A method for predicting Alzheimer's disease based on gene sequences and brain gray matter characteristics
Through multimodal prediction methods of gene sequence and brain gray matter characteristics, the number of dynamic neurons and cross-modal attention mechanisms are used to solve the problem of identification of MCI in the precursor stage of Alzheimer's disease, and the stable and progressive MCI are achieved. The accurate distinction of the prediction accuracy and model performance are improved.
Patent Information
- Application Number
- CN202410509044.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-26
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-04-26
AI Technical Summary
The prior art is difficult to accurately identify whether mild cognitive impairment (MCI) in the prodromal stage of Alzheimer's disease will stabilize or progress to Alzheimer's disease, and there is a lack of effective prediction methods, which affects the effectiveness of clinical treatment.
A method for predicting gene sequences and brain gray matter characteristics, including multimodal feature extraction network, feature fusion module and prediction output layer, multimodal feature extraction network that dynamically adjusts the number of neurons and network parameters, combined with cross-modal Attention mechanism and 2D CNN model capture the interaction of morphological features, sMRI and SNPs, achieve feature fusion, and use the SoftMax activation function for prediction.
It improves the accuracy and reliability of Alzheimer's disease prediction, can effectively distinguish between stable MCI and progressive MCI, enhances the adaptability and prediction ability of the model, and provides more accurate emotion recognition results.
Smart Images

Figure CN118352082B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer science and technology, and in particular to a method for predicting Alzheimer's disease based on gene sequences and brain gray matter characteristics. Background Art
[0002] Alzheimer's disease (AD) is a neurodegenerative disease that usually leads to a gradual decline in cognitive abilities in the elderly, affecting their daily life and social functions. Due to the lack of effective treatments, it is extremely difficult to prevent the deterioration of AD. To solve this problem, researchers have focused on mild cognitive impairment (MCI), the prodromal stage of AD, and believe that intervention during MCI is crucial to delay or even prevent the patient's condition from progressing to the AD stage. In the classification of MCI, it can be divided into two types according to its progression characteristics: stable MCI (sMCI) and progressive MCI (pMCI). sMCI is defined as a diagnosis that is the same as the baseline MCI during a 36-month follow-up period, which means that during this period, the patient's cognitive function level has not deteriorated significantly, remaining in the MCI state without further progression to AD. In contrast, pMCI is defined as a transition from the MCI state at baseline to AD during a 36-month follow-up period. Therefore, accurate identification of MCI patients as sMCI and pMCI is of great significance for clinical treatment.
[0003] To this end, we propose a method for predicting Alzheimer's disease based on gene sequences and brain gray matter characteristics. Summary of the invention
[0004] The purpose of the present invention is to provide a method for predicting Alzheimer's disease based on gene sequences and brain gray matter characteristics to solve the problems raised in the above background technology.
[0005] To achieve the above object, the present invention provides the following technical solutions: a method for predicting Alzheimer's disease based on gene sequences and brain gray matter features, comprising a prediction framework, wherein the prediction framework is composed of a multimodal feature extraction network, a feature fusion module and a prediction output layer;
[0006] The multimodal feature extraction network includes three hidden layers with variable parameters, which can dynamically adjust the number of neurons and network parameters according to the characteristics of the input data;
[0007] The feature fusion module captures the interaction between morphological features, sMRI and SNPs through a cross-modal attention mechanism to achieve feature fusion;
[0008] The prediction output layer is composed of a fully connected layer with two output nodes, and a SoftMax activation function is used to ensure the probability distribution of the output value.
[0009] Preferably, the multimodal feature extraction network specifically includes: the number of neurons in the first hidden layer is the characteristic number of the data, which is used to receive the original data, and the number of neurons in the second and third hidden layers is half of the characteristic number of the data.
[0010] Preferably, the sMRI image processing adopts an artificial neural network model combined with 2D CNN, which uses three layers of Conv2D for feature extraction, each layer uses a 3×3 convolution kernel, and uses Relu as an activation function to gradually capture local features in the image.
[0011] Preferably, the multi-dimensional feature map is expanded into a one-dimensional feature vector through a Flatten layer to provide input data for the subsequent artificial neural network.
[0012] Preferably, the artificial neural network comprises two Dense layers, the first Dense layer has 100 neurons and is activated using a ReLU function, and the second Dense layer has 50 neurons and also uses a ReLU activation function.
[0013] Preferably, a Dropout layer with a dropout rate of 0.5 is included between the two Dense layers.
[0014] Preferably, the cross-modal attention mechanism in the feature fusion module is implemented by multi-head attention, and the multi-head attention output matrix is:
[0015] MultiHead(Q,K,V)=Concat(head1,…,head h )W Q
[0016] The parameter matrix Q, K, V represent the query, key, and value of the input modality, respectively, and h is the number of heads;
[0017] In the process of performing cross-modal attention mechanism, keys (Q) and values (K) are calculated for the same modality while querying (V) for the other modality.
[0018] Preferably, after the multimodal feature extraction, the feature fusion module generates three feature matrices, which are the morphological feature matrix M, the sMRI feature matrix I, and the SNPs feature matrix P.
[0019] Preferably, after generating the three feature matrices, the modal attention mechanism performs the following operations:
[0020] concatenate(cross attention(M,I))
[0021] concatenate(cross attention(M,P))
[0022] concatenate(cross attention(I,P))
[0023] The feature fusion module begins to integrate the key information of different modal data.
[0024] Preferably, the prediction output layer maps the 245-dimensional feature space after feature fusion into two output nodes, and the two nodes represent the prediction results of sMCI and pMCI respectively.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] 1. The present invention uses this method to simultaneously consider data from multiple modalities, including morphological features, sMRI and SNPs. This method can capture more comprehensive and rich information, thereby improving the accuracy and reliability of emotion recognition. The fusion of multimodal data helps to capture the complementarity between different data, making the recognition results more accurate.
[0027] 2. The present invention utilizes a multimodal feature extraction network to dynamically adjust the number of neurons and network parameters according to the characteristics of the input data, which makes the network highly adaptable and flexible. No matter how the characteristics of the input data change, the network can effectively extract key information and provide strong support for subsequent emotion recognition.
[0028] 3. The present invention uses a cross-modal attention mechanism. This method can capture the interaction between different modalities and achieve efficient feature fusion. This fusion method not only retains the original information of each modality, but also can establish connections between different modalities and mine deeper information, thereby improving the performance of emotion recognition.
[0029] 4. The prediction output layer uses a fully connected layer and a SoftMax activation function, which can map the fused features into the results of emotion recognition, making this method have strong prediction capabilities and can accurately distinguish different emotional states. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 It is a schematic diagram of the overall prediction framework of the present invention. DETAILED DESCRIPTION
[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] Embodiment 1
[0033] See also Figure 1 , the present invention provides a technical solution: a method for predicting Alzheimer's disease based on gene sequences and brain gray matter features, the present invention makes corresponding improvements to the technical problems mentioned in the background technology, including a prediction framework, the prediction framework is composed of a multimodal feature extraction network, a feature fusion module and a prediction output layer;
[0034] The multimodal feature extraction network contains three hidden layers with variable parameters, which can dynamically adjust the number of neurons and network parameters according to the characteristics of the input data;
[0035] The feature fusion module captures the interaction between morphological features, sMRI and SNPs through a cross-modal attention mechanism to achieve feature fusion;
[0036] The prediction output layer consists of a fully connected layer with two output nodes, and uses the SoftMax activation function to ensure the probability distribution of the output value;
[0037] The multimodal feature extraction network specifically includes: the number of neurons in the first hidden layer is the number of characteristics of the data, which is used to receive the original data; the number of neurons in the second and third hidden layers is half of the number of characteristics of the data;
[0038] For sMRI image processing, an artificial neural network model combined with 2D CNN is used. The model uses three layers of Conv2D for feature extraction. Each layer uses a 3×3 convolution kernel and uses Relu as the activation function to gradually capture local features in the image.
[0039] The multi-dimensional feature map is expanded into a one-dimensional feature vector through the Flatten layer to provide input data for the subsequent artificial neural network;
[0040] The artificial neural network contains two Dense layers. The first Dense layer has 100 neurons and uses the ReLU function for activation. The second Dense layer has 50 neurons and also uses the ReLU activation function.
[0041] There is a Dropout layer with a dropout rate of 0.5 between the two Dense layers;
[0042] The cross-modal attention mechanism in the feature fusion module is implemented by multi-head attention, and the multi-head attention output matrix is:
[0043] MultiHead(Q,K,V)=Concat(head1,...,head h ) O
[0044] The parameter matrix Q, K, V represent the query, key, and value of the input modality, respectively, and h is the number of heads;
[0045] In the process of performing cross-modal attention mechanism, keys (Q) and values (K) are calculated for the same modality, while queries (V) are performed on the other modality;
[0046] After multimodal feature extraction, the feature fusion module generates three feature matrices, which are the morphological feature matrix M, the sMRI feature matrix I, and the SNPs feature matrix P;
[0047] After generating the three feature matrices, the modal attention mechanism performs the following operations:
[0048] concatenate(cross attention(M,I))
[0049] concatenate(cross attention(M,P))
[0050] concatenate(cross attention(I,P))
[0051] The feature fusion module begins to synthesize key information from different modal data;
[0052] The prediction output layer maps the 245-dimensional feature space after feature fusion into two output nodes, which represent the prediction results of sMCI and pMCI respectively.
[0053] The specific implementation of this embodiment is as follows:
[0054] (1) The multimodal feature extraction network adopts different processing strategies for different data types to better adapt the feature extraction effect of each modality data. For sequence data, such as SNPs and morphological features, an adaptive artificial neural network is designed. The network contains three hidden layers with variable parameters. It can dynamically adjust the number of neurons and network parameters according to the characteristics of the input data, so as to effectively identify key features. The number of neurons in the first hidden layer is the number of data features, which is used to receive the original data. The number of neurons in the second and third hidden layers is half of the number of data features, in order to enhance the efficiency of feature processing. For image data, especially sMRI, a combination of 2D CNN artificial neural network model, first, the model uses three layers of Conv2D for feature extraction, each layer uses a 3×3 convolution kernel and adopts Relu as the activation function to gradually capture the local features in the image. Next, the multi-dimensional feature map is expanded into a one-dimensional feature vector through the Flatten layer to provide input data for the subsequent artificial neural network. The subsequent artificial neural network contains two Dense layers. The first Dense layer has 100 neurons and is activated by the ReLU function. The second Dense layer has 50 neurons and also uses the ReLU activation function. Between the two Dense layers, we designed a Dropout layer with a dropout rate of 0.5. These layers are designed to more accurately capture the key features in the image data and improve the generalization ability of the model.
[0055] (2) The feature fusion module aims to fuse data features of different modalities. We use the cross-modal attention mechanism to capture the interaction between morphological features, sMRI and SNPs, thereby achieving feature fusion. The cross-modal attention is implemented by multi-head attention, where the multi-head attention output matrix is:
[0056] MultiHead(Q,K,V)=Concat(head1,...,head h ) O
[0057] The parameter matrix Q, K, V represent the query, key, and value of the input modality respectively, h is the number of heads,
[0058] In the process of performing the cross-modal attention mechanism, for the same modality, we calculate the key (Q) and value (K), and query (V) for the other modality at the same time, so as to obtain the correlation between the modalities. After multimodal feature extraction, we generate three feature matrices, namely, the morphological feature matrix M, the sMRI feature matrix I, and the SNPs feature matrix P. After that, the cross-modal attention mechanism performs the following operations:
[0059] concatenate(cross attention(M,I))
[0060] concatenate(cross attention(M,P))
[0061] concatenate(cross attention(I,P))
[0062] Through the above steps, the feature fusion module can integrate the key information of different modal data, thereby enhancing the prediction performance of the model.
[0063] (3) The prediction output layer consists of a fully connected layer with two output nodes, and uses the SoftMax activation function to ensure the probability distribution of the output value. In our research task, this layer maps the 245-dimensional feature space after feature fusion into two output nodes, which represent the prediction results of sMCI and pMCI respectively.
[0064] Embodiment 2
[0065] The proposed method is compared with four other state-of-the-art methods. The following table lists the comparison results with the most advanced models for classification using multimodal data. The average accuracy, precision, recall and F1 score of this method under five randomly initialized test sets reached 95%, 95.55%, 95% and 94.98% respectively, surpassing the results of other advanced methods for sMCI and pMCI classification using multimodal data. This generation of method can effectively distinguish sMCI from pMCI.
[0066]
[0067]
[0068] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting Alzheimer's disease based on gene sequences and brain gray matter characteristics, characterized by: The method comprises a prediction framework, wherein the prediction framework is composed of a multimodal feature extraction network, a feature fusion module and a prediction output layer; The multimodal feature extraction network includes three hidden layers with variable parameters, which can dynamically adjust the number of neurons and network parameters according to the characteristics of the input data; The feature fusion module captures the interaction between morphological features, sMRI and SNPs through a cross-modal attention mechanism to achieve feature fusion; The prediction output layer is composed of a fully connected layer with two output nodes, and a SoftMax activation function is used to ensure the probability distribution of the output value; The cross-modal attention mechanism in the feature fusion module is implemented by multi-head attention, and the multi-head attention output matrix is: ; The parameter matrix , , , Represents the query, key, and value of the input modal, respectively. is the number of heads; In the process of performing the cross-modal attention mechanism, for the same modality, the key is calculated ( ) and value ( ), and query the other modality at the same time ( ); After the multimodal feature extraction, the feature fusion module generates three feature matrices, which are the morphological feature matrix M, the sMRI feature matrix I, and the SNPs feature matrix P; After generating the three feature matrices, the modal attention mechanism performs the following operations: concatenate (cross attention (M, I)) concatenate (cross attention (M, P)) concatenate (cross attention (I, P)) The feature fusion module begins to integrate the key information of different modal data.
2. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 1, characterized in that: The multimodal feature extraction network specifically includes: the number of neurons in the first hidden layer is the characteristic number of the data, which is used to receive the original data; the number of neurons in the second and third hidden layers is half of the characteristic number of the data.
3. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 1, characterized in that: For sMRI image processing, an artificial neural network model combined with 2D CNN is used. The model uses three layers of Conv2D for feature extraction. Each layer uses a 3×3 convolution kernel and adopts Relu as the activation function to gradually capture local features in the image.
4. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 3, characterized in that: The multi-dimensional feature map is expanded into a one-dimensional feature vector through the Flatten layer to provide input data for the subsequent artificial neural network.
5. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 4, characterized in that: The artificial neural network includes two Dense layers, the first Dense layer has 100 neurons and is activated using the ReLU function, and the second Dense layer has 50 neurons and also uses the ReLU activation function.
6. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 5, characterized in that: There is a Dropout layer with a dropout rate of 0.5 between the two Dense layers.
7. The method for predicting Alzheimer's disease based on gene sequence and brain gray matter characteristics according to claim 1, characterized in that: The prediction output layer maps the 245-dimensional feature space after feature fusion into two output nodes, and the two nodes represent the prediction results of sMCI and pMCI respectively.
