An alzheimer's prediction system
By using a deep learning model that combines multi-layer perceptron and graph neural network, multi-modal data is integrated for Alzheimer's disease risk assessment, which solves the problem of low accuracy of single-modal assessment and achieves higher-dimensional risk assessment and early identification.
Patent Information
- Application Number
- CN202510829543.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2026-08-25
- Estimated Expiration
- 2045-06-20
AI Technical Summary
Existing Alzheimer's disease risk assessment methods are based on single-modal data, which makes it difficult to fully reflect an individual's true risk level and fails to fully consider the intrinsic relationships between different modal data, resulting in limited prediction accuracy and insufficient model generalization ability.
A deep learning prediction model combining multilayer perceptron and graph neural network is adopted, integrating magnetic resonance images, biochemical biomarkers and neuropsychological data. By constructing a relationship graph between phenotypic and genotypic features through multi-head attention layer and graph neural network, the fusion of multimodal information and coupled feature extraction are realized.
It improves the predictive accuracy of Alzheimer's disease risk assessment, enhances the effectiveness of early intervention, can more accurately characterize changes in patients' brain and cognitive states, and improves the sensitivity of early identification and the model's individualized predictive ability.
Smart Images
Figure CN120748772B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to an Alzheimer's disease prediction system. Background Technology
[0002] Alzheimer's disease (AD) is a common neurodegenerative disease, mainly characterized by cognitive decline, memory impairment, and behavioral abnormalities, severely impacting patients' quality of life. Due to its complex pathogenesis and long course, effective early intervention methods are currently lacking in clinical practice. Studies have shown that timely identification of high-risk groups and implementation of intervention measures in the early stages of the disease can help slow disease progression and reduce the burden on healthcare.
[0003] Existing Alzheimer's disease risk assessment methods are typically based on single-modality data analysis, such as using magnetic resonance imaging (MRI) to assess brain atrophy or genetic risk factors like APOE through gene testing. However, the risk of developing Alzheimer's disease is influenced by multiple factors, and single-modality information often fails to comprehensively reflect an individual's true risk level, resulting in limited predictive accuracy. Furthermore, existing methods fail to adequately consider the intrinsic relationships between different modalities of data and lack modeling of data coupling relationships, leading to insufficient model generalization ability and making them difficult to apply to personalized and accurate prediction scenarios. Summary of the Invention
[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides an Alzheimer's disease prediction system.
[0005] To achieve the above objectives, the main technical solutions adopted by the present invention include:
[0006] This invention provides an Alzheimer's disease prediction system, comprising:
[0007] The data receiving module is used to receive the first information data input from the user; the first information data includes phenotypic information data and genotypic information data; the phenotypic information data includes: magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB); the genotypic information data includes: pre-acquired SNP data on chromosome 19 where the user's genetic risk factor APOE is located.
[0008] The data prediction module is used to input the user's first information data into a pre-trained prediction model, and the pre-trained prediction model outputs a prediction label corresponding to the first information data.
[0009] The prediction model includes: a first multilayer perceptron for processing magnetic resonance imaging (MRI) data to obtain feature information of MRI; a second multilayer perceptron for processing biochemical biomarker (BMR) data to obtain feature information of BMR; a third multilayer perceptron for processing neuropsychological data (NPB) data to obtain feature information of NPB; an integration unit for integrating the feature information of MRI, BMR, and NPB data to obtain a phenotypic fusion feature vector; a first extraction component for obtaining a genotype feature vector based on genotype information data; a first fusion unit for obtaining a relationship graph corresponding to the phenotypic fusion feature vector and the genotype feature vector based on the phenotypic fusion feature vector and the genotype feature vector; a graph neural network for processing the relationship graph and outputting the coupling features corresponding to the relationship graph; and a prediction head for processing the coupling features and outputting the prediction label corresponding to the first information data.
[0010] Preferably, the trained prediction model is obtained by training the prediction model in advance using a training dataset;
[0011] The training dataset includes multiple sample data;
[0012] Each sample data includes the user's primary information data collected at a historical moment and its corresponding real label.
[0013] Preferably, the integration unit integrates the feature information from magnetic resonance imaging (MRI), biochemical biomarker (BMR), and neuropsychological data (NPB) to obtain a phenotypic fusion feature vector, specifically including:
[0014] The integration unit uses formula (1) to integrate the feature information of magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB) to obtain a phenotypic fusion feature vector.
[0015] The formula (1) is:
[0016] Z Phe =ψ Phe ([Z M ||Z B ||Z N ]);
[0017] Where, ψ Phe Z is the softmax gate function; M For the characteristic information of magnetic resonance imaging data; Z BCharacteristic information for biochemical biomarker data (BMR); Z N For the feature information of NPB in neuropsychological data; [Z M ||Z B ||Z N [] indicates the channel direction connection along the characteristic information of MRI image data, BMR biochemical biomarker data, and NPB neuropsychological data; Z Phe This is the phenotypic fusion feature vector.
[0018] Preferably,
[0019] The first extraction component includes:
[0020] A fourth multilayer perceptron is used to extract the embedded features from the genotype information data;
[0021] The fifth multilayer perceptron is used to obtain the identity encoding of the sites on the embedded features.
[0022] in,
[0023] n s d represents the number of loci in the genotype information data. e Indicates the embedding dimension of the embedded feature;
[0024] The multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (2);
[0025] The formula (2) is:
[0026]
[0027] MHSA is a multi-head attention layer; Z Gen This is a genotype feature vector; For embedded features;
[0028]
[0029] Preferably, the first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, obtains a relationship graph corresponding to the phenotypic fusion feature vector and the genotype feature vector, specifically including:
[0030] The first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, uses formula (3) to obtain the correlation between the phenotypic fusion feature vector and the genotype feature vector;
[0031] The formula (3) is:
[0032]
[0033] Γ[i,j] represents the genotype feature vector Z Gen The i-th dimension of the fused feature vector Z of embedded features and phenotypic features Phe The correlation between the j-th dimension embedded features Represents the set of real numbers;
[0034] It is the genotype feature vector Z Gen The embedding feature value of the i-th dimension at the k-th site;
[0035] It is the phenotypic fusion feature vector Z Phe The embedding feature value of the j-th dimension at the k-th site; It is the average value of the embedded feature values in the i-th dimension of the genotype feature vector;
[0036] It is the average value of the embedded feature value of the j-th dimension of the phenotypic fusion feature vector;
[0037] Based on the correlation between the phenotypic fusion feature vector and the genotype feature vector, the adjacency matrix of the relation graph is constructed using formula (4);
[0038] The formula (4) is:
[0039]
[0040] Where τ is a pre-set threshold;
[0041] A[i,j] is the adjacency matrix corresponding to the correlation between the phenotypic fusion feature vector and the genotype feature vector;
[0042] Based on the adjacency matrix A[i, j], obtain the corresponding relation graph G;
[0043] Where G = {V, E};
[0044] V represents the set of nodes in graph G; E represents the set of edges in graph G;
[0045] V j Let be the j-th node in the relation graph G.
[0046] Preferably, specifically, the graph neural network processes the relationship graph using formula (5) to obtain the coupling features corresponding to the relationship graph;
[0047] The formula (5) is:
[0048] Z = GCN(G, v);
[0049] GCN stands for Graph Neural Network;
[0050] v represents the feature set of all nodes in the node set V of the relation graph G;
[0051]
[0052] v j V represents the j-th node in the relation graph G. j eigenvalues, and v j Fusion feature vector Z with phenotypic features Phe The j-th value is the same;
[0053] Z represents the coupling characteristic.
[0054] Preferably, the prediction head performs prediction based on the coupling features and outputs the predicted label corresponding to the first information data, specifically including:
[0055] The prediction head uses formula (6) to make a prediction based on the coupling features to obtain the prediction label corresponding to the first information data;
[0056] The formula (6) is:
[0057]
[0058] in, For predicting labels;
[0059] MLP pred-head For prediction head.
[0060] Preferably, the loss function of the trained prediction model is the cross-entropy loss function, and training is completed when the loss function converges.
[0061] Preferably,
[0062] The real label is a label that identifies whether the user corresponding to the sample data actually has Alzheimer's disease;
[0063] The predicted label is the prediction result of the prediction model on the user's risk of developing Alzheimer's disease, expressed in the form of a probability value.
[0064] Preferably,
[0065] Magnetic resonance imaging (MRI) data includes: average cortical thickness values for multiple cortical regions, including but not limited to multiple cortical regions such as the frontal lobe, parietal lobe, temporal lobe, occipital lobe, and insula; volume parameters of multiple subcortical structures, including but not limited to multiple subcortical regions of the hippocampus, amygdala, thalamus, and putamen; and cerebrospinal fluid volume, white matter volume, and gray matter volume parameters.
[0066] Biochemical biomarker data (BMR) include the concentration levels of β-amyloid protein, total Tau protein, and phosphorylated Tau protein in cerebrospinal fluid.
[0067] The neuropsychological data NPB includes: memory composite score, executive function composite score, verbal cognition composite score, and visuospatial skills composite score.
[0068] The beneficial effects of this invention are:
[0069] This invention provides an Alzheimer's disease prediction system. By employing a deep learning prediction model that combines multimodal information fusion and graph neural networks, it achieves a higher-dimensional Alzheimer's risk assessment compared to existing technologies, thus improving prediction accuracy and enhancing the effectiveness of early intervention. Specifically, the beneficial effects of its technical solution include:
[0070] By combining phenotypic and genotypic information as input, with phenotypic information including MRI images, biochemical biomarkers, and neuropsychological scale data, and genotypic information including SNP data related to the APOE gene, this approach can comprehensively reflect the user's physiological, cognitive, and genetic characteristics compared to prediction methods based on only a single data source, thus improving the comprehensiveness of predictions.
[0071] By setting up multiple multilayer perceptrons in the prediction model to extract features from MRI, biochemical markers, and psychological data, and by using a softmax gating mechanism to achieve effective feature integration, compared with traditional simple splicing or averaging methods, it can highlight the key information of each type of data, enhance the model's ability to identify phenotypic features of different dimensions, and achieve the effect of optimizing phenotypic feature fusion expression.
[0072] By introducing a genotype information embedding feature extraction mechanism and forming high-quality genotype feature vectors through a multi-head attention layer, it can more effectively mine the semantic relationships between gene SNPs compared to traditional static coding methods, thus improving the ability to model genetic information.
[0073] By constructing a relationship graph based on the correlation between phenotypic fusion features and genotypic features, and using graph neural networks for deep semantic fusion, it can uncover complex interactive relationships between features of different modalities, achieving the effect of generating highly discriminative coupled features, compared to the shortcomings of traditional neural networks that ignore the relationships between features.
[0074] By employing a graph convolutional network (GCN)-based structure to propagate and aggregate features from the relational graph, it can enhance global feature representation while preserving local structure, compared to traditional models that do not use graph structures. This results in an improved understanding of the underlying mechanisms of Alzheimer's disease.
[0075] Because MRI data contains detailed parameters such as cortical thickness, brain region volume, cerebrospinal fluid, white matter, and gray matter, and biochemical markers cover key proteins of multiple neurodegenerative diseases, while neuropsychological data encompasses multiple dimensions of cognition, it can more accurately depict changes in the patient's brain and cognitive state compared to existing single-dimensional feature prediction methods, thus achieving the technical effect of improving the sensitivity of early identification of Alzheimer's disease. Attached Figure Description
[0076] Figure 1 This is a schematic diagram of the structure of an Alzheimer's disease prediction system according to the present invention;
[0077] Figure 2 This is a schematic diagram of the prediction model architecture in Embodiment 1 of the present invention. Detailed Implementation
[0078] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0079] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.
[0080] Example 1
[0081] See Figure 1 This embodiment provides an Alzheimer's disease prediction system, including:
[0082] The data receiving module is used to receive the first information data input from the user; the first information data includes phenotypic information data and genotypic information data; the phenotypic information data includes: magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB); the genotypic information data includes: pre-acquired SNP data on chromosome 19 where the user's genetic risk factor APOE is located.
[0083] It should be noted that the SNP data on chromosome 19, where the user's genetic risk factor APOE is located, obtained in advance in this embodiment is in one-hot encoded format.
[0084] In this embodiment, the magnetic resonance imaging (MRI) data includes: average cortical thickness values for multiple cortical regions, including but not limited to multiple cortical regions such as the frontal lobe, parietal lobe, temporal lobe, occipital lobe, and insula; volume parameters of multiple subcortical structures, including but not limited to multiple subcortical regions of the hippocampus, amygdala, thalamus, and putamen; cerebrospinal fluid volume, white matter volume, and gray matter volume parameters; biochemical biomarker data (BMR) including: concentration levels of β-amyloid protein, total Tau protein, and phosphorylated Tau protein in cerebrospinal fluid; and neuropsychological data (NPB) including: memory composite score, executive function composite score, verbal cognition composite score, and visuospatial skill composite score.
[0085] Specifically, the magnetic resonance imaging data (MRI) includes mean cortical thickness (TA) of 66 cortical regions and subcortical volumes (CV) of 28 regions, as well as cerebrospinal fluid volume, white matter volume, and gray matter volume.
[0086] The data prediction module is used to input the user's first information data into a pre-trained prediction model, and the pre-trained prediction model outputs a prediction label corresponding to the first information data.
[0087] See Figure 2 The prediction model includes: a first multilayer perceptron, used to process magnetic resonance imaging (MRI) data to obtain feature information of the MRI data; specifically, Z... M =MLP1(X M ); X M For MRI image data, MLP1 is the first multilayer sensor, Z M The feature information of MRI image data;
[0088] The second multilayer perceptron is used to process biochemical biomarker data (BMR) and obtain its feature information; specifically, Z... B =MLP2(X B );X BFor biochemical biomarker data BMR, MLP2 is the second multilayer perceptron, Z B Feature information for biochemical biomarker data (BMR);
[0089] The third multilayer perceptron is used to process neuropsychological data NPB and obtain its feature information; specifically, Z... N =MLP3(X N );X N The data is from neuropsychological data NPB, MLP3 is the third multilayer perceptron, and Z... N This refers to the feature information of NPB in neuropsychological data;
[0090] The integration unit is used to integrate the feature information from MRI image data, BMR biomarker data, and NPB neuropsychological data to obtain a phenotypic fusion feature vector; specifically, it includes:
[0091] The integration unit uses formula (1) to integrate the feature information of MRI image data, BMR biochemical biomarker data, and NPB neuropsychological data to obtain a phenotypic fusion feature vector; the formula (1) is: Z Phe =ψ Phe ([Z M ||Z B ||Z N ]);
[0092] Where, ψ Phe Z is the softmax gate function; M For the characteristic information of magnetic resonance imaging data; Z B Characteristic information for biochemical biomarker data (BMR); Z N For the feature information of NPB in neuropsychological data; [Z M ||Z B ||Z N [] indicates the channel direction connection along the characteristic information of MRI image data, BMR biochemical biomarker data, and NPB neuropsychological data; Z Phe This is the phenotypic fusion feature vector.
[0093] The first extraction component is used to obtain genotype feature vectors based on genotype information data;
[0094] The first extraction component includes:
[0095] A fourth multilayer perceptron is used to extract embedded features from the genotype information data; specifically, X SFor genotype information data, MLP4 is the fourth multilayer perceptron. For embedded features;
[0096] The fifth multilayer perceptron is used to obtain the identity encoding of the sites on the embedded features.
[0097] Specifically, MLP5 is the fifth multilayer perceptron;
[0098] in,
[0099] n s Genotype information data X S Number of upper positions, d e Representing embedded features Embedding dimension;
[0100] The multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (2);
[0101] The formula (2) is:
[0102]
[0103] MHSA is a multi-head attention layer; Z Gen This is a genotype feature vector; For embedded features;
[0104]
[0105] In this embodiment, the multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (2). That is, in this embodiment, the identity encoding of the sites is used to obtain the genotype feature vector. As an embedded feature Location encoding and embedding features The features are added together, and then the genotype feature vector Z is extracted through a set of multi-head attention layers (MHSA). Gen .
[0106] The first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, obtains the correlation between the phenotypic fusion feature vector and the genotype feature vector, specifically including:
[0107] The first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, uses formula (3) to obtain the correlation between the phenotypic fusion feature vector and the genotype feature vector;
[0108] The formula (3) is:
[0109]
[0110] Γ[i,j] represents the genotype feature vector Z Gen The i-th dimension of the fused feature vector Z of embedded features and phenotypic features Phe The correlation between the j-th dimension embedded features Represents the set of real numbers;
[0111] It is the genotype feature vector Z Gen The embedding feature value of the i-th dimension at the k-th site;
[0112] It is the phenotypic fusion feature vector Z Phe The embedding feature value of the j-th dimension at the k-th site; It is the average value of the embedded feature values in the i-th dimension of the genotype feature vector;
[0113] It is the average value of the embedded feature value of the j-th dimension of the phenotypic fusion feature vector;
[0114] Based on the correlation between the phenotypic fusion feature vector and the genotype feature vector, the adjacency matrix of the relation graph is constructed using formula (4);
[0115] The formula (4) is:
[0116]
[0117] Wherein, τ is a preset threshold; in this embodiment, τ is 0.2.
[0118] A[i,j] is the adjacency matrix corresponding to the correlation between the phenotypic fusion feature vector and the genotype feature vector;
[0119] Based on the adjacency matrix A[i, j], obtain the corresponding relation graph G;
[0120] Where G = {V, E};
[0121] V represents the set of nodes in the relation graph G; E represents the set of edges in the relation graph G; in this embodiment, the nodes in the node set represent the nodes in the relation graph G (that is, the elements on the diagonal of the adjacency matrix of the relation graph); the edges in the relation graph G represent the non-zero elements on the off-diagonal of the adjacency matrix.
[0122] V j Let be the j-th node in the relation graph G.
[0123] Graph neural networks are used to process relational graphs and output coupling features corresponding to the relational graph.
[0124] Specifically, the graph neural network uses formula (5) to process the relationship graph to obtain the coupling features corresponding to the relationship graph;
[0125] The formula (5) is:
[0126] Z = GCN(G, v);
[0127] GCN stands for Graph Neural Network;
[0128] v represents the feature set of all nodes in the node set V of the relation graph G;
[0129]
[0130] v j V represents the j-th node in the relation graph G. j eigenvalues, and v j Fusion feature vector Z with phenotypic features Phe The j-th value is the same;
[0131] Z represents the coupling characteristic.
[0132] The prediction head processes the coupled features and outputs the predicted label corresponding to the first information data. Specifically, it includes:
[0133] The prediction head uses formula (6) to make a prediction based on the coupling features to obtain the prediction label corresponding to the first information data;
[0134] The formula (6) is:
[0135]
[0136] in, For predicting labels;
[0137] MLP pred-head This is the prediction head. In this embodiment, the prediction head uses a multilayer perceptron to complete the final prediction task.
[0138] In practical applications, a well-trained prediction model is obtained by training the prediction model with a training dataset in advance.
[0139] The training dataset includes multiple sample data;
[0140] In this embodiment, the number of sample data is greater than or equal to 10,000.
[0141] Each sample data includes the user's primary information data collected at a historical moment and its corresponding real label.
[0142] The loss function of the trained prediction model is the cross-entropy loss function, and training is completed when the loss function converges.
[0143] The real label is a label that identifies whether the user corresponding to the sample data actually has Alzheimer's disease;
[0144] The predicted label is the prediction result of the prediction model on the user's risk of developing Alzheimer's disease, expressed in the form of a probability value.
[0145] Example 2
[0146] This embodiment provides an Alzheimer's disease prediction system, including:
[0147] The data receiving module is used to receive the first information data input from the user; the first information data includes phenotypic information data and genotypic information data; the phenotypic information data includes: magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB); the genotypic information data includes: pre-acquired SNP data on chromosome 19 where the user's genetic risk factor APOE is located.
[0148] It's important to note that SNPs are static and do not change over time. They are typically represented by discrete 0, 1, and 2 codes, reflecting combinations of different alleles. Although the number of SNP features is vast, they are usually sparse, and their relationship with disease is more about revealing genetic risk than directly reflecting specific disease manifestations. Genotypic features, on the other hand, are mostly discrete and sparse, and involve the interaction effects of multiple SNPs.
[0149] In this embodiment, the magnetic resonance imaging (MRI) data includes the mean cortical thickness (TA) of 66 cortical regions and the subcortical volume (CV) of 28 regions, as well as the cerebrospinal fluid volume, white matter volume, and gray matter volume. The biochemical biomarker data (BMR) includes the concentration levels of β-amyloid protein, total Tau protein, and phosphorylated Tau protein in the cerebrospinal fluid; the neuropsychological data (NPB) includes the memory composite score, executive function composite score, verbal cognition composite score, and visuospatial skill composite score.
[0150] In this embodiment, magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), neuropsychological data (NPB), and genotype information (SNP loci) are used in combination. Features are extracted and integrated through a multilayer perceptron to fully explore the information features of individuals at multiple levels, including structural, physiological, and genetic aspects, overcoming the problems of low accuracy and poor generalization of single-modality prediction.
[0151] The data prediction module is used to input the user's first information data into a pre-trained prediction model, and the pre-trained prediction model outputs a prediction label corresponding to the first information data.
[0152] The prediction model includes:
[0153] The first multilayer perceptron is used to process magnetic resonance image data (MRI) and obtain the feature information of the MRI image data.
[0154] The second multilayer perceptron is used to process biochemical biomarker data (BMR) and obtain the feature information of the biochemical biomarker data (BMR).
[0155] The third multilayer perceptron is used to process neuropsychological data NPB and obtain the feature information of neuropsychological data NPB;
[0156] The integration unit is used to integrate the feature information of magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB) to obtain a phenotypic fusion feature vector.
[0157] The first extraction component is used to obtain genotype feature vectors based on genotype information data;
[0158] The first extraction component in this second embodiment includes:
[0159] A fourth multilayer perceptron is used to extract embedded features from the genotype information data; specifically, X S For genotype information data, MLP4 is the fourth multilayer perceptron. For embedded features;
[0160] The fifth multilayer perceptron is used to obtain the identity encoding of the sites on the embedded features.
[0161] Specifically, MLP5 is the fifth multilayer perceptron;
[0162] in,
[0163] n s Genotype information data X S Number of upper positions, d e Representing embedded features Embedding dimension;
[0164] The counterfactual strategy processing unit encodes the identity based on the location on the embedded features. Gaussian noise g with the same mean and standard deviation is used, and then a counterfactual identity identifier is inferred based on the Gaussian noise g using a sixth multilayer perceptron (MLP6).
[0165] Specifically,
[0166] in, Counterfactual identifier;
[0167] Then, the first identifier ID is obtained based on formula (a). S ;
[0168] The formula (a) is:
[0169] The multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (b);
[0170] Formula (b) is:
[0171]
[0172] MHSA is a multi-head attention layer; Z Gen This is a genotype feature vector; For embedded features;
[0173]
[0174] In this embodiment, the multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (b). That is, in this embodiment, the first identifier ID of the site is used. S As an embedded feature Location encoding and embedding features Genotype feature vectors are extracted using a set of multi-head attention layers (MHSA).
[0175] This embodiment introduces a counterfactual identity generation strategy, adding Gaussian noise to gene embedding features and inferring counterfactual identity codes, thereby constructing a change vector between counterfactual and factual information (i.e., the first identifier ID). S This approach effectively simulates the impact of mutations at key gene loci on phenotypic risk prediction. This strategy enhances the model's sensitivity to and generalization ability to potential gene perturbations, thereby improving the interpretability of genetic risk modeling.
[0176] The first fusion unit is used to obtain a relationship graph corresponding to the phenotypic fusion feature vector and the genotypic feature vector based on the phenotypic fusion feature vector and the genotypic feature vector.
[0177] Graph neural networks are used to process relational graphs and output coupling features corresponding to the relational graph.
[0178] This embodiment constructs a relationship graph between phenotypic fusion feature vectors and genotypic feature vectors, and uses a graph neural network to process this graph to uncover the coupling structure and semantic associations between different features. This allows for a more comprehensive modeling of the potential influence paths between phenotypic performance and genetic risk, thereby improving the global learning ability for prediction.
[0179] The prediction head is used to process the coupled features and output the predicted label corresponding to the first information data.
[0180] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0181] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0182] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.
[0183] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0184] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. An Alzheimer's disease prediction system, characterized in that, include: The data receiving module is used to receive the initial information data input by the user; The first information data includes phenotypic information data and genotypic information data; The phenotypic information data includes: magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB); the genotypic information data includes: pre-acquired SNP data on chromosome 19, where the user's genetic risk factor APOE is located. The data prediction module is used to input the user's first information data into a pre-trained prediction model, and the pre-trained prediction model outputs a prediction label corresponding to the first information data. The prediction model includes: a first multilayer perceptron for processing magnetic resonance imaging (MRI) data to obtain feature information of MRI; a second multilayer perceptron for processing biochemical biomarker (BMR) data to obtain feature information of BMR; a third multilayer perceptron for processing neuropsychological data (NPB) data to obtain feature information of NPB; an integration unit for integrating the feature information of MRI, BMR, and NPB data to obtain a phenotypic fusion feature vector; a first extraction component for obtaining a genotype feature vector based on genotype information data; a first fusion unit for obtaining a relationship graph corresponding to the phenotypic fusion feature vector and the genotype feature vector based on the phenotypic fusion feature vector and the genotype feature vector; a graph neural network for processing the relationship graph and outputting the coupling features corresponding to the relationship graph; and a prediction head for processing the coupling features and outputting the prediction label corresponding to the first information data. The first extraction component includes: A fourth multilayer perceptron is used to extract embedded features from the genotype information data; specifically, ; For genotype information data, MLP4 is the fourth multilayer perceptron. For embedded features; The fifth multilayer perceptron is used to obtain the identity encoding of the sites on the embedded features. ; Specifically, MLP5 is the fifth multilayer perceptron. in, ; Genotype information data Number of upper positions, Representing embedded features Embedding dimension; The counterfactual strategy processing unit encodes the identity based on the location on the embedded features. Gaussian noise g with the same mean and standard deviation is used, and then a counterfactual identity identifier is inferred based on the Gaussian noise g using a sixth multilayer perceptron (MLP6). Specifically, ; in, Counterfactual identifier; Then, the first identifier is obtained based on formula (a). ; The formula (a) is: ; The multi-head attention unit obtains the genotype feature vector based on the embedded features and the identity encoding of the sites on the embedded features using formula (b); Formula (b) is: ； MHSA is a multi-head attention layer; This is a genotype feature vector; For embedded features; .
2. The Alzheimer's disease prediction system according to claim 1, characterized in that, in, A well-trained prediction model is obtained by training the prediction model in advance using a training dataset; The training dataset includes multiple sample data; Each sample data includes the user's primary information data collected at a historical moment and its corresponding real label.
3. The Alzheimer's disease prediction system according to claim 2, characterized in that, The integration unit integrates the feature information from MRI image data, BMR biomarker data, and NPB neuropsychological data to obtain a phenotypic fusion feature vector, specifically including: The integration unit uses formula (1) to integrate the feature information of magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB) to obtain the phenotypic fusion feature vector. The formula (1) is: ； in, It is the softmax gate function; The feature information of MRI image data; Feature information for biochemical biomarker data (BMR); This refers to the feature information of NPB in neuropsychological data; This indicates the channel direction connection along the feature information of magnetic resonance imaging data (MRI), biochemical biomarker data (BMR), and neuropsychological data (NPB). This is the phenotypic fusion feature vector.
4. The Alzheimer's disease prediction system according to claim 3, characterized in that, The first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, obtains the correlation between the phenotypic fusion feature vector and the genotype feature vector, specifically including: The first fusion unit, based on the phenotypic fusion feature vector and the genotype feature vector, uses formula (3) to obtain the correlation between the phenotypic fusion feature vector and the genotype feature vector; The formula (3) is: ； Representative genotype feature vector The i-th dimension of the embedded feature and the phenotypic fusion feature vector The correlation between the j-th dimension embedded features ; Represents the set of real numbers; It is a genotype feature vector The Middle k The embedding feature value of the i-th dimension of each site; It is a phenotypic fusion feature vector The Middle k The embedding feature value of the j-th dimension of each site; It is the average value of the embedded feature values in the i-th dimension of the genotype feature vector; It is the average value of the embedded feature value of the j-th dimension of the phenotypic fusion feature vector; Based on the correlation between the phenotypic fusion feature vector and the genotype feature vector, the adjacency matrix of the relation graph is constructed using formula (4); The formula (4) is: ； in, For a pre-set threshold; This is the adjacency matrix corresponding to the correlation between the phenotypic fusion feature vector and the genotype feature vector; Based on the adjacency matrix Obtain the corresponding relationship graph G; in, ; V represents the set of nodes in graph G; E represents the set of edges in graph G; ； Let be the j-th node in the relation graph G.
5. The Alzheimer's disease prediction system according to claim 4, characterized in that, Specifically, the graph neural network uses formula (5) to process the relationship graph to obtain the coupling features corresponding to the relationship graph; The formula (5) is: ； GCN stands for Graph Neural Network; Let V be the set of features of all nodes in the node set V of the relation graph G; ； This represents the j-th node in the relation graph G. eigenvalues, and Feature vector fused with phenotypic features The j-th value is the same; Z represents the coupling characteristic.
6. The Alzheimer's disease prediction system according to claim 5, characterized in that, The prediction head performs predictions based on the coupling features and outputs the predicted label corresponding to the first information data, specifically including: The prediction head uses formula (6) to make a prediction based on the coupling features, and obtains the prediction label corresponding to the first information data; The formula (6) is: ； in, For predicting labels; For prediction head.
7. The Alzheimer's disease prediction system according to claim 6, characterized in that, The loss function of the trained prediction model is the cross-entropy loss function, and training is completed when the loss function converges.
8. The Alzheimer's disease prediction system according to claim 7, characterized in that, The real label is a label that identifies whether the user corresponding to the sample data actually has Alzheimer's disease; The predicted label is the prediction result of the prediction model on the user's risk of developing Alzheimer's disease, expressed in the form of a probability value.
9. The Alzheimer's disease prediction system according to claim 8, characterized in that, Magnetic resonance imaging (MRI) data includes: average cortical thickness values for multiple cortical regions, including but not limited to multiple cortical regions such as the frontal lobe, parietal lobe, temporal lobe, occipital lobe, and insula; volume parameters of multiple subcortical structures, including but not limited to multiple subcortical regions of the hippocampus, amygdala, thalamus, and putamen; and cerebrospinal fluid volume, white matter volume, and gray matter volume parameters. Biochemical biomarker data (BMR) include the concentration levels of β-amyloid protein, total Tau protein, and phosphorylated Tau protein in cerebrospinal fluid. The neuropsychological data NPB includes: memory composite score, executive function composite score, verbal cognition composite score, and visuospatial skills composite score.