A hybrid deep learning-based classification system for children's viral encephalitis

Abstract

The application discloses a kind of classification systems of children viral encephalitis based on hybrid deep learning, wherein data acquisition module acquires the electroencephalogram data of clinical children viral encephalitis patient;Data preprocessing module carries out filtering processing to artifact and noise of electroencephalogram data;Data segmentation module carries out wavelet packet transform to the data after filtering, extracts electroencephalogram rhythm signal;Feature representation module converts continuous electroencephalogram rhythm time domain signal to time-frequency domain by continuous wavelet transform, and represents characteristics by time-frequency TF image;Feature extraction module extracts the time-space-frequency fusion features represented in TF image;Training classification module inputs different domain fusion features into feature extraction module, trains the neural network in feature extraction module, and obtains the classification result of different symptoms VE patient.The application solves the problem of frequency band difference between patients of different degrees, and improves the efficiency and accuracy of classification.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence, specifically relating to a classification system for viral encephalitis in children based on hybrid deep learning. Background Technology

[0002] Encephalitis is a serious disease caused by inflammation of the brain parenchyma, leading to neurological dysfunction. It is a common pediatric disease requiring timely diagnosis and treatment. Delayed treatment or misdiagnosis can severely threaten a child's life and health. Therefore, early identification and warning of severe encephalitis are of significant clinical importance. Electroencephalography (EEG) is a common diagnostic method for viral encephalitis due to its non-invasiveness, low cost, and ability to accurately assess the degree of brain function damage in the early stages. The more severe the clinical symptoms, the higher the rate and the more pronounced the abnormalities on the EEG. Therefore, the study of patients' EEG signals is of great value in the detection of viral encephalitis.

[0003] Currently, EEG analysis for viral encephalitis (VE) in clinical practice still relies on observation and analysis by professional physicians based on their experience and expertise. This process is not only time-consuming and labor-intensive but also highly subjective. Because viral encephalitis progresses rapidly, a large number of severe cases occur because doctors fail to detect abnormalities in early EEG changes, leading to delays in clinical treatment and ultimately the development of severe encephalitis.

[0004] With the development of artificial intelligence and deep learning, researchers in the medical field have already used AI algorithms to automatically analyze and identify EEG signals from encephalitis. Luca et al. fused the time-domain features related to the slow biphasic complex (SBC) in the EEG of encephalitis patients with the frequency-domain features of the relative power spectrum of EEG rhythms in different frequency bands, and used the optimal binary tree model (BTM) to train the extracted features, finally classifying patients with different degrees of encephalitis.

[0005] However, despite the good results achieved by existing EEG classification methods for childhood encephalitis, conventional deep learning methods such as convolutional neural networks cannot retain the "memory" of adjacent time series models when extracting features from continuous, large-scale EEG signals. They are prone to ignoring certain important feature information and cannot comprehensively analyze the changes in features between adjacent EEG segments, leading to a decrease in classification accuracy. Summary of the Invention

[0006] To address the shortcomings of existing methods, this invention proposes a hybrid deep learning-based classification system for viral encephalitis in children. This system can effectively identify and correctly classify EEG signals from children with viral encephalitis exhibiting different symptoms, achieving high accuracy and efficiency. It also eliminates cognitive differences and subjective influences among clinicians, making the classification results more objective.

[0007] A classification system for viral encephalitis in children based on hybrid deep learning includes: a data acquisition module, a data preprocessing module, a data segmentation module, a feature representation module, a feature extraction module, and a training classification module.

[0008] The data acquisition module is used to collect EEG signal data from pediatric VE patients in clinical practice, obtaining EEG data from patients with different symptoms.

[0009] The data preprocessing module is used to filter artifacts and noise from the raw EEG data of VE patients.

[0010] The data segmentation module is used to extract low-frequency EEG rhythm signals from the filtered EEG data using wavelet packet transform (WPT), and then segment the EEG rhythm signals to form fixed-length EEG rhythm segment data.

[0011] The feature representation module is used to convert the continuous EEG rhythm time-domain signal to the time-frequency domain through continuous wavelet transform (CWT) and represent the features through a time-frequency (TF) image.

[0012] The feature extraction module is used to extract the temporal-spatial-frequency fusion features represented in TF images.

[0013] The training classification module is used to input the extracted features from different domains into the feature extraction module to train the neural network in the feature extraction module and obtain the classification results of VE patients with different symptoms.

[0014] Furthermore, the data acquisition module is specifically used to: collect electroencephalogram (EEG) scalp recording data of children with persistent VE episodes in clinical practice, and obtain the original EEG time-series signals of children with mild and severe encephalitis.

[0015] Furthermore, the data preprocessing module is specifically used to: perform noise filtering on the acquired EEG signal, extract the 1-48Hz EEG signal using a bandpass filter, and remove interference from power frequency noise such as eye movement and artifacts.

[0016] Furthermore, the data segmentation module is specifically used for:

[0017] Wavelet packet transform was used to extract the low-frequency EEG rhythms of the EEG signal, including three subbands: Alpha (8-12Hz), Theta (4-8Hz), and Delta (1-4Hz).

[0018] A fixed-size sliding time window is used to segment the EEG rhythm signal into non-overlapping segments. Each patient's EEG data sample S is divided into n segments of duration T, resulting in n data sets for each sample S, i.e., {S1, S2, ..., S...}. n}

[0019] Furthermore, the feature representation module is specifically used to: in order to gain a deeper understanding of the signal, convert the segmented EEG rhythm time-domain signal into a time-frequency domain TF image using CWT, simultaneously capturing the details of the time-domain signal in the frequency domain. The TF image can display the spectral transformation of the signal over time, allowing simultaneous observation of the signal in both the time and frequency domains, and fully representing the characteristics of different domains.

[0020] Furthermore, the feature extraction module includes a spatial-frequency feature extraction unit and a sequence feature extraction unit:

[0021] (1) The spatial-frequency feature extraction unit learns the frequency-spatial features in the input TF feature image by designing a convolutional neural network (CNN) sub-module containing three "learners". Finally, it flattens the two-dimensional feature matrix and outputs a one-dimensional feature vector for input to the serialization feature extraction module.

[0022] Furthermore, the specific architecture of the three "learners" CNN sub-modules consists of, in sequence, a first convolutional layer, a normalization layer, a first pooling layer, a dropout layer, a second convolutional layer, a normalization layer, a second pooling layer, a dropout layer, a third convolutional layer, a normalization layer, an SE module, a third pooling layer, a dropout layer, and a Flatten layer.

[0023] (2) Sequence feature extraction unit: It extracts the relationship change features between adjacent features through bidirectional long short-term memory network Bi-LSTM, and effectively combines them with the features extracted by the spatial-frequency feature extraction module, so that they are completely input into the classification module for training and classification.

[0024] Furthermore, the specific architecture of the sequence feature extraction unit consists of a Bi-LSTM layer, a Dropout layer, another Bi-LSTM layer, and a fully connected layer.

[0025] Furthermore, the training classification module is specifically used to: train the neural network in the feature extraction module; during training, cross-entropy is used as the loss function to calculate the loss value; and the Adam gradient descent algorithm is used to find the minimum loss value to update the network parameters. The trained neural network is then input with EEG signal data of different symptoms to obtain the output classification result.

[0026] The present invention has the following beneficial effects:

[0027] The method of this invention analyzes EEG signals of viral encephalitis. Against the backdrop of significant differences in the frequency bands exhibited by encephalitis of different severities, it solves the problem of frequency band differences between patients with different severities by extracting multiple low-frequency EEG rhythms such as δ, θ, and α.

[0028] This invention utilizes time-frequency feature images to represent EEG data based on the multi-dimensional features of EEG signals, such as time, space, and frequency. This fully leverages the feature information across different dimensions and avoids the accuracy drop caused by insufficient feature extraction.

[0029] This invention addresses the challenge of handling large volumes of time-frequency feature images of encephalitis by designing a hybrid deep learning model that combines CNN and Bi-LSTM. This model effectively extracts feature information from different dimensions within the feature images and performs feature fusion, significantly improving the efficiency and accuracy of classification. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the structural process of the present invention;

[0031] Figure 2 This is a channel distribution diagram of the EEG acquisition device of the present invention;

[0032] Figure 3 This is a structural diagram of the hybrid deep learning model proposed in this invention;

[0033] Figure 4 This is a comparison chart of classification indices for waveforms in different frequency bands using the algorithm of this invention. Detailed Implementation

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments: This embodiment is implemented under the premise of the technical solution of the present invention, and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0035] like Figure 1 As shown, a classification system for viral encephalitis in children based on hybrid deep learning is provided, the system comprising:

[0036] The data acquisition module is used to collect raw EEG time-series signal data from clinical pediatric VE patients.

[0037] Specifically, the data used in this method were electroencephalogram (EEG) data from patients with viral encephalitis at the Children's Hospital Affiliated to Zhejiang University School of Medicine. The dataset collected EEG signals from 30 patients with viral encephalitis, including 15 patients with mild encephalitis and 15 patients with severe encephalitis, aged between 6 and 12 years.

[0038] In the experiment, clinical patients were continuously monitored for several days, and scalp electroencephalogram (EEG) data of patients with viral encephalitis experiencing continuous seizures were recorded. During data acquisition, measurements were taken using a monopolar lead method, and recording and labeling were performed according to the internationally standardized 10-20 EEG electrode placement and naming system, with a sampling frequency of 256 Hz. Due to variations in the electrodes used to record EEG data from patients, 21 common electrodes (Fp1, Fp2, F3, F4, C3, C4, P3, P4, Fz, Cz, Pz, F7, F8, T3, T4, T5, T6, A1, A2, O1, O2) were selected as experimental electrodes for ease of data analysis. Figure 2 As shown.

[0039] The data preprocessing module is used to perform noise filtering on the acquired raw EEG time-series signals.

[0040] Specifically, because clinically acquired EEG signals can be affected by artifacts caused by patient movement or body posture, as well as noise added during the acquisition and transmission from the scalp of patients with viral encephalitis, this type of EEG signal may bias data analysis and lead to erroneous conclusions.

[0041] The data preprocessing module uses a bandpass filter to filter the given EEG signal, removing unwanted frequency components. The filter range is 0.1-48Hz, minimizing noise interference with the research results.

[0042] The data segmentation module is used to extract low-frequency EEG rhythm signals from the filtered signal by wavelet packet transform (WPT), and then segment the signal to form fixed-length EEG rhythm fragment data.

[0043] Specifically, since the EEG signals corresponding to different severities of encephalitis show significant differences across different frequency bands, this method extracts the EEG rhythm signals from the δ, θ, and α sub-bands in the low-frequency range of the EEG signal, enabling them to clearly characterize the features of encephalitis patients with different symptoms. Wavelet packet transform is a time-scale analysis method with multi-resolution analysis capabilities, capable of decomposing both low-frequency and high-frequency components of the signal. This decomposition is neither redundant nor incomplete, allowing for better time-frequency localization analysis of the signal. The filtered EEG signal was subjected to a 5-level wavelet packet decomposition using the Daubechies 4 (db4) wavelet function. In the 5th level decomposition, the EEG rhythm signals from the Alpha (8-12Hz), Theta (4-8Hz), and Delta (1-4Hz) sub-bands were selected.

[0044] To reduce computational complexity and achieve a balance between capturing specific patterns and maintaining stability, a sliding time window was designed to process the experimental data after extracting low-frequency EEG rhythm signals, extracting EEG segments from continuous signals. Long-duration EEG rhythm signals were divided into smaller duration segments using a fixed-size window, with a sliding time window length of 5 seconds and no overlap between adjacent segments. Each patient's EEG data sample S was equally divided into n segments of duration T, resulting in n data sets for each sample S, i.e., {S1, S2, ..., S...}. n The length T of the data in each set is 5.

[0045] The feature representation module is used to convert the continuous EEG rhythm time-domain signal to the time-frequency domain and represent the features through a time-frequency (TF) image.

[0046] Specifically, because EEG signals contain amplitude and frequency components with violent oscillations, they are difficult to analyze directly. Furthermore, the degree of abnormality in EEG signals from encephalitis is closely related to low-frequency EEG rhythms such as alpha, theta, and delta. Therefore, to gain a deeper understanding of the signal, it is necessary to convert the EEG rhythm signal from one domain to another. Time-frequency transduction (CWT) can simultaneously capture the details of the time-domain signal into the frequency domain. Time-frequency representation can show the spectral changes of the signal over time, allowing observation of the signal in both the time and frequency domains simultaneously. This is particularly useful for continuous, non-steady EEG signals.

[0047] Considering these facts, the extracted EEG rhythm signals of the three frequency bands were processed. After being segmented into EEG rhythm signal segments by sliding time windows, time-frequency processing was performed using CWT to convert them into EEG rhythm TF image representations.

[0048] For EEG rhythm signals (f∈R), use the continuous wavelet transform function of f with scaling and shift parameters:

[0049]

[0050] Through this transformation, a one-dimensional time-series signal f(t) can be converted into a two-dimensional form Wf(u,v), also known as a scaling plot. These scaling plots are used to detect and represent the most significant frequencies (scales) of the signal in its time-scale representation. The scaling plot of the signal f(t) can be calculated using the following formula:

[0051]

[0052] The function φ represents the energy scale map of signal Wf at scale u and time position v. Wavelet transform has various families of wavelet functions, which differ from each other based on compactness and smoothness. In the proposed method, the Morlet wavelet (cmor3-3) is used as the mother wavelet for continuous wavelet transform, which can extract features with equal variance in frequency and time. The general mathematical form of the Morlet wavelet is:

[0053]

[0054] Where e is the natural logarithm, C is a constant, and Morlet wavelet is a Gaussian-weighted sine curve that can extract the time features of various signals and adapt to their time-frequency resolution.

[0055] The feature extraction module is used to extract the temporal-spatial-frequency fusion features represented in TF images.

[0056] The detailed structure diagram of the module is as follows: Figure 3 As shown, it is divided into two parts: a spatial-frequency feature extraction unit and a sequence feature extraction unit. The spatial-frequency feature extraction unit can automatically select representative EEG channels to extract depth features. Not only are the frequency and spatial features of each segment important for the analysis of time-frequency images, but since these images are generated along continuous EEG segments, the time-series features contained in adjacent images can also provide very useful information. Therefore, a sequence feature extraction unit is added after the spatial-frequency feature extraction unit to learn the serialized features contained in continuous time-frequency image segments. The frequency and spatial features are fused to improve classification accuracy.

[0057] The spatial-frequency feature extraction unit comprises three sub-blocks, each consisting of a two-dimensional kernel "learner": a long-time learner, a spatial learner, and a short-time learner. The learners primarily include convolutional layers (Conv), batch normalization layers (BN), max-pooling layers, and dropout layers. The third block's learner also incorporates a SE module, which suppresses currently useless features by setting weights for each feature channel. Each part of the learner has a different structure and parameters, sequentially learning the frequency-spatial features from the input feature image. Through multiple learners, hidden features in the EEG time-frequency image can be automatically learned.

[0058] In the first block, the learner size is set to (5,5), and the input image size is uniformly adjusted to 128*128*3 (W*H*C), where W, H, and C represent the width, height, and number of channels of the input image, respectively. Data normalization is performed to accelerate deep network training by reducing internal covariate bias. Let X be the input time-frequency feature image of encephalitis EEG. Where N is the number of EEG image segments, C is the frequency band of the EEG signal, and T is the length of the EEG image segment. The process of Block 1 can be described as follows:

[0059] Z1=f drop (f bn (Conv 2D(X,F1)))

[0060] Where Z1 is the feature extracted from the first block, Conv 2D(·) is a 2D convolution operation with a long kernel F1, and f bn (·) is the batch normalization operation. drop (·) represents the dropout neuron function, using a dropout rate p of 0.25 to improve the model's generalization performance and avoid overfitting.

[0061] In the second block, a spatial learner of size (4,4) is used to learn spatial filters inspired by applications of deep convolutional neural networks in computer vision. The spatial learner not only reduces the number of trainable parameters to be fitted but also provides a direct method for learning spatial filters for each temporal filter. The features Z1 extracted from the first block are input into the second block for feature extraction, as follows:

[0062] Z2 = f drop (σ(f bn (Conv2D(Z1,F2))))

[0063] Where σ(·) represents the activation operation, and the ReLU activation function is used here.

[0064] In the third block, inspired by depthwise convolutions in EEGNet, a learner with short-time kernels is used to extract features. This reduces the number of parameters to be fitted and allows for the exploration of deep features within the feature map. In block 3, the feature learning representation is as follows:

[0065] Z3 = f drop (σ(f bn (Conv2D(Z2,F3))))

[0066] The parameters of the third block are the same as those of the second block. In addition, a compression and excitation (SE) module is added to the third block to highlight feature information. The SE module is a channel-based attention mechanism that adaptively calibrates feature weights in different channels by modeling interdependencies.

[0067] At the end of the spatial-frequency feature extraction unit, the extracted two-dimensional feature matrix is ​​flattened by the Flatten layer into a one-dimensional feature vector, which can then be input into the sequence feature extraction module for serialized feature extraction.

[0068] In the sequence feature extraction unit, the feature vector x after the Flatten layer is flattened t The input is fed into an LSTM unit, which consists of an input layer, hidden layers, and an output layer. The input layer receives a one-dimensional feature vector output from the spatial-frequency feature extraction module, and the hidden layers store the feature x. t (x here) t This refers to Z3 mentioned earlier, and the output layer is used to output the classification results of viral encephalitis. The specific steps are as follows:

[0069] f t =σ(ω) f [h t-1 ,x t ]+b f )

[0070] Where, x t There are n time series X = (x1, x2, ..., xn) n The characteristics of time t in ) h t-1 Let f represent the hidden state output of the previous Bi-LSTM unit, ω and b represent the weight matrix and bias vector, respectively, and σ is the sigmoid function that determines the degree to which information is forgotten. In the forgetting gate, f t The sigmoid function is used to determine which information is discarded from the previous cell state. The selected features are then fed into the input gate.

[0071] i t =σ(ω) i [ht-1 ,x t ]+b i )

[0072]

[0073]

[0074] in, It is for x t and h t-1 The normalized temporary state, c t-1 It represents the state of the previous unit. In the input gate, c t The process involves merging the acquired temporary information with the information output by the forget gate to obtain a new cell state.

[0075] o t =σ(ω) o [h t-1 ,x t ]+b o )

[0076] h t =o t ·tanh(c t )

[0077] Among them, h t This is the hidden state of the current cell, and also the output. In the output gate, we use the tanh activation function to activate the output h. t Perform nonlinear mapping.

[0078] Finally, the time-series features extracted by the Bi-LSTM unit and the spatial-frequency features from the previous part are input together into a fully connected layer (FC) to achieve the combination and mapping between different features. For the features X = [x1, x2, ..., x] output by the LSTM unit... n If ] represents the spatial-frequency-temporal features of n time-frequency images, then the output Y after the features pass through a fully connected layer is:

[0079] y = f(W*x + b)

[0080]

[0081] Here, f(·) is the Softmax activation function that assigns a specific probability to each category to classify patients with mild encephalitis and severe encephalitis.

[0082] In the sequence feature extraction unit, Bi-LSTM modules containing 128 and 64 LSTM units respectively were used to learn sequence features.

[0083] The training classification module is used to input the extracted fusion features from different domains into the proposed hybrid deep learning classification model for training, so as to obtain the classification results of VE patients with different symptoms.

[0084] Specifically, the EEG rhythm TF image dataset of 30 children was divided into three sets: 60% as the training set, 20% as the validation set, and 20% as the test set. The samples from each set were then randomly shuffled.

[0085] The hybrid deep learning model is trained using labeled training data, validated using validation data, and then used to classify the test set, that is, to classify each sample in the test set to distinguish between patients with mild encephalitis and those with severe encephalitis.

[0086] The model uses the following parameter settings during training. Binary cross-entropy is chosen as the loss function to evaluate model performance. After calculating the loss, gradient descent is used to find the minimum loss value, thereby updating the network parameters and optimizing the network model. This invention uses the Adaptive Momentum Estimation (Adam) optimizer to optimize the parameters, with a learning rate of 3 × 10⁻⁶. -4 The network model was trained for 200 epochs, with a batch size of 80.

[0087] Finally, to evaluate the performance of the proposed hybrid deep learning model, values ​​for accuracy, precision, sensitivity, specificity, and classification performance were calculated:

[0088] The accuracy rate (ACC) is as follows:

[0089]

[0090] Precision (Pre) is:

[0091]

[0092] Sensitivity (Sen) is:

[0093]

[0094] Specificity (Spe) is:

[0095]

[0096] The classification performance value (F1) is:

[0097]

[0098] In the formula, TP, FP, TN, and FN represent true positives, false positives, true negatives, and false negatives, respectively.

[0099] Training and testing were performed on the δ, θ, and α rhythm time-frequency images in the low-frequency band, as shown in Table 1. Figure 4 The classification results of the three rhythms on the proposed model are presented. It can be seen that the average accuracy, precision, sensitivity, specificity, and F1 score of the three rhythms in the proposed model are 90.60%, 90.52%, 90.52%, 90.78%, and 90.32%, respectively. Among them, the delta rhythm shows the best results in all indicators, with accuracy, precision, sensitivity, specificity, and F1 score of 94.18%, 94.20%, 94.37%, 94.52%, and 94.15%, respectively, which is superior to the training results of the theta and alpha frequency band rhythm images. This result demonstrates that lower EEG rhythm frequency bands are more effective in analyzing encephalitis of different severities and can accurately distinguish encephalitis patients with different symptoms.

[0100] Table 1 shows the performance of the proposed model across different rhythms (in %).

[0101]

[0102] In summary, the proposed classification system for childhood viral encephalitis based on a hybrid deep learning model combines CNN and Bi-LSTM network models to construct an end-to-end network. The fusion of these two networks improves the accuracy and stability of the algorithm. This method, trained in a supervised manner, can accurately and efficiently classify EEG signals from patients with mild and severe encephalitis.

[0103] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A classification system for viral encephalitis in children based on hybrid deep learning, characterized in that, It includes a data acquisition module, a data preprocessing module, a data segmentation module, a feature representation module, a feature extraction module, and a training classification module; The data acquisition module is used to collect electroencephalogram (EEG) signal data of clinical pediatric patients with viral encephalitis (VE) to obtain EEG data of patients with different symptoms. The data preprocessing module is used to filter EEG data for artifacts and noise. The data segmentation module is used to perform wavelet packet transform on the filtered EEG data, extract the EEG rhythm signal, and segment the EEG rhythm signal to form fixed-length EEG rhythm segment data. Specifically, it is implemented as follows: Wavelet packet transform was used to extract the EEG rhythm of the signal, which includes three subbands: Alpha (8-12Hz), Theta (4-8Hz), and Delta (1-4Hz). Using a fixed-size sliding time window, the EEG rhythm signals were segmented non-overlappingly, dividing the EEG data samples from each patient. Divide into equal parts The duration is The data, thus for each sample get A data set, namely ; The feature representation module is used to convert the continuous EEG rhythm time-domain signal to the time-frequency domain through continuous wavelet transform (CWT) and represent the features through a time-frequency (TF) image. The feature extraction module is used to extract the temporal-spatial-frequency fusion features represented in the TF image, including a spatial-frequency feature extraction unit and a sequence feature extraction unit; The spatial-frequency feature extraction unit learns the frequency-spatial features in the input TF feature image sequentially through a convolutional neural network (CNN) submodule containing three learners, and finally flattens the two-dimensional feature matrix to output a one-dimensional feature vector for input to the serialization feature extraction module. The sequence feature extraction unit extracts the relationship change features between adjacent features through a bidirectional long short-term memory network (Bi-LSTM), and combines them with the features extracted by the spatial-frequency feature extraction unit, and inputs them completely into the classification module for training and classification. The training classification module is used to input the extracted features from different domains into the feature extraction module to train the neural network in the feature extraction module, thereby obtaining the classification results of VE patients with different symptoms.

2. The classification system for viral encephalitis in children based on hybrid deep learning according to claim 1, characterized in that, The data preprocessing module is specifically implemented as follows: noise filtering is performed on the acquired EEG signal, and a bandpass filter is used to extract the 1-48Hz EEG signal to remove power frequency noise interference from eye movement and artifacts.

3. The classification system for childhood viral encephalitis based on hybrid deep learning according to claim 2, characterized in that, The specific architecture of the CNN sub-modules of the three learners consists of, in sequence, a first convolutional layer, a normalization layer, a first pooling layer, a dropout layer, a second convolutional layer, a normalization layer, a second pooling layer, a dropout layer, a third convolutional layer, a normalization layer, an SE module, a third pooling layer, a dropout layer, and a Flatten layer.

4. The classification system for childhood viral encephalitis based on hybrid deep learning according to claim 3, characterized in that, The specific architecture of the sequence feature extraction unit consists of a Bi-LSTM layer, a Dropout layer, a Bi-LSTM layer, and a fully connected layer.

5. The classification system for childhood viral encephalitis based on hybrid deep learning according to any one of claims 1 to 4, characterized in that, The training classification module is specifically implemented as follows: training the deep neural network of the feature extraction module, using cross-entropy as the loss function to calculate the loss value during the training process, and using the Adam gradient descent algorithm to find the minimum loss, updating the network parameters, and completing the training.

Images