Disease risk prediction method

Abstract

The present application relates to the technical field of computer, and provides a disease risk prediction method, device, equipment, medium and product, the method comprises the following steps: constructing N characteristic matrices based on metagenomic data of a to-be-tested sample; constructing sample correlation graphs based on the N characteristic matrices respectively, and performing graph feature extraction on the N characteristic matrices based on the corresponding sample correlation graphs to obtain N characteristic representations; performing cross-type fusion on the N characteristic representations to obtain a fusion representation of the to-be-tested sample; performing disease risk prediction based on the fusion representation to obtain a prediction result; wherein N is a positive integer not less than 2. The method can perform prediction based on the fusion representation which retains statistical characteristics and topological structure characteristics of biological information at each level, effectively improves the utilization degree of the original metagenomic data, and improves the accuracy and robustness of disease risk prediction.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method for predicting disease risk. Background Technology

[0002] In recent years, research on the relationship between the host microbiota and host health has revealed a bidirectional regulatory relationship between microorganisms and host health. Metagenomic sequencing technology can comprehensively analyze the composition and function of the host microbiota. However, metagenomic data contains biological information at multiple levels, such as species classification, functional annotation, and enzyme activity. Existing methods mainly construct graph structures based on single-type features or simply splice multi-level features and then encode them uniformly. This fails to fully utilize the heterogeneous topological structures of different types of features and makes it difficult to achieve adaptive fusion of multi-level information within a sample, resulting in poor disease prediction performance based on microbiota metagenomic data.

[0003] Therefore, methods for disease prediction still require further research. Summary of the Invention

[0004] This invention aims to address, to a certain extent, the technical problems existing in the prior art. This invention proposes a method, apparatus, device, medium, and product for disease risk prediction. By extracting features from multi-level biological information of metagenomic data and adaptively interacting and fusing these multi-level features, a fused representation of the sample metagenomic data is obtained while effectively preserving the statistical and topological features of biological information at each level. This fused representation is then used for disease risk prediction, effectively improving the utilization of original metagenomic data and enhancing the accuracy and robustness of disease risk prediction.

[0005] In a first aspect, the present invention provides a method for predicting disease risk. According to an embodiment of the present invention, the method includes: constructing N feature matrices based on metagenomic data of a sample to be tested, wherein the N feature matrices correspond to microbial features at different biological levels; constructing sample association graphs based on the N feature matrices respectively, and extracting graph features from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations; performing cross-type fusion on the N feature representations to obtain a fused representation of the sample to be tested; and performing disease risk prediction based on the fused representation to obtain a prediction result; wherein N is a positive integer not less than 2.

[0006] According to embodiments of the present invention, the above-described disease risk prediction method may further include at least one of the following additional technical features: According to an embodiment of the present invention, the construction of N feature matrices based on metagenomic data of the sample to be tested includes: preprocessing the metagenomic data of the sample to be tested to obtain microbial features at N biological levels; constructing adjacency matrices and correlation matrices for the microbial features at the N biological levels respectively, wherein the adjacency matrix is ​​used to characterize the feature similarity relationship between the sample to be tested and the reference sample, and the correlation matrix is ​​used to characterize the feature common relationship between the sample to be tested and the reference sample.

[0007] According to an embodiment of the present invention, the preprocessing of the metagenomic data of the sample to be tested to obtain microbial features at N biological levels includes: performing transformation processing on the N biological level data in the metagenomic data of the sample to be tested to obtain N transformed features; and performing screening processing on the N transformed features to obtain the N biological level microbial features, wherein the biological level microbial features include K screened microbial features, and the K screened microbial features are the top K transformed features with the highest variance among the transformed features, where K is a positive integer not less than 1.

[0008] According to an embodiment of the present invention, constructing sample association graphs based on the N feature matrices includes: establishing connection edges between the test sample and its nearest neighbor samples based on the feature similarity relationship, adjusting the connection edges to obtain an adjacency graph, wherein the nearest neighbor samples include the reference samples that have similar features to the test sample; establishing association edges between the test sample and common samples based on the feature common relationship, adjusting the association edges to obtain an association graph, wherein the common samples include the reference samples that have common features to the test sample; and weighted mixing the adjacency graph and the association graph to obtain the sample association graph.

[0009] According to an embodiment of the present invention, the step of extracting graph features from the N feature matrices based on the corresponding sample association graph to obtain N feature representations includes: encoding the sample association graph to obtain a hidden representation; performing M-stage graph convolution aggregation processing on the hidden representation to obtain the feature representation, where M is a positive integer not less than 1, and the graph convolution aggregation processing is used to integrate neighborhood information of the node features of the sample association graph.

[0010] According to an embodiment of the present invention, cross-type fusion of the N feature representations to obtain a fused representation of the test sample includes: weighted fusion of the N feature representations based on a multi-head attention mechanism to obtain a sample representation; fusion of the N sample association graphs based on global type weights to obtain a global graph; feature propagation of the sample representation on the global graph to obtain a global representation; and weighted calibration of the global representation based on category prototypes to obtain a fused representation.

[0011] According to an embodiment of the present invention, performing cross-type fusion on the N feature representations to obtain a fused representation of the sample to be tested further includes: encoding the clinical modality data of the sample to be tested to obtain a clinical modality representation; and fusing the weighted calibration result with the clinical modality representation to obtain a fused representation.

[0012] According to an embodiment of the present invention, the step of performing disease risk prediction based on the fusion representation to obtain a prediction result further includes: performing a correction process on the prediction result to obtain a corrected risk prediction result, wherein the correction process includes at least one of the following: performing a correction process on the prediction result based on a sparse prior model to obtain a corrected risk prediction result; and / or performing a correction process on the prediction result based on a probability calibration model to obtain a corrected risk prediction result.

[0013] According to an embodiment of the present invention, the correction process includes: generating a first risk probability based on the prediction result; generating a second risk probability based on the sparse prior model and / or the probability calibration model; and generating a final risk prediction probability based on the first risk probability and the second risk probability.

[0014] Secondly, the present invention provides a disease risk prediction device. According to an embodiment of the present invention, the device includes: a feature matrix construction module, a feature representation module, a fusion module, and a prediction module. The system includes the following modules: a feature matrix construction module for constructing N feature matrices based on metagenomic data of the sample to be tested, wherein the N feature matrices correspond to microbial features at different biological levels; a feature representation module for constructing sample association graphs based on the N feature matrices and extracting graph features from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations; a fusion module for performing cross-type fusion of the N feature representations to obtain a fused representation of the sample to be tested; and a prediction module for performing disease risk prediction based on the fused representation to obtain a prediction result.

[0015] Thirdly, the present invention provides an electronic device, according to an embodiment of the present invention, the device comprising: a processor and a memory; the aforementioned memory for storing a computer program; the aforementioned processor for executing the computer program to implement the disease risk prediction method as provided in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium, according to an embodiment of the invention, wherein the medium stores computer instructions or programs that, when executed on a computer, cause the disease risk prediction method provided in the first aspect to be performed.

[0017] Fifthly, the present invention provides a computer program product, according to an embodiment of the invention, the product comprising computer instructions that, when some or all of the computer instructions are executed on a computer, cause the disease risk prediction method as provided in the first aspect to be performed.

[0018] In summary, the solution provided by the embodiments of the present invention, on the one hand, by screening multi-level biological information of metagenomics and independently constructing feature matrices and graphs for different levels of biological information, while eliminating invalid data, reducing data redundancy and data computation, effectively retains the statistical and topological features of biological information at different levels, providing more accurate and richer feature information for subsequent prediction.

[0019] On the other hand, by performing cross-type attention fusion on the aforementioned multi-level features and dynamically allocating attention weights for each level of feature type, information complementarity and adaptive fusion of features at different levels are achieved. This suppresses redundant noise, avoids redundant feature superposition in simple feature splicing methods, and overcomes the deficiency of its inability to model the intrinsic relationships between different feature types. Furthermore, based on the constructed category prototype, the fused features are calibrated for category and noise correction, avoiding the inaccuracy caused by feature dispersion and bias in small samples. Further, clinical modality representation is introduced to achieve complementarity with multi-dimensional information on internal physiological and physicochemical aspects or external behavioral manifestations, effectively improving the representation quality and robustness of the fused representation and enhancing the generalization performance of prediction. Moreover, a correction mechanism is used to improve the accuracy of prediction results, ensuring the practicality and deployment capability of the prediction method and device, and providing an objective and quantifiable method for disease prediction. Attached Figure Description

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart illustrating a disease risk prediction method. Figure 2 A schematic block diagram of a disease risk prediction device; Figure 3 A schematic block diagram of an electronic device; Figure 4 A schematic block diagram of an electronic device including a transceiver component; Figure 5 Receiver operating characteristic curves for the test set; Figure 6 Calibration curves for disease risk prediction methods. Detailed Implementation

[0021] The embodiments of the present invention are described in detail below. These embodiments are exemplary and are only used to explain the present invention, and should not be construed as limiting the invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Reagents or instruments used, unless otherwise specified, are all commercially available conventional products.

[0022] It should be noted that in the embodiments of the present invention, the terms "comprising" or "including" are open-ended expressions, that is, they include the contents specified in the present invention, but do not exclude other aspects.

[0023] In embodiments of the present invention, the terms “optionally,” “optionally,” or “optionally” generally refer to events or conditions described subsequently that may but may not occur, and the description includes both cases in which the event or condition occurs and cases in which the event or condition does not occur.

[0024] In this embodiment of the invention, the term "feature similarity relationship" refers to the pairwise pairing and close relationship formed between sample individuals based on the spatial distance and value matching degree of their own feature vectors.

[0025] In this embodiment of the invention, the term "feature common relationship" refers to the group-based association constraint relationship formed by multiple samples having the same type of feature attributes, shared feature patterns, or the same representational rules.

[0026] In this embodiment of the invention, the term "feature propagation" refers to the process of transmitting, circulating, and aggregating the feature information of each node / sample between individuals based on structural relationships, thereby achieving iterative updates and representation enhancement of features through information interaction.

[0027] In this embodiment of the invention, the term "boundary margin" refers to the difference in similarity between a sample and the positive / negative class prototype.

[0028] In this embodiment of the invention, the term "early stopping patience value" refers to the maximum number of iterations the model is allowed to continue iterating and waiting during model training when the performance of the validation set no longer improves.

[0029] In this embodiment of the invention, the term "Brier score" is an indicator that measures the accuracy of probability prediction. Essentially, it is the mean square error between the predicted probability and the true label, used to evaluate the model's probability calibration and discrimination ability. The smaller the value, the better.

[0030] The technical problem to be solved, the inventive concept, and the system architecture of the embodiments of the present invention will be described below: Currently, for disease prediction, especially for mental illnesses like postpartum depression which are difficult to diagnose, traditional methods mainly rely on various psychological rating scales for diagnosis. However, these methods have the following obvious limitations: (1) strong subjectivity, easily affected by the cognitive state of the subject or the cognitive environment around the subject; (2) obvious lag, making it difficult to provide early warning before the appearance of clinical symptoms; (3) time-consuming assessment, unsuitable for large-scale screening, etc., resulting in extremely unsatisfactory diagnostic results. Based on recent research on the relationship between the host's gut microbiota and host health, it is clear that microorganisms are not only closely related to the host's physiological health, but also deeply involved in the regulation of the host's mental state. In particular, the gut microbiota-brain axis research has confirmed the intrinsic law of mutual influence and bidirectional regulation between gut microbiota and the host's mental health. However, existing research methods mostly rely on single-class features to construct graph topology structures, or simply splice multi-layer features and then implement unified encoding. This cannot fully explore the heterogeneous topological characteristics inherent in various features, nor can it complete the deep integration of multi-dimensional hierarchical information within the sample. This easily leads to the loss of feature information and insufficient representation of heterogeneous associations, ultimately resulting in limited discrimination and representation capabilities.

[0031] To address the aforementioned issues, the inventors have solved the problems of existing prediction methods that rely solely on single-class features or simply concatenate multiple layers of features, resulting in the loss of effective information and the inability to accurately predict disease risk. This is achieved by independently extracting features from different levels of biological information based on metagenomic data and then adaptively fusing multi-level features for prediction.

[0032] Specifically, firstly, the multi-level biological information contained in the sample's metagenomics is transformed and filtered, and matrix construction and feature extraction are performed independently on different levels of biological information to obtain type-level feature representations of biological information at different levels. This can filter out invalid data, reduce redundancy and computational complexity, and retain the unique features of each level of biological information. Compared with single-level feature methods, it can provide more accurate and complete differentiated feature support for subsequent prediction tasks.

[0033] Secondly, based on a multi-head attention mechanism, cross-type fusion is performed on the aforementioned type-level feature representations at different levels. The fused sample representation is then propagated on the fused global graph, and further calibrated based on category prototypes to obtain a fused representation containing multiple levels of features and the topological relationships between them. Compared to simple concatenation, which only superficially superimposes feature information and easily loses hierarchical associations and structural semantics, cross-type fusion can fully model the intrinsic associations and topological dependencies of features at different levels, avoid information redundancy and feature fragmentation, and retain multi-dimensional fine-grained feature semantics. At the same time, the regularity and discriminativeness of the representation are improved through category prototype calibration, outputting a high-quality representation that fuses multi-level features and their mutual topological relationships, providing more stable and discriminative feature support for subsequent prediction tasks.

[0034] Furthermore, the prediction results are further post-calibrated to reduce the prediction disturbances caused by sample noise and structural bias. The prediction bias is corrected from the output end, and the sparsity and stability of the results are improved. This effectively alleviates the problems of insufficient feature utilization, easy prediction distortion, and large fluctuations in results caused by simple splicing of single-level features or multi-level features.

[0035] This effectively eliminates the prediction bias and instability problems caused by insufficient utilization of information features in existing technologies, thereby achieving higher accuracy, stability and versatility in disease risk prediction.

[0036] The technical solution of this application will be described in detail below: In one aspect of the invention, a disease risk prediction method is proposed, referring to... Figure 1 As shown, the method may include: S1000: Based on the metagenomic data of the sample to be tested, construct N feature matrices, each corresponding to a different biological level of microbial characteristics, where N is a positive integer not less than 2.

[0037] In some embodiments, the microbial sources of the metagenomic data of the sample to be tested include, but are not limited to, the host’s gut, oral cavity, skin, respiratory tract, vagina, and urethra, with gut microbiota being preferred.

[0038] In some embodiments, metagenomic data includes, but is not limited to, species abundance profiles, KEGG functional hierarchy profiles, CAZy carbohydrate enzyme profiles, eggNOG functional annotation profiles, antibiotic resistance gene profiles, virulence factor profiles, secondary metabolite profiles, transporter protein profiles, and protein domain profiles, preferably species abundance profiles, KEGG functional hierarchy profiles, and CAZy carbohydrate enzyme profiles. Species abundance profiles preferably include phylum-level abundance profiles and genus-level abundance profiles, and KEGG functional hierarchy profiles preferably include KEGG Level 2 functional profiles and KEGG pathway functional profiles.

[0039] S1100: Preprocess the metagenomic data of the sample to be tested to obtain microbial characteristics at N biological levels.

[0040] S1110: Transform the N biological level data in the metagenomic data of the test sample respectively to obtain N transformed features. In some embodiments, the CLR transformation is used to perform a central logarithmic transformation on the N biological level data to eliminate the compositional constraint of fixed sum. Referring to Equation (1), for the j-th data component of the t-th feature under the i-th sample, a pseudo-count is added to the original count before transformation to avoid the logarithmic zero value problem. In Equation (1), This represents the transformed data of the j-th data component of the t-th feature under the i-th sample. Let j represent the original data of the j-th data component of the t-th feature in the i-th sample. ε The value is 10 -6 The pseudo-count, d t The dimension of the t-th class feature is represented by Let represent the geometric mean of the feature of class t under the i-th sample after pseudo-counting integration of all original components.

[0041] (1).

[0042] S1120: Screen the N transformation features respectively to obtain N biological level microbial features, where the biological level microbial features include K screened microbial features, which are the top K transformation features with the highest variance among the transformation features, and K is a positive integer not less than 1.

[0043] In some embodiments, a variance-based feature selection method is used to screen out features with low variability and retain microbial feature variables with high information content and strong discriminative ability. Referring to equation (2), K microbial features are selected and retained for different categories of features. In equation (2), Let d represent the set of selected features in the t-th feature class, K represent the number of features retained, and d tThis represents the total dimension of the t-th class feature. Let V represent the vector of values ​​of the j-th feature in the t-th class across all samples, and Var(·) represent the variance function.

[0044] (2).

[0045] In some embodiments, the microbial features at the N biological levels are further standardized, including but not limited to Z-score standardization, Min-Max normalization, MaxAbs normalization, Robust Scaling, mean normalization, L2 normalization, and L1 normalization. For example, referring to equation (3), Z-score standardization is performed on the selected features. This represents the standardized value after the standardization transformation. This represents the filtering features obtained after filtering. This represents the calculated average of all selected features in the i-th sample. σ j Let represent the standard deviation of the j-th feature.

[0046] (3).

[0047] S1200: Construct adjacency matrices and correlation matrices for microbial features at N biological levels respectively. The adjacency matrix is ​​used to characterize the feature similarity relationship between the test sample and the reference sample, and the correlation matrix is ​​used to characterize the feature common relationship between the test sample and the reference sample.

[0048] In some embodiments, the reference samples include, but are not limited to, some or all of the training set samples, some or all of the validation set samples, and category prototype samples constructed based on the training set samples.

[0049] In some embodiments, the adjacency matrix is ​​constructed based on a preceding matrix, which includes, but is not limited to, a similarity matrix, a distance matrix, or a correlation matrix, and node connections are determined by threshold filtering or nearest neighbor methods. Similarity matrices include, but are not limited to, Gaussian kernel similarity matrices and cosine similarity matrices; distance matrices include, but are not limited to, Euclidean distance matrices, Manhattan distance matrices, and Mahalanobis distance matrices; correlation matrices include, but are not limited to, Pearson correlation coefficient matrices and Spearman rank correlation coefficient matrices. For example, a cosine similarity matrix is ​​used to construct the preceding matrix, and the cosine similarity between samples is calculated as shown in equation (4). This represents the cosine similarity between sample i and sample j under the t-th class feature. This represents the transformed feature of the t-th class of features of sample i after the aforementioned transformation and filtering. Let represent the transformed features of the t-th type of feature of sample j after the aforementioned transformation and filtering. Then, based on the cosine similarity matrix, adjacency matrices are constructed for microbial features at different biological levels. The matrix size is K×K, where K is the number of microbial features at a certain biological level. Referring to equation (5), for the t-th type of feature, if sample j belongs to one of the m nearest neighbors of sample i, then the matrix elements... The value is the cosine similarity between the two. Otherwise, the value is 0. In equation (5), Let m represent the set of m nearest neighbor samples of sample i.

[0050] (4); (5).

[0051] In some embodiments, the established feature-similar connection edges are symmetricized to make the connection edges between any two samples bidirectionally equivalent, ensuring the symmetry of the connection relationship. The propagation weights of each sample node are balanced, and the influence of the connection edges is adjusted so that the influence of sample nodes with different numbers of connections is balanced in the graph structure, ultimately forming a normalized symmetric adjacency matrix. For example, referring to Equation (6), the adjacency matrix is ​​added to its transpose to obtain a symmetric matrix, and then its degree matrix is ​​calculated. Then, the degree matrix is ​​multiplied by the left and right sides by the negative 1 / 2 power of the degree matrix. In Equation (6), Represents a normalized symmetric adjacency matrix. Degree matrix The corresponding inverse square root diagonal matrix, This represents the adjacency matrix constructed from features of class t. This represents the transpose of the adjacency matrix constructed from the t-th type of features, obtained by interchanging the rows and columns.

[0052] (6).

[0053] In some embodiments, correlation matrices are constructed for microbial characteristics at different biological levels, with each matrix consisting of Y rows and E columns. t A real matrix of columns, where Y is the number of nodes, E t Let be the number of hyperedges for the t-th type of feature.

[0054] In some embodiments, the common feature edges corresponding to the constructed association matrix are organized to clarify the sample range connected by each edge, ensuring that the association relationship can accurately reflect the common feature attributes between samples. The association strength between sample nodes and association edges is balanced, and the influence of different association edges is adjusted to make the propagation characteristics of the association graph structure more stable, ultimately forming a normalized association matrix. For example, the propagation matrix of the association matrix is ​​defined with reference to Equation (7), where, Let represent the propagation matrix of the t-th type of feature, used for information propagation and node feature aggregation on the hypergraph. This represents a diagonal matrix of node degrees, where the diagonal elements represent the weighted sum of the hyperedges in which each node participates. Indicates to The normalized matrix obtained by raising each diagonal element to the power of negative 1 / 2. Represents a hyperdiagonal matrix. Indicates to The normalized matrix obtained by taking the reciprocal of each diagonal element. This represents the correlation matrix corresponding to the t-th type of feature. This represents the transpose of the incidence matrix.

[0055] (7).

[0056] S2000: Construct sample association graphs based on N feature matrices respectively, and extract graph features from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations.

[0057] The sample association graph constructed based on N feature matrices can include: S2110: Based on feature similarity, establish connection edges between the test sample and its nearest neighbor samples, adjust the connection edges, and obtain an adjacency graph. The nearest neighbor samples include reference samples with similar features to the test sample.

[0058] In some embodiments, adjacency graphs include, but are not limited to, k-NN graphs, ε-neighborhood graphs, mutual k-NN graphs, fully connected graphs, relative neighborhood graphs, Delaunay graphs, and minimum spanning trees.

[0059] S2120: Based on the common relationship of features, establish an association edge between the sample to be tested and the common sample, adjust the association edge to obtain an association graph, wherein the common sample includes the reference sample that has common features with the sample to be tested.

[0060] In some embodiments, based on variance characteristics, the top r samples of the t-th type of features that are greater than the variance threshold are connected to construct a hyperedge, and an association graph is further obtained. For example, r is 64.

[0061] In some embodiments, a bipartite graph is constructed based on the t-th type of feature, where one type of nodes are sample nodes and the other type of nodes are feature nodes, and the edges of the bipartite graph represent that a sample possesses that feature. The bipartite graph is then projected, converting each feature node into a hyperedge containing all sample nodes connected to that feature node, thus obtaining an association graph.

[0062] S2130: Weighted blending of adjacency graph and association graph to obtain sample association graph.

[0063] In some embodiments, a weighted mixture is performed based on the adjacency matrix corresponding to the adjacency graph and the association matrix corresponding to the association graph to obtain a hybrid graph matrix, and a hybrid graph is constructed based on the hybrid graph matrix. Refer to equation (8). This represents the sample association graph corresponding to the mixture graph matrix obtained after weighted mixing of features of class t. This represents the adjacency matrix corresponding to the adjacency graph of the t-th type of feature. Let represent the propagation matrix corresponding to the association graph of the t-th feature, and α represent the mixed weight parameter, with a value of 0 ≤ α ≤ 1. For example, α has a value of 0.6.

[0064] (8).

[0065] Specifically, graph feature extraction is performed on the N feature matrices based on the corresponding sample association graph to obtain N feature representations, which may include: S2210: Encode the sample association graph to obtain the hidden representation.

[0066] In some embodiments, the encoding process includes linear transformation projection and optional layer normalization, optional function activation, and optional random deactivation on the association graph of the t-th type of feature samples to obtain a hidden representation. For example, the encoding process sequentially performs linear projection, GELU activation, and layer normalization as shown in equation (9), or sequentially performs linear projection, layer normalization, function activation, and random deactivation as shown in equation (10) to obtain a hidden representation. The hidden representation of the t-th class feature. This represents the projection matrix corresponding to the t-th type of feature. This represents the mixture graph matrix corresponding to the sample association graph of the t-th feature. This represents the activation function. Representation layer normalization, This indicates random inactivation.

[0067] (9); (10).

[0068] S2220: Perform M-stage graph convolution aggregation processing on the hidden representation to obtain the feature representation, where M is a positive integer not less than 1. Graph convolution aggregation processing is used to integrate neighborhood information of node features in the sample association graph.

[0069] In some embodiments, graph convolutional aggregation processing includes convolution and optional function activation, optional residual connections, optional layer normalization, and optional random deactivation to obtain feature representations. For example, a two-stage graph convolutional aggregation process employing the same operations is performed, with the first stage graph convolutional aggregation process shown in Equation (11), the second stage graph convolutional aggregation process shown in Equation (12), and the operation definition for either stage graph convolutional aggregation process shown in Equation (13). Wherein, , The first and second stages of processing represent the processed feature representations of the t-th class, respectively. GNA represents graph convolution aggregation processing. This represents the hidden representation of the aforementioned t-th type of feature. The sample association graph representing the t-th class feature corresponds to the mixture graph matrix. This represents graph convolution used to aggregate neighborhood information. This indicates that the original input hybrid graph matrix and the features of the graph data processing are residually linked at corresponding positions.

[0070] (11); (12); (14).

[0071] S3000: Perform cross-type fusion on N feature representations to obtain the fused representation of the sample to be tested. This step may include: S3100: Based on a multi-head attention mechanism, it performs weighted fusion of N feature representations to obtain sample representations.

[0072] In some embodiments, the N feature representations of all samples are stacked to obtain a stacked vector containing dimensions of sample number, feature representation category number, and feature number. Then, a multi-head attention mechanism is used to capture the correlation between the N feature representations based on the stacked vector to update their respective feature representations. Finally, the importance of different feature representations to the samples is used to obtain the fused features of the samples through weighted summation.

[0073] In some embodiments, the N feature representations of all samples are stacked to obtain a stacked vector containing the sample number dimension, the feature class number dimension, and the feature number dimension. Then, a multi-head attention mechanism is used to capture the correlation between the N feature representations based on the stacked vector to update their respective feature representations. The updated feature representations are then subjected to optional residual connections and optional layer normalization. Finally, the fused features of the samples are obtained by weighted summation based on the importance of different feature representations to the samples. For example, multi-head attention processing is performed with reference to equation (15), residual connections and layer normalization processing are performed with reference to equation (16), and weighted summation is performed with reference to equation (17), where the weight allocation of different feature types of the samples is shown in equation (18). This represents the updated feature representation. This represents the original feature representation in the stacked vectors. This indicates multi-head attention processing. This indicates that residual joins and layer normalization are being performed. This represents the feature representation after residual connection and layer normalization. Let α represent the sample representation of the i-th sample. it w represents the attention weight distribution of the t-th class feature representation of the i-th sample. type Let represent the parameter vector, whose length is the feature dimension of the t-th class feature representation. This means ensuring that the sum of the attention weights of the t-class feature representations of the i-th sample is 1, where T represents the total number of feature representations.

[0074] (15); (16); (17); (18).

[0075] S3200: Based on global type weights, fuse the association graphs of N samples to obtain a global graph.

[0076] In some embodiments, referring to equation (19), the hybrid graph matrices corresponding to the N sample association graphs are weighted and summed based on the global type weights to obtain a global matrix, and a global graph is constructed based on the global matrix. Wherein, Represents the global matrix. This represents the global type weight of the hybrid graph matrix corresponding to the association graph of the t-th class of samples. Let T represent the hybrid graph matrix corresponding to the association graph of the t-th class of samples, where T represents the total number of feature representations.

[0077] (19).

[0078] S3300: Propagate the sample representation on the global graph to obtain the global representation.

[0079] In some embodiments, feature propagation includes convolution and optional function activation, optional residual connections, optional layer normalization, and optional random deactivation to obtain a global representation. Exemplarily, as shown in Equation (20), feature propagation including convolutional function activation, residual connections, layer normalization, and random deactivation is performed to integrate complete neighborhood information. The operation definition of GNA is shown in Equation (14). Indicates global representation, Represents the global matrix. This represents the sample representation.

[0080] (20).

[0081] S3400: Weighted calibration of the global representation based on the category prototype to obtain the fused representation.

[0082] In some embodiments, category prototypes are obtained based on training set samples. The similarity between the current sample and each category prototype is calculated through an attention mechanism to obtain calibrated category prototypes. Then, a fusion representation of the sample is generated through a gating mechanism. For example, an attention mechanism is introduced to calibrate the category prototypes. The fusion representation of the current sample is used as a query, and the category prototypes are used as keys and values. Referring to equations (21), (22), and (23), the weight matrices are transformed respectively. Then, referring to equation (24), the attention similarity score between the sample and each category prototype is calculated. After scaling, the category prototype matching score is obtained. Then, referring to equation (25), the matching score is normalized to obtain the attention weight of the sample to each category prototype. Finally, referring to equation (26), the calibration of each category prototype is completed. Wherein, q i Let i represent the query vector for the i-th sample. This represents the fusion representation of the i-th sample. This represents the query weight matrix. This represents the update key vector of the category prototype c. The original vector representing the prototype c of the category. Represents the key weight matrix. This represents the updated value vector of the category prototype c. d represents the value weight matrix. h Indicates the dimension length of the query / key / value vector. Let represent the attention similarity between the i-th sample and the class prototype c. This represents the attention weights of the i-th sample to each category prototype. This indicates that the attention weights for each type of prototype have been normalized. This represents the calibrated category prototype obtained by weighting the category prototypes.

[0083] (twenty one); (twenty two); (twenty three); (twenty four); (25); (26).

[0084] In some embodiments, a gating mechanism is used to fuse the global representation of the sample, the calibration class prototype, the difference vector between the sample features and the calibration class prototype, and the optional boundary margin to obtain a fused representation of the sample. For example, referring to equations (27) and (28), a gating mechanism is used to fuse the global representation of the sample, the calibration class prototype, the difference vector between the sample features and the calibration class prototype, and the boundary margin to obtain a fused representation of the sample. Indicates the gating coefficient. This represents the Sigmoid activation function, whose output value is between 0 and 1. Represents the gate weight matrix. This represents the fusion representation of the i-th sample. Indicates the prototype of the calibration category. This represents the difference vector between the fused representation and the calibration category prototype. This represents the boundary margin for the two calibration category prototypes. The fusion representation of the samples This indicates element-wise multiplication.

[0085] (27); (28).

[0086] In some embodiments, the aforementioned fusion representation is further supplemented by: encoding the clinical modality data of the test sample to obtain a clinical modality representation; and fusing the aforementioned weighted calibration result with the clinical modality representation to obtain a fusion representation. Exemplarily, the clinical modality data includes, but is not limited to, biochemical data, blood marker data, clinical symptom data, vital sign data, muscle activity state data, limb behavior data, and rhythmic behavior data.

[0087] S4000: Based on fusion representation, disease risk prediction is performed to obtain prediction results.

[0088] In some embodiments, during the training phase of prediction, multi-task learning is employed to optimize the classification and regression tasks. For example, the graph score of the graph network branch is obtained through the fusion representation of the samples, and the prototype score of the prototype calibration branch is obtained by interacting with the category prototype through an attention mechanism using the fusion representation as input. Referring to Equation (29), the classification score is obtained by weighting the aforementioned graph score and prototype score, then normalized to obtain the classification probability, and finally backpropagated through the total loss function to optimize the parameters of each type. This represents the final classification score. This represents the graph score result of the graph network branch. This represents the prototype score result of the prototype calibration branch. This represents the prototype calibration branch weight, with values ​​ranging from 0 to 1. For example, This represents the prototype calibration branch weight with a value of 0.35. The sources of the loss parameters in the total loss function include, but are not limited to, classification loss, Zong scale regression loss, EPDS scale regression loss, prototype branch supervision loss, and regularization loss. The classification loss represents the difference between the final classification prediction result and the true label; the Zong scale regression loss represents the difference between the predicted Zong scale score and the true score; the EPDS scale regression loss represents the difference between the predicted EPDS scale score and the true score; the prototype branch supervision loss represents the difference between the prototype calibration branch classification prediction result and the true label; and the regularization loss represents the complexity of the model parameters to prevent overfitting. The loss parameters in the total loss function include, but are not limited to, cross-entropy, Focal Loss, GHM Loss, CB Loss, symmetric cross-entropy, Triplet Loss, and binary cross-entropy. For example, the total loss function is defined as shown in Equation (30). Indicates the total loss. Represents classification loss. This indicates the regression loss of the Zong scale. This indicates the regression loss of the EPDS scale. This represents the prototype branch supervision loss. Indicates the regularization loss. This represents the prototype consistency loss. This indicates the sparse a priori distillation loss. , , , , , represents the weight coefficients of the regression loss of the Zong scale, the regression loss of the EPDS scale, the prototype branch supervision loss, the regularization loss, the prototype consistency loss, and the sparse prior distillation loss, respectively.

[0089] (29); (30).

[0090] In some embodiments, the prediction results include at least one of disease risk probability, binary classification results, and psychological scale score regression results, wherein the psychological scale score regression results include Zong score regression and / or EPDS score regression, which represent the degree of anxiety or depression.

[0091] In some embodiments, the aforementioned prediction results are corrected to obtain corrected risk prediction results, including at least one of the following: correcting the prediction results based on a sparse prior model to obtain corrected risk prediction results; and / or correcting the prediction results based on a probability calibration model to obtain corrected risk prediction results.

[0092] In some embodiments, the correction process includes: generating a first risk probability based on the prediction result; generating a second risk probability based on a sparse prior model and / or a probability calibration model; and generating a final risk prediction probability based on the first risk probability and the second risk probability.

[0093] In some embodiments, the prediction results are corrected based on a sparse prior model. For example, the fused features of the samples are first input into the sparse prior logistic regression branch, as shown in Equation (31), to perform prediction processing and obtain the sparse prior probability corresponding to L1 regularized logistic regression. Then, the output probability of the graph network branch and the sparse prior probability are weighted and fused as shown in Equation (32) to obtain the corrected risk prediction result. This represents sparse prior probability. Let represent the Sigmoid function, x represent the fusion feature of the samples, w represent the coefficient vector of logistic regression, and b represent the bias term of logistic regression. This represents the dot product of the coefficient vector and the fused features. This indicates the corrected risk prediction result. This represents the output probability of a branch in a graph network. This indicates the dilution of prior fusion weights.

[0094] (31); (32).

[0095] In some embodiments, the prediction results are corrected based on a probabilistic calibration model. The calibration method for the probabilistic calibration model is selected from among various probabilistic calibration methods with the smallest Brier score, and the correction of the prediction results is completed with reference to equation (33). The aforementioned various probabilistic calibration methods include, but are not limited to, Platt scaling, temperature scaling, and conformal regression. This indicates the corrected risk prediction result. This represents the output probability of a branch in a graph network. This represents the output probability of the probabilistic calibration model. This represents the probability calibration weight.

[0096] (33).

[0097] In some embodiments, a variety of anti-overfitting strategies are employed to improve the applicability and stability of the prediction method. The aforementioned anti-overfitting strategies include, but are not limited to: (1) label smoothing in the loss calculation stage to soften hard labels and reduce the risk of overfitting; (2) input perturbation in the forward propagation data input stage to enhance robustness; (3) gradient clipping in the reverse propagation stage to limit the gradient norm to stabilize the training process; (4) R-Drop in the training iteration stage to constrain the consistency of the same batch of data in two forward propagations through KL divergence; and (5) gap constraint in the training monitoring stage to limit the performance difference between the training set and the validation set.

[0098] In another aspect of the invention, a disease risk prediction device is provided, with reference to Figure 2 As shown, the device includes: a feature matrix construction module 100, used to construct N feature matrices based on metagenomic data of the sample to be tested, wherein the N feature matrices correspond to microbial features at different biological levels; a feature representation module 200, used to construct sample association graphs based on the N feature matrices respectively, and to extract graph features from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations; a fusion module 300, used to perform cross-type fusion of the N feature representations to obtain a fused representation of the sample to be tested; and a prediction module 400, used to perform disease risk prediction based on the fused representations to obtain prediction results.

[0099] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, they will not be repeated here. Specifically, this device can execute the above-described corresponding disease risk prediction method embodiments individually or as a whole, and the aforementioned and other operations and / or functions of each unit in the device are respectively for implementing the corresponding processes in the above methods, which will not be repeated here for the sake of brevity.

[0100] In another aspect of the invention, an electronic device is provided, with reference to Figure 3 As shown. The electronic device includes a memory 500 and a processor 600. The memory 500 is used to store computer programs; the processor 600 is used to execute the aforementioned disease risk prediction method. Figure 3 The dashed line indicates that the memory 500 and the processor 600 can transmit data and instructions through dual links, specifically including a physical connection link and a virtual information transmission link.

[0101] For example, processor 600 can be used to execute the steps in the above method according to the instructions in the computer program.

[0102] Processor 600 may include, but is not limited to: General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0103] Memory 500 includes, but is not limited to: Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0104] A computer program can be divided into one or more modules, which are stored in memory 500 and executed by processor 600 to perform the method provided by the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.

[0105] refer to Figure 4 As shown, the electronic device may further include: Transceiver 700, which can be connected to processor 600 or memory 500.

[0106] The processor 600 can control the transceiver 700 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 700 may include a transmitter and a receiver. The transceiver 700 may further include antennas, and the number of antennas may be one or more.

[0107] It should be understood that the various components in an electronic device are connected through a bus system, which includes not only a data bus, but also a power bus, a control bus, and a status signal bus.

[0108] In another aspect, the present invention provides a computer-readable storage medium storing computer instructions or programs that, when executed on a computer, cause the aforementioned disease risk prediction method to be performed.

[0109] In another aspect of the invention, a computer program product is provided, comprising computer instructions that, when some or all of the computer instructions are run on a computer, cause the aforementioned disease risk prediction method to be executed.

[0110] The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0111] Example 1: Prediction of postpartum depression 1. Dataset Postpartum depression was predicted based on gut microbiota metagenomic data from 201 postpartum women. A fixed partitioning strategy was used: 140 cases in the training set, 30 cases in the validation set, and 31 cases in the test set. The metagenomic data included abundance profiles at the phylum and genus levels, KEGG Level 2 functional profiles, KEGG pathway functional profiles, and CAZy carbohydrate enzyme profiles.

[0112] 2. Processing flow 1) CLR transformation and feature screening based on inter-sample variance were performed on metagenomic data. The screening results of microbial features at each biological level are shown in Table 1. In the phylum level abundance spectrum, the top 120 phylum level features with descending variance were retained. In the genus level abundance spectrum, the top 260 genus level features with descending variance were retained. In the Level 2 functional spectrum, all functional features were retained. In the Pathway functional spectrum, the top 90 pathway features with descending variance were retained. In the carbohydrate enzyme spectrum, the top 140 enzyme features with descending variance were retained.

[0113] 2) Project the microbial features of the five biological levels, with a hidden layer dimension of 128, and construct the adjacency matrix and association matrix for the microbial features of the five biological levels independently.

[0114] 3) Construct k-NN and hypergraphs independently for the microbial features of the five biological levels, and then weight and mix the k-NN and hypergraphs of each biological level to obtain the sample association graph.

[0115] 4) Perform feature type-level graph neighborhood aggregation on the sample association graphs of different biological levels to obtain 5 feature representations of the samples.

[0116] 5) Perform cross-type attention fusion and global graph aggregation on the five feature representations to obtain global representations for different samples.

[0117] 6) The global representation is weighted and calibrated based on the category prototype, and the weighted calibration result is enhanced based on clinical modality data to obtain the fusion representation corresponding to different samples.

[0118] 7) Compress the dimension of the fused representation to 64, perform multi-task output, and obtain preliminary prediction results.

[0119] 8) The preliminary prediction results are corrected by using a sparse prior logistic regression model and a probability calibration model to obtain the final prediction result of the disease risk probability.

[0120] Table 1

[0121] 3. Model parameter settings The AdamW optimizer was used, with an initial learning rate of 8×10⁻⁶. -4 A cosine annealing scheduling strategy is used, with a training cycle of 220 rounds. The weight decay coefficient is set to 1×10. -3The early stopping patience value was set to 70. The hidden layer dimension was set to 128, and the dropout rate was set to 0.35. The weight settings are shown in Table 2. The sparse prior logistic regression regularization strength is a parameter applied to the sparse prior logistic regression model to adjust the constraint strength of the L1 regularization term, which determines the degree to which the sparse prior logistic regression model compresses redundant feature weights. The regularization loss weight corresponds to the weight decay coefficient of the AdamW optimizer.

[0122] Table 2

[0123] 4. Evaluation Indicators For classification tasks, accuracy, precision, recall, specificity, F1 score, AUC, and balanced accuracy are used as evaluation metrics. For probability calibration, Brier score is used as the evaluation metric.

[0124] 5. Learning Outcomes The average attention weights for each feature type learned by the model in the proposed method are shown in Table 3. Among them, the abundance spectrum at the genus level contributes the most (0.347), followed by the KEGG Level 2 functional spectrum (0.258), while the CAZy carbohydrate activity spectrum contributes the least (0.065). This result is consistent with biological understanding: species classification information directly reflects the composition and structure of the microbial community, functional spectrum reflects metabolic potential, while CAZy enzyme spectrum information is relatively indirect.

[0125] Table 3

[0126] 6. Test Results (1) Prediction evaluation of the test set Based on the trained prediction device, disease risk prediction was performed on the test set of the aforementioned 31 cases. The evaluation metrics for the classification task are shown in Table 4, and the confusion matrix between the prediction results and the actual situation is shown in Table 5. Among the evaluation metrics for the test set, accuracy and AUC (the ROC curve for the test set is shown in Table 5) are also important. Figure 5 As shown in the figures, the method of the present invention has good classification performance, and the precision, recall, and F1 score indicate that the method of the present invention has balanced and stable performance on both classes of samples. (See reference figures for further details.) Figure 6 As shown, the calibration curve indicates that the predicted probability of the model is close to the true probability, and the Brier score is 0.170, indicating that the probability prediction reliability of the method of the present invention is high.

[0127] Table 4

[0128] Table 5

[0129] (2) Model configuration contribution assessment To verify the effectiveness of different configurations in the method of this invention, ablation experiments were conducted. The experimental results are shown in Table 6. Feature type-level graph aggregation and cross-type attention fusion are the key contributors to model performance. The category prototype attention mechanism improves the model's classification accuracy and F1 score. Although sparse prior fusion cannot improve the classification result evaluation metrics such as accuracy, F1 score, and AUC value, its main function is to supplement the constraints on output probabilities and enhance the stability and interpretability of the model during deployment.

[0130] Table 6

[0131] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is 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.

[0132] 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 changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting disease risk, characterized in that, include: Based on the metagenomic data of the sample to be tested, N feature matrices are constructed, each corresponding to microbial characteristics at a different biological level. A sample association graph is constructed based on the N feature matrices, and graph features are extracted from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations; Cross-type fusion is performed on the N feature representations to obtain the fused representation of the sample to be tested; Based on the fusion representation, disease risk prediction is performed to obtain prediction results; Where N is a positive integer not less than 2.

2. The method according to claim 1, characterized in that, The metagenomic data based on the sample to be tested is used to construct N feature matrices, including: The metagenomic data of the sample to be tested are preprocessed to obtain microbial characteristics at N biological levels; Adjacency matrices and correlation matrices are constructed for the microbial features at the N biological levels, respectively. The adjacency matrix is ​​used to characterize the feature similarity relationship between the test sample and the reference sample, and the correlation matrix is ​​used to characterize the feature common relationship between the test sample and the reference sample.

3. The method according to claim 2, characterized in that, The metagenomic data of the sample to be tested is preprocessed to obtain microbial characteristics at N biological levels, including: The N biological level data in the metagenomic data of the sample to be tested are transformed to obtain N transformation features; The N transformation features are screened to obtain the microbial features at the N biological levels. The microbial features at the biological levels include K screened microbial features, which are the top K transformation features with the highest variance among the transformation features, and K is a positive integer not less than 1.

4. The method according to claim 2 or 3, characterized in that, The construction of sample association graphs based on the N feature matrices includes: Based on the aforementioned feature similarity relationship, connection edges are established between the test sample and its nearest neighbor samples, and the connection edges are adjusted to obtain an adjacency graph. The nearest neighbor samples include the reference samples that have similar features to the test sample. Based on the common relationship of the features, an association edge is established between the sample to be tested and the common sample. The association edge is adjusted to obtain an association graph. The common sample includes the reference sample that has common features with the sample to be tested. The adjacency graph and the association graph are weighted and mixed to obtain the sample association graph.

5. The method according to claim 1, characterized in that, The graph feature extraction is performed on the N feature matrices based on the corresponding sample association graph to obtain N feature representations, including: The sample association graph is encoded to obtain a hidden representation; The hidden representation is subjected to an M-stage graph convolution aggregation process to obtain the feature representation, where M is a positive integer not less than 1. The graph convolution aggregation process is used to integrate neighborhood information of the node features of the sample association graph.

6. The method according to claim 1, characterized in that, Cross-type fusion is performed on the N feature representations to obtain the fused representation of the sample to be tested, including: Based on the multi-head attention mechanism, the N feature representations are weighted and fused to obtain the sample representation; The N sample association graphs are fused based on global type weights to obtain a global graph; The sample representation is propagated on the global graph to obtain a global representation; The global representation is weighted and calibrated based on the category prototype to obtain the fused representation.

7. The method according to claim 6, characterized in that, The cross-type fusion of the N feature representations to obtain the fused representation of the sample to be tested further includes: The clinical modality data of the sample to be tested are encoded to obtain a clinical modality representation; The weighted calibration result is fused with the clinical modality representation to obtain a fused representation.

8. The method according to claim 1, characterized in that, The step of performing disease risk prediction based on the fusion representation to obtain prediction results further includes: The prediction results are corrected to obtain corrected risk prediction results. The correction process includes at least one of the following: The prediction results are corrected based on a sparse prior model to obtain corrected risk prediction results; and / or The prediction results are corrected based on the probability calibration model to obtain the corrected risk prediction results.

9. The method according to claim 8, characterized in that, The correction process includes: A first risk probability is generated based on the prediction results; A second risk probability is generated based on the sparse prior model and / or the probability calibration model; The final risk prediction probability is generated based on the first risk probability and the second risk probability.

10. A disease risk prediction device, characterized in that, include: The feature matrix construction module is used to construct N feature matrices based on the metagenomic data of the sample to be tested, wherein the N feature matrices correspond to microbial features at different biological levels. The feature representation module is used to construct sample association graphs based on the N feature matrices respectively, and to extract graph features from the N feature matrices based on the corresponding sample association graphs to obtain N feature representations; The fusion module is used to perform cross-type fusion on the N feature representations to obtain the fused representation of the sample to be tested; The prediction module is used to predict disease risk based on the fused representation and obtain the prediction result.

11. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the disease risk prediction method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions or programs that, when executed on a computer, cause the disease risk prediction method as described in any one of claims 1 to 9 to be performed.

13. A computer program product, characterized in that, The computer program product includes computer instructions that, when some or all of the computer instructions are run on a computer, cause the disease risk prediction method as described in any one of claims 1 to 9 to be executed.

Images