CN-122000022-A - Hypothalamic hamartoma epileptic subtype identification method based on interpretable graph neural network

CN122000022ACN 122000022 ACN122000022 ACN 122000022ACN-122000022-A

Abstract

The invention discloses a hypothalamic hamartoma epilepsy subtype identification method based on an interpretable graph neural network, and belongs to the technical field of epilepsy diagnosis. Acquiring preoperative electroencephalogram signals of patients with hypothalamic hamartoma accompanied by epilepsia, preprocessing and extracting features to obtain time-frequency feature data, inputting the time-frequency feature data into an ST-AttnNet model, cooperatively capturing spatial topological features, multi-scale time-frequency features and dynamic time sequence features of brain functional connection through a trunk layer, a graph attention network module, a multi-scale frequency-time attention module and a two-way long-short-term memory network, realizing subtype identification of giggle type epilepsia and non-smiling epilepsia, extracting an adjacent matrix after training convergence of a GAT module, and mining potential biomarkers through network topology quantitative analysis. The method breaks through the limitation of traditional subjective identification, improves identification accuracy and stability through a multi-module collaborative architecture, provides objective basis for clinic, and has important clinical conversion value.

Inventors

JIANG BIN
CHEN YIN
LIU WEI
ZHAO YUN
HE DONGYI
XIA QINGLING
ZHOU ZHIHONG
WANG XIANGKAI
LI MINGDA

Assignees

重庆理工大学

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (10)

1. A method for identifying hypothalamic hamartoma epileptic subtypes based on an interpretable graph neural network, comprising: acquiring preoperative electroencephalogram signals of patients with hypothalamic hamartoma and epilepsy, and preprocessing and extracting features of the electroencephalogram signals to obtain time-frequency feature data; Inputting the time-frequency characteristic data into an ST-AttnNet model, and cooperatively capturing the spatial topological characteristic, the multi-scale time-frequency characteristic and the dynamic time sequence characteristic of the brain function connection through a trunk layer, a drawing meaning network module, a multi-scale frequency-time attention module and a two-way long-short-term memory network of the model; Based on the output result of the model, the subtype identification of giggle epilepsy related to hypothalamic hamartoma and non-smiling epilepsy is realized; extracting an adjacency matrix after the GAT module is trained and converged, and mining potential biomarkers for distinguishing two types of epileptic subtypes through network topology quantitative analysis.
2. The method for identifying hypothalamic hamartoma epileptic subtype based on the interpretable graphic neural network according to claim 1, wherein the electroencephalogram signals are recorded by a multi-channel electroencephalogram acquisition device, and channel signals covering frontal pole, frontal lobe, central area, parietal lobe, temporal lobe, occipital lobe and midline area are covered, so that full electroencephalogram physiological activities are comprehensively captured; Synchronously recording the sleep state of a patient in the acquisition process, intercepting a signal segment without epileptic burst during the deep sleep, and avoiding the interference of abnormal discharge in the epileptic burst on the characteristic extraction and classification result; the sampling rate of the electroencephalogram acquisition equipment is set according to the frequency characteristic of nerve discharge signals, so that electrophysiological activities in a preset frequency band can be completely captured, the acquired original signals are subjected to anti-interference treatment, the influences of physiological artifacts such as myoelectricity and electrooculography and environmental noise are reduced, and the data quality is ensured.
3. The method for identifying hypothalamic hamartoma epileptic subtype based on the interpretable graphic neural network according to claim 1, wherein the preprocessing and the feature extraction of the electroencephalogram signals comprise the following steps: The electroencephalogram signals are subjected to sectional processing according to a set window length and a step length, after slow drift is eliminated through baseline correction, high-pass filtering and low-pass filtering are sequentially carried out by adopting a cascading digital filtering strategy, and band-pass filtering signals of a preset frequency band are obtained; Performing short-time Fourier transform on the filtered signals, inhibiting spectrum leakage through a hanning window, and converting time domain signals into time-frequency characteristic data, wherein the mathematical formula of the short-time Fourier transform is as follows: , wherein, Is the time-frequency coefficient of the signal, Representing at discrete points in time A sampling value is obtained at the location, Is a hanning window function, which is a hanning window function, For the fast fourier transform window length, The step size of the overlap is chosen to be the same, For the segment sequence numbers corresponding to the timing dimension, Discrete frequency sequence number for frequency domain dimension; After the absolute value conversion is carried out on the time-frequency characteristic data, the time-frequency characteristic data is mapped to a preset interval by adopting Z-fraction normalization processing, and a normalization formula is as follows: , wherein, In order to make the data after the normalization, Is the absolute value of the original time-frequency coefficient, Is the mean value of the original time-frequency data, Is the standard deviation of the original time-frequency data.
4. The hypothalamic hamartoma epileptic subtype identification method based on the interpretable graph neural network, which is disclosed by claim 1, is characterized in that the trunk layer comprises a depth separable convolution layer, a batch normalization layer and GELU activation functions, the convolution kernel size of the depth separable convolution layer is designed according to the dimension adaptation of time-frequency characteristic data, the channel dimension and the space dimension of input characteristics are separated, key characteristic information is reserved while the complexity of model parameters is reduced, the batch normalization layer performs standardization processing on convolution output characteristics to eliminate the influence of characteristic distribution offset on model training, and the GELU activation functions enhance the fitting capacity of the model on complex characteristics through nonlinear transformation to realize the preliminary extraction and dimension reduction of the time-frequency characteristics and provide low-dimension and high-recognition characteristic input for subsequent modules.
5. The method for identifying hypothalamic hamartoma epileptic subtype based on interpretable graphic neural network according to claim 4, wherein the working process of the graphic force network module comprises the following steps: based on the feature vectors output by the trunk layer, constructing node feature matrixes corresponding to all channels of the electroencephalogram, and initializing a learnable adjacent matrix to represent potential functional connection relations among the channels; Calculating attention coefficients among nodes through a self-attention mechanism, dynamically evaluating the importance of each connection based on the feature similarity, carrying out normalization processing on the attention coefficients through a Softmax function, and adaptively adjusting adjacent matrix element values to generate a dynamic brain network reflecting the functional coupling strength among channels; Any two channel nodes in the adjacency matrix And (3) with Functional connection strength of (a) Defined as the attention coefficient after the min-max normalization process The formula is: , wherein, For the normalized connection strength, the connection strength is calculated, As the original attention coefficient(s), And Respectively minimum and maximum values in the attention coefficient matrix; The node characteristics weighted by the attention are processed through linear transformation and an activation function and then transmitted to a downstream module, and the adjacency matrix after training convergence is used for subsequent network topology analysis.
6. The method for identifying hypothalamic hamartoma epileptic subtypes based on interpretable graphic neural networks according to claim 5, wherein the multi-scale frequency-time attention module includes a main path connected with a residual, the main path including a time maximizing pooling layer, a multi-scale depth separable convolution unit, and a channel-frequency attention module in order; the time maximization layer performs time sequence dimension downsampling on the input features, and key time sequence information is reserved; The separable convolution unit extracts different receptive field features through four parallel branches, wherein the four branches are respectively 1×1 point-by-point convolution branches, 3×3 depth separable convolution branches connected with 1×1 point-by-point convolution branches, 1×1 point-by-point convolution branches connected with 3×3 maximum pooling branches, two cascaded 3×3 depth separable convolution branches, and the outputs of the branches are spliced in the channel dimension to realize the fusion of multi-scale context information; The frequency attention module calculates frequency attention and channel attention in parallel, a frequency path learns key frequency band weights through one-dimensional convolution and an activation function, a channel path adopts a Squeeze-and-specification mechanism, channel importance weights are generated through global average pooling, a full connection layer and the activation function, input features are multiplied by the two attention weights element by element, and double enhancement is completed, wherein the formula is as follows: , wherein, In order to input the characteristics of the feature, For the channel attention weighting to be used, As the weight of the attention to the frequency, Representing multiplication by element; And after the channel dimension is adjusted through 1X 1 point-by-point convolution, the residual connection is added with the output characteristics of the CFA module, so that the gradient degradation problem of the deep network is relieved, and the smooth transfer of information is ensured.
7. The method for identifying hypothalamic hamartoma epileptic subtype based on the interpretable graph neural network according to claim 6, wherein the two-way long-short-term memory network comprises a forward LSTM layer and a backward LSTM layer, and the characteristic graph processed by the MS-CFA module is subjected to frequency dimension global average pooling and transposed to form a time sequence input; The Bi-LSTM output is mapped to a class space through a full connection layer, probability distribution of each class is output through a Softmax activation function, class corresponding to the maximum probability is used as subtype identification result, giggle epilepsy and non-smile epilepsy are classified, a cross-validation and early stop method is adopted in the model training process, overfitting is avoided, and model generalization capability is improved.
8. The method for identifying hypothalamic hamartoma epileptic subtypes based on the interpretable graph neural network according to claim 1, wherein the network topology quantitative analysis comprises node clustering coefficient calculation and inter-channel connection strength analysis, and the difference significance of the two types of epileptic subtypes on the indexes is evaluated by adopting paired t test; the node clustering coefficient The formula for characterizing the local aggregations of the nodes is: , wherein, Is a node Is used to determine the number of neighbor nodes, And Traversing nodes Is provided for the pair of all the neighboring nodes, 、 And Respectively normalizing function connection weights of corresponding node pairs, wherein the node pairs jointly form nodes Is a triangular structure; and screening out the connection strength index and the node clustering coefficient with obvious difference through statistical test, and taking the connection strength index and the node clustering coefficient as potential biomarker candidate sets for distinguishing two types of epileptic subtypes.
9. The method of identifying hypothalamic hamartoma epileptic subtypes based on interpretable graphics neural network according to claim 8, wherein the screening process of the potential biomarker includes: performing multiple test correction on the connection strength index and the node clustering coefficient in the candidate set, and eliminating the influence of false positive results; Calculating the contribution degree of each candidate index to subtype identification through a feature importance evaluation method, and selecting TopN indexes after sequencing according to the contribution degree; and (3) combining neurophysiologic mechanism analysis, removing indexes without explicit physiological significance, and finally forming a biomarker set containing specific inter-channel connection strength and key node clustering coefficients, wherein the biomarker set can reflect the essential difference of two types of epileptic subtypes on a brain network topological structure.
10. The method for identifying hypothalamic hamartoma epileptic subtypes based on interpretable graphic neural networks according to claim 1, characterized in that the mining distinguishes potential biomarkers of two classes of epileptic subtypes, comprising: Based on the training convergence post-adjacency matrix output by the GAT module, respectively calculating intra-group average adjacency matrices of giggle epilepsy groups and non-smiling epilepsy groups, and constructing two groups of representative brain function connection diagrams; obtaining two groups of differential connection matrixes by element-by-element subtraction, screening out channel pairs with obvious connection strength differences, and forming a differential connection set; Combining the inter-group difference analysis results of the node clustering coefficients, and integrating the node clustering coefficients with obvious differences with channel pairs in the difference connection set; redundant information is removed through a feature selection algorithm, core indexes with strong discrimination are reserved, and finally potential biomarkers capable of effectively distinguishing two types of epileptic subtypes are obtained, so that objective electrophysiological basis is provided for clinical discrimination.

Description

Hypothalamic hamartoma epileptic subtype identification method based on interpretable graph neural network Technical Field The invention relates to the technical field of epilepsy diagnosis, in particular to identification of preoperative electroencephalogram signals and biomarker mining of patients with hypothalamic hamartoma and epilepsy based on an ST-AttnNet model. Background Hypothalamic Hamartoma (HH) is a rare but serious congenital brain lesion, often causing seizures directly as a primary epileptogenic focus, or by inducing secondary epileptogenic focus to cause seizures, with a incidence of about one twentieth. Clinically, epilepsy related to HH is mainly divided into giggle epilepsy (GS) and non-smiling epilepsy (NGS), and whether giggle symptoms of patients appear or not is not only a key basis of disease typing, but also is closely related to disease evolution, treatment scheme selection and prognosis effect. The accurate identification of the two epileptic subtypes has important significance for improving the pertinence of surgical treatment and improving the prognosis of patients. However, the current clinical preoperative evaluation mainly depends on manual epileptic diary records of patients and nursing staff, and the subjective symptom description-based mode is easily influenced by factors such as individual cognitive difference, record timeliness and the like, and is difficult to objectively and accurately capture fine characteristics of epileptic seizures, so that the subtype identification is extremely challenging. Imaging technology is the main means of HH focus detection and analysis at present, and the technology including Magnetic Resonance Imaging (MRI) and the like can effectively identify the structural morphological change of HH, thereby providing support for clinical diagnosis, operation planning and curative effect evaluation. However, for the GS subtype and NGS subtype without obvious structural difference, the imaging technology is difficult to capture the essential difference, and the requirement of clinical accurate typing cannot be met. In contrast, electroencephalogram (EEG) can directly capture abnormal synchronous discharge processes of brain neurons, has the unique advantages of high time resolution and quantitative evaluation of cerebral cortex function changes, and has been widely used in epileptic localization, classification and diagnosis. However, in the existing research of HH epileptic subtype based on EEG, most of the research relies on manual design characteristics to carry out statistical analysis, and excessive focusing is carried out on temporal lobe areas related to HH pathology, so that complex nerve modes which are difficult to manually extract are easily omitted, key effects of multi-brain area interaction under the whole brain network view angle are ignored, and high-dimensional and complex nerve activity differences between GS and NGS subtypes are difficult to comprehensively capture. With the rapid development of deep learning technology, the method based on the structures such as Convolutional Neural Network (CNN), cyclic neural network (RNN) and transducer shows strong high-dimensional space-time pattern recognition and long-term dependency modeling capability in EEG signal processing, and the epileptic focus positioning accuracy and real-time early warning accuracy are remarkably improved. However, the brain electrical activity of HH-related epilepsy is of significant specificity, and its inter-seizure spike is probably originated from subcortical regions near the hamartoma, accompanied by late cortical diffusion, and is essentially different from the cortical origin pattern of conventional epilepsy. Meanwhile, due to the rarity of HH diseases, high-quality EEG data which can be used for distinguishing the GS subtype from the NGS subtype are scarce, the requirement of a deep learning model on a large sample is difficult to meet, the model based on conventional epileptic signal training is difficult to accurately capture an HH specific electrophysiological mechanism, and the effective distinction of the GS subtype and the NGS subtype cannot be realized. The Graph Neural Network (GNN) is used as an emerging deep learning framework, node neighborhood information can be aggregated through a message transmission mechanism, and the efficient learning graph nodes are embedded to represent and capture global topological attributes, so that the method has unique advantages in the aspect of mining epileptic nerve biomarkers. However, there is currently no study on the application of GNN to the identification and biomarker mining of GS and NGS subtypes in HH patients. Aiming at the problems of strong subjective dependence, limited feature extraction, insufficient model suitability and the like in HH epileptic subtype identification in the prior art, a deep learning method specially adapting to the characteristics of HH electroencephalogram signals is needed