quantitative features; The quantitative characteristics of these three rhythms are then used to generate descriptive text about background activities according to fixed templates and descriptive terms.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: The processor is used to execute the computer program to implement the EEG report generation method combining a hybrid neural network and an analysis experience algorithm as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the EEG report generation method combining a hybrid neural network and an analysis experience algorithm as described in any one of claims 1 to 7.

Images