Gastric cancer / pancreatic cancer risk assessment method and system based on periodontal data

Abstract

The invention relates to the technical field of medical information processing, in particular to a gastric cancer / pancreatic cancer risk assessment method and system based on periodontal data, and the method comprises the steps: obtaining periodontal index data and clinical index data of a user, and extracting a periodontal image feature vector in the periodontal image data; fusing features reflecting similar information in the periodontal index data, the periodontal image feature vector and the clinical index data, and mapping the fused data into nodes in a heterogeneous graph; on the basis of statistical correlation, domain knowledge and biological association, defining connecting edges between different nodes, and constructing initial heterogeneous graph structure data; carrying out aggregation updating processing on the initial heterogeneous graph structure data by utilizing a heterogeneous graph neural network model to obtain feature representation of each node; and performing classification processing based on the updated feature representations of all the nodes, and outputting a risk assessment result that the user belongs to gastric cancer, pancreatic cancer or non-cancer people. And constructing a relational graph model by using the multi-modal periodontal data to realize quantitative evaluation of cancer risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical information processing, in particular to a gastric cancer / pancreatic cancer risk assessment method and system based on periodontal data. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Gastric cancer and pancreatic cancer are common malignant tumors in the digestive system with extremely high malignancy. Due to the insidious early symptoms and the lack of effective early screening methods, most patients are diagnosed at an advanced stage, resulting in poor prognosis and extremely low five-year survival rate.

[0004] Currently, the risk assessment of gastric cancer and pancreatic cancer in clinical practice mainly relies on the detection of specific serum biomarkers, for example, the detection of serum gastrin 17 (G-17), Helicobacter pylori (H. pylori) infection and pepsinogen ratio (PGR) to assess the risk of gastric cancer, and the detection of cancer antigen 19-9 (CA19-9) to assess the risk of pancreatic cancer.

[0005] However, these methods have obvious limitations: on the one hand, the detection cost is high, requiring professional laboratory equipment and reagents, which brings heavy economic burden to large-scale population screening and long-term dynamic monitoring; on the other hand, the detection process is complex and time-consuming, which cannot realize rapid and convenient on-site screening, and is difficult to popularize and apply in areas with relatively scarce medical resources. These limitations have seriously restricted the effective popularization of existing technologies in early screening scenarios.

[0006] Periodontitis is a chronic infectious disease caused by plaque microorganisms that destroys periodontal supporting tissues, which can cause systemic inflammation and immune response. Current studies have shown that periodontitis, as a chronic infectious disease caused by plaque microorganisms, is significantly associated with a variety of systemic diseases, including gastric cancer and pancreatic cancer. Periodontal pathogenic bacteria and their metabolites can migrate to distant organs through the digestive tract or blood circulation, induce local chronic inflammation and activate related carcinogenic signaling pathways, thereby promoting the occurrence and development of tumors.

[0007] However, although the biological association between periodontitis and gastric cancer / pancreatic cancer has been confirmed by multiple studies, most existing researches are limited to qualitative association analysis, lacking specific technical solutions for effectively integrating multi-modal periodontal data (numerical indicators, images) with cancer risk and quantitative assessment. The difficulty lies in: Periodontal data is not a single type; it includes structured data (such as plaque index, probing depth, etc.), unstructured data (periodontal imaging), and clinical data (age, smoking history). These data have completely different scales, dimensions, and semantic spaces, and using conventional simple splicing will lead to information loss and difficulty in model convergence. The relationship between periodontitis and cancer risk is not a simple linear one or an "if-then" rule. It involves complex, non-linear interactions. For example, the cancer risk of an elderly person with mild periodontitis and a young person with severe periodontitis may be completely different. Traditional models such as logistic regression cannot effectively capture this complexity; Qualitative studies alone cannot provide a "risk score," and an uninterpretable score lacks practical value, making it impossible for doctors to make appropriate clinical decisions based on the results of qualitative studies.

[0008] Furthermore, medical data typically has a limited sample size but very high feature dimensionality (especially image features). Training a model directly on a limited sample size can easily lead to overfitting, causing the model to memorize the training data but fail to generalize to new patients. Summary of the Invention

[0009] To address the technical problems mentioned above, this invention provides a method and system for assessing gastric / pancreatic cancer risk based on periodontal data. It constructs a unified relational graph model from multimodal periodontal data (clinical indicators, index scores, and images), and automatically mines the complex nonlinear correlation between periodontitis features and the risk of gastric / pancreatic cancer through deep learning, ultimately achieving a quantitative assessment of cancer risk. This method overcomes the limitations of traditional methods, such as high detection costs and complex procedures, transforming biological qualitative correlations into operable quantitative risk scores, providing a new approach for low-cost and convenient early screening.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for assessing the risk of gastric / pancreatic cancer based on periodontal data, comprising the following steps: Acquire users' periodontal index data, periodontal imaging data, and clinical index data; extract features from periodontal imaging data to obtain periodontal imaging feature vectors; Based on the principle of semantic consistency, features reflecting similar information in periodontal index data, periodontal image feature vectors, and clinical index data are fused, and the fused data is mapped to nodes in a heterogeneous graph. Based on statistical correlation, domain knowledge, and biological association, the edges between different nodes are defined to construct the initial heterogeneous graph structure data. The initial heterogeneous graph structure data is aggregated and updated using a heterogeneous graph neural network model. The aggregation and update process includes: projecting the features of different types of nodes to the same feature space using a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node. Based on the updated feature representations of all nodes, classification processing is performed to output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer, or non-cancer population.

[0011] Furthermore, feature extraction is performed on the periodontal image data to obtain periodontal image feature vectors. Specifically, the pre-trained residual network model takes the periodontal image as input, processes it through multiple residual blocks, and performs feature extraction and nonlinear transformation on the input periodontal image through the convolutional layer, batch normalization layer and activation function in each residual block to obtain the periodontal image feature vector.

[0012] Furthermore, based on the principle of semantic consistency, features reflecting similar information in periodontal index data, periodontal image feature vectors, and clinical index data are fused. Specifically, information related to the set periodontal index is selected from periodontal image features, and the selected information is weighted and averaged with the periodontal index data to generate a comprehensive feature node, and the original feature data is removed.

[0013] Furthermore, periodontal indicators can be defined, including periodontal probing depth or the extent of attachment loss.

[0014] Furthermore, the features of different types of nodes are projected onto the same feature space using a projection matrix, specifically as follows: For periodontal image feature nodes, the first mapping matrix is ​​used to linearly transform the high-dimensional features to the target feature space; For numerical periodontal index data nodes, after Min-Max normalization, they are projected onto the target feature space through the second mapping matrix; For categorical clinical indicator data nodes, after one-hot encoding, they are projected onto the target feature space using a third mapping matrix.

[0015] Furthermore, the first mapping matrix, the second mapping matrix, and the third mapping matrix correspond to different types of nodes, respectively.

[0016] Furthermore, a node-level attention mechanism is used to learn the weights between different nodes, resulting in the feature representation of each node, as shown in the following equation: ; ; in, Represents the weight matrix. and For nodes and Node features after feature projection Represents a node For nodes attention, This represents the activation function. Indicates a connection operation. For another weight matrix, Represents a node and all its neighboring nodes.

[0017] Furthermore, the node-level attention mechanism is extended to a multi-head attention mechanism, which involves repeating the independent attention calculation process multiple times and concatenating the results of the multiple calculations to form the final feature representation of the node.

[0018] Furthermore, classification is performed based on the updated feature representations of all nodes, specifically as follows: Flatten and concatenate the updated feature representations of all nodes to form a one-dimensional feature vector; The one-dimensional feature vector is input into a fully connected layer containing an activation function to obtain the final risk assessment result.

[0019] A second aspect of the present invention provides a gastric / pancreatic cancer risk assessment system based on periodontal data, comprising: The patient information entry module is configured to: acquire the user's periodontal index data, periodontal imaging data, and clinical index data; The data processing module is configured to: extract features from periodontal image data to obtain periodontal image feature vectors; The data processing module is also configured to: fuse features reflecting similar information in periodontal index data, periodontal image feature vectors and clinical index data according to the principle of semantic consistency, and map the fused data into nodes in the heterogeneous graph; define the edges between different nodes based on statistical correlation, domain knowledge and biological association, and construct the initial heterogeneous graph structure data. The data processing module is also configured to: aggregate and update the initial heterogeneous graph structure data using a heterogeneous graph neural network model; the aggregation and update process includes: projecting the features of different types of nodes to the same feature space through a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node; The data processing module is also configured to: perform classification processing based on the updated feature representations of all nodes, and output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer or non-cancer population; The data storage module is configured to store users' periodontal index data, periodontal imaging data, clinical index data, and risk assessment results.

[0020] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: 1. Periodontal indicators and imaging data, which can be conveniently and cost-effectively obtained during routine oral examinations, can replace traditional and expensive serum biomarker tests (such as G-17 and CA19-9). This effectively overcomes the core limitations of existing technologies, which are characterized by "high testing costs and the need for specialized laboratory equipment," resulting in "heavy economic burden" and "difficulty in promoting in resource-scarce areas." It provides a feasible technical approach for large-scale population censuses and community screenings.

[0021] 2. To address the heterogeneity of periodontal data, which includes "structured data, unstructured data, and clinical data" and "completely different scales, dimensions, and semantic spaces," a heterogeneous graph neural network model is constructed to map different types of data into semantically unified nodes. Relationships are defined based on medical knowledge to avoid information loss caused by simple splicing, which is conducive to the deep integration and efficient utilization of multi-source data.

[0022] 3. To address the issue that "the risk relationship between periodontitis and cancer involves complex nonlinear interactions," the node-level attention mechanism in heterogeneous graph neural networks can be used to adaptively learn complex interaction patterns between different features, thereby more accurately quantifying the risk and overcoming the shortcomings of traditional linear models (such as logistic regression) in effectively modeling such complex relationships.

[0023] 4. This solution not only outputs a quantitative risk score, but also reveals the basis behind the score through the structure of the heterogeneous graph and attention weights (for example, indicating that the risk mainly comes from severe attachment loss or specific radiographic features), thereby solving the problems that "purely qualitative studies cannot provide risk scores" and "uninterpretable scores lack practical value", and providing doctors with clear decision support.

[0024] 5. To address the risk of overfitting caused by "limited sample size of medical data but very high feature dimensionality", a pre-trained model is used to extract image features, and an inductive bias is introduced into the model using a graph structure built based on medical prior knowledge. This significantly improves the training efficiency and generalization ability of the model on limited data and ensures the reliability of the evaluation results. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 This is a schematic diagram of a gastric / pancreatic cancer risk assessment process based on periodontal data provided in one or more embodiments of the present invention; Figure 2 This is a schematic diagram of the residual network architecture provided in one or more embodiments of the present invention; Figure 3 This is a schematic diagram of an initial heterogeneous diagram provided in one or more embodiments of the present invention; Figure 4 This is a schematic diagram of the evaluation of a fully connected layer provided in one or more embodiments of the present invention; Figure 5 This is a schematic diagram of the periodontitis data processing procedure provided in one or more embodiments of the present invention; Figure 6 This is a schematic diagram of a gastric / pancreatic cancer risk assessment system based on periodontal data provided in one or more embodiments of the present invention; Figure 7 This is a schematic diagram of the results of the model provided by one or more embodiments of the present invention after 300 rounds of training. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] Example 1: To address the challenge of fusing multimodal heterogeneous data caused by the non-single-type nature of periodontal data, this solution introduces a heterogeneous graph neural network, treating each data type (image features, periodontal indices, and clinical indices) as different types of nodes in the graph. Relationships between nodes are defined based on statistical correlation and medical knowledge (e.g., a strong correlation exists between the "periodontal image feature node" and the "periodontal probing depth node"). Through a specific type of projection matrix, data from different modalities are projected onto the same latent feature space, allowing pixel features of images and numerical features of indices to interact and be compared on the same dimension.

[0030] Addressing the challenge of capturing the risk relationship between periodontitis and cancer, this approach leverages the powerful nonlinear feature mapping capabilities of deep neural networks to extract deep, complex patterns (such as subtle patterns of bone resorption) from periodontal images that are difficult for the human eye to discern. Through a node-level attention mechanism, the model automatically learns the strength (weights) of the associations between different nodes. Because it is data-driven, it eliminates the need for manually pre-defining all rules, thus capturing complex nonlinear relationships.

[0031] Addressing the issue of qualitative research's inability to quantify, the model's decision-making process is made transparent to some extent because its construction is based on medical knowledge. It allows for tracing which nodes (data features) participated in the decision-making process. The importance of nodes can be visualized, and when analyzing results, the abstract concept of "risk" can be transformed into a ranking of the importance of specific, actionable clinical indicators. For example, "The high risk assessment score is mainly due to the very high attention weight given to the nodes of periodontal probing depth and attachment loss, while imaging features indicate extensive bone resorption," thus achieving a leap from qualitative to quantitative analysis.

[0032] To address the issue of limited medical data samples, this approach utilizes a ResNet model pre-trained on a large image dataset (such as ImageNet) for feature extraction. This is a transfer learning strategy that enables the extraction of effective features even on small-scale periodontal imaging data, significantly reducing the risk of overfitting. By introducing a graph structure (definitions of nodes and edges) based on medical knowledge, the model is infused with prior knowledge. This is equivalent to telling the model "these relationships should be followed," reducing the possibility of the model learning spurious associations, thereby improving learning efficiency and generalization ability on small datasets.

[0033] like Figure 1 As shown, the gastric / pancreatic cancer risk assessment method based on periodontal data includes the following steps: Acquire users' periodontal index data, periodontal imaging data, and clinical index data; extract features from periodontal imaging data to obtain periodontal imaging feature vectors; Based on the principle of semantic consistency, features reflecting similar information in periodontal index data, periodontal image feature vectors, and clinical index data are fused, and the fused data is mapped to nodes in a heterogeneous graph. Based on statistical correlation, domain knowledge, and biological association, the edges between different nodes are defined to construct the initial heterogeneous graph structure data. The initial heterogeneous graph structure data is aggregated and updated using a heterogeneous graph neural network model. The aggregation and update process includes: projecting the features of different types of nodes to the same feature space using a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node. Based on the updated feature representations of all nodes, classification processing is performed to output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer, or non-cancer population.

[0034] (1) Periodontal data input: We collected periodontal imaging data, periodontal-related index data (plaque index, soft deposit index, gingival bleeding index, presence or absence of furcation lesions, periodontal probing depth, and extent of attachment loss, etc.), and clinical index data (age, smoking status, skeletal muscle mass, etc.) from three groups of people: those excluded from having cancer, those diagnosed with pancreatic cancer, and those diagnosed with gastric cancer. Individuals without a prior cancer record were randomly selected from the physical examination hall to exclude cancer, and data from those diagnosed with pancreatic / gastric cancer through pathological examination were also collected. (Note: Before data collection, patients were fully informed that the data obtained would be used solely for medical research.)

[0035] Taking "Assessing the relevant risks of gastric cancer" as an example, data from people diagnosed with gastric cancer through pathological examination were selected as the basis.

[0036] (2) Periodontal imaging data processing: Residual networks (ResNet) are used to process periodontal imaging data and extract key image features, providing a foundation for subsequent heterogeneous graph neural network processing.

[0037] ResNet is a deep neural network architecture that uses multiple residual blocks (composed of convolutional layers, normalization layers, and activation functions) to gradually extract deep features from the input data.

[0038] The architecture of ResNet is as follows Figure 2 As shown, during feature extraction, the periodontal image, as input to the ResNet network, undergoes processing through multiple residual blocks. Each residual block contains a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function layer to extract features and perform nonlinear transformations on the input data. It then proceeds to a second 3×3 convolutional layer and a batch normalization layer. The output of each residual block is added to its initial input before entering the next residual block, making it easier for the network to learn the residual mapping. After processing through multiple residual blocks, the final result (image features) is used as nodes in the graph neural network for subsequent data processing.

[0039] (3) Construction of heterogeneous graph structure data: The extracted periodontal imaging features, periodontal related index data (plaque index, soft deposit index, gingival bleeding index, presence or absence of root bifurcation lesions, periodontal probing depth and attachment loss range, etc.) and clinical index data (age, smoking status, skeletal muscle content, etc.) are integrated into nodes of a heterogeneous graph neural network according to the principle of semantic consistency.

[0040] During the integration process, features reflecting similar information in different modal data are fused. For example, the information related to periodontal probing depth in periodontal images and the periodontal probing depth in periodontal index data are weighted and averaged to obtain a comprehensive periodontal probing depth. The original probing depth data in periodontal images and periodontal indexes are removed, and the obtained comprehensive periodontal probing depth is used as a node in the heterogeneous graph.

[0041] After obtaining all heterogeneous graph nodes, the initial heterogeneous graph structure data is constructed by defining edges based on statistical correlation, domain knowledge, and biological associations.

[0042] Graph Neural Networks (GNNs) are deep learning models used to process graph-structured data. Unlike traditional neural networks, GNNs can learn the embedded representations of nodes by utilizing the structural information in the graph. By aggregating and updating the information of nodes and their neighbors, they obtain the feature representation of each node.

[0043] Heterogeneous Graph Neural Networks (HNNs) are a variant of graph neural networks that can effectively handle complex graph structures with different types of nodes and edges.

[0044] In heterogeneous graph neural networks, heterogeneous graphs It consists of a set of nodes and a set of edges It is constructed such that each node and each edge corresponds to a type. (Using...) A collection representing node types. As a set of connected edges.

[0045] A heterogeneous graph can be represented by two mapping functions: one that projects each node to the corresponding type. And project each edge onto the corresponding type This means that the graph contains more than one type of nodes and edges. The initial heterogeneous graph structure data constructed is as follows: Figure 3 As shown.

[0046] (4) Heterogeneous graph neural network aggregation update: In heterogeneous graph neural networks, the weights of each node are learned through aggregation update processing and node-level attention mechanism. Nodes with different weights are aggregated to form node embeddings.

[0047] Due to the heterogeneity of nodes in the heterogeneous graph, different types of nodes (periodontal imaging feature nodes, periodontal related index data nodes, and clinical index data nodes) have different feature spaces. Therefore, before performing aggregation and update, it is necessary to perform feature projection on each type of node and project the features of different types of nodes into the same feature space through a specific type of transformation matrix.

[0048] Specifically, for periodontal imaging feature nodes, their original features are extracted by residual networks, which have high dimensionality and complex structure; the features of periodontal related index data nodes are the values ​​of various periodontal indices; and the features of clinical index data nodes are information such as the patient's age and smoking status.

[0049] For different types of nodes, design corresponding projection matrices.

[0050] For periodontal imaging feature nodes, a single The mapping matrix linearly transforms its high-dimensional features to the target feature space; For periodontal-related index data nodes, after Min-Max normalization processing based on the dimensions and characteristics of each index, the data is mapped using a matrix (…). Projected onto the target feature space; For clinical indicator data nodes, OneHot encoding is first used to uniformly encode the patient's clinical indicator data, and then a mapping matrix is ​​used ( Project it onto the target feature space.

[0051] This targeted projection method enables the effective interaction and learning of features from different types of nodes in periodontal data within a shared feature space. The projection process can be represented by the following formula: ; in, and These are nodes Original features and projected features It is a mapping matrix of a specific type.

[0052] After feature projection is completed, a node-level multi-head attention mechanism is introduced to learn the weights between different nodes.

[0053] Specifically, periodontal imaging feature nodes are correlated with periodontal related indicator data nodes (such as periodontal probing depth nodes) and clinical indicator data nodes (such as age nodes).

[0054] Through the attention mechanism, the model can automatically assign high attention weights to periodontal imaging feature nodes and periodontal probing depth nodes based on potential correlation patterns in the data, or emphasize the correlation weights of periodontal probing depth and attachment loss range, two periodontal indicator nodes, based on existing medical knowledge.

[0055] The formula for calculating node embedding under the attention mechanism is: ; in, Represents the weight matrix. and For nodes and Node features after feature projection Represents a node For nodes Attention, its calculation formula is: ; in, This represents the activation function. Indicates a connection operation. For another weight matrix, Represents a node And all its neighboring nodes. Because heterogeneous graphs are scale-free, the variance of the graph data is too large. To reduce the variance of the graph data, node-level attention is extended to multi-head attention.

[0056] Specifically, the attention of each node is calculated K times, and the results of multiple attention calculations are concatenated to form an embedding vector. The calculation method is as follows: ; in, To represent calculation k The weight matrix at this time. and For nodes and Node features after feature projection Represents a node For nodes attention.

[0057] (5) Processing of fully connected layers: The graph structure data processed by the heterogeneous graph neural network is flattened, and the resulting one-dimensional flattened vector is input into the fully connected layer for further processing.

[0058] Fully connected layers are a basic layer structure in neural networks, often used to integrate and classify features extracted from the previous layers of the neural network.

[0059] In a fully connected layer, each neuron is fully connected to all neurons in the previous layer. That is, each neuron in the layer receives the outputs of all neurons in the previous layer, obtains its own output through weighted summation and nonlinear transformation, and finally connects to the activation function layer to achieve the risk assessment task.

[0060] like Figure 4As shown, the graph structure data obtained after multiple aggregation updates of the heterogeneous graph neural network is flattened. A one-dimensional flattened vector is obtained by concatenating each node. The resulting one-dimensional flattened vector is connected to a fully connected layer, and a preliminary risk assessment result is obtained through forward propagation and Softmax activation. The forward propagation formula is as follows: ; in, It is the first The output vector of the layer, It is the first The weight matrix of the layer, It is the first Layer bias vector, It is the first The activation function of the layer.

[0061] After obtaining the evaluation results of the fully connected layer, the population's own detection results (excluding cancer, diagnosed with pancreatic cancer, diagnosed with gastric cancer) are used as the target value and compared with the results of the fully connected layer. Backpropagation is used to update the weights and biases of the fully connected layer, and finally the gastric cancer / pancreatic cancer risk assessment task is completed.

[0062] Backpropagation refers to the process of calculating gradients layer by layer from the output layer back to the input layer, used to update the network's weights and biases. Its core lies in calculating the gradient information of each layer using the chain rule. The gradient of the output layer is: ; in, For loss function, It is the gradient of the loss function with respect to the output. This is the derivative of the activation function of the output layer.

[0063] The gradient of the hidden layer is: ; in, It is the first The weight matrix of the layer, It is the first The output gradient of the layer, This is the derivative of the activation function of the output layer.

[0064] The update formulas for weights and biases are as follows: ; ; in, For learning rate, , .

[0065] Figure 5The process flow of the periodontal data processing method in this scheme is shown, which is mainly divided into four parts.

[0066] The first part collects patient data, including periodontal imaging, periodontal-related index data (plaque index, soft deposit index, gingival bleeding index, etc.), and clinical index data (age, gender, etc.).

[0067] The second part uses a ResNet residual network to extract features from the patient's periodontal images.

[0068] The third part uses the periodontal image features obtained in the second part, as well as the collected periodontal-related index data and clinical index data, as nodes to construct a heterogeneous graph structure data about periodontal data. The features of each node are projected to the same feature space through feature projection, and a multi-head attention mechanism is introduced to perform multiple aggregation updates.

[0069] The fourth part flattens the output of the third part to obtain a one-dimensional flattened vector. This flattened vector is then used as the input to a fully connected layer for weighted and nonlinear transformation, followed by a Softmax activation layer to obtain the final gastric / pancreatic cancer risk assessment result. Data from 200 subjects is used to preliminarily validate the effectiveness of the proposed method. The results of 300 rounds of model training are shown below. Figure 7 As shown in Table 1, Figure 7 The horizontal axis represents the number of rounds, and the vertical axis represents the percentage of each indicator.

[0070] Table 1 Experimental verification results

[0071] The periodontal data processing method designed in this scheme significantly reduces the economic cost and barriers to early risk assessment of gastric / pancreatic cancer. It avoids the traditional approach that relies on expensive serum biomarkers such as G-17 and CA19-9, instead utilizing readily available and low-cost periodontal indicators (such as plaque index and probing depth) and imaging data from routine oral examinations as the core assessment basis. This successfully overcomes the core limitation of "relatively high testing costs" mentioned in the background section, making large-scale population screening and long-term dynamic monitoring possible.

[0072] This solution achieves effective fusion and deep feature mining of multimodal and heterogeneous periodontal data, solving the data integration challenge. Addressing the heterogeneity challenge of periodontal data, which includes structured, unstructured, and clinical data with completely different scales, dimensions, and semantic spaces, a Heterogeneous Graph Neural Network (HGNN) architecture is used. Numerical indicators, clinical data, and imaging features are mapped to heterogeneous nodes in a graph, and related edges are defined based on medical knowledge. An attention mechanism is then used to dynamically learn the nonlinear interaction weights between different features. This design can capture complex risk patterns that cannot be reflected by a single data type. It solves the technical challenge of "how to effectively integrate multimodal and heterogeneous periodontal health data," overcoming the limitations of traditional methods that can only analyze single types of data in isolation or simply piece together data, thus achieving more accurate quantitative assessment.

[0073] This approach accurately captures the complex nonlinear relationship between periodontitis and cancer risk, improving the accuracy of the assessment model. Addressing the challenge of the complex, nonlinear interaction between periodontitis and cancer risk, it utilizes the node-level attention mechanism in heterogeneous graph neural networks to adaptively learn complex interaction patterns between different features. This allows for more precise risk quantification, overcoming the limitation of traditional linear models (such as logistic regression) in effectively modeling such complex relationships.

[0074] This approach provides a highly interpretable quantitative risk score, offering a clear basis for clinical decision-making. The structure of the heterogeneous graph incorporates domain knowledge, and the node-level attention mechanism outputs the contribution (weights) of different features to the final risk score. This allows physicians to understand the rationale behind the model's decisions (e.g., "a high risk score is mainly due to severe attachment loss and specific radiographic findings"). It overcomes the bottleneck of existing research, which "mostly focuses on qualitative analysis and lacks specific quantitative assessment of risk levels," transforming biological qualitative associations into an intuitive quantitative tool that can serve clinical practice, thus overcoming the difficulty of purely qualitative analysis guiding clinical practice.

[0075] This approach demonstrates strong generalization ability, effectively avoiding overfitting on small-scale medical data. Addressing the overfitting risk caused by the characteristic of medical data having limited sample size but very high feature dimensionality, it employs pre-trained deep networks (such as ResNet) to extract image features and utilizes a graph structure built based on prior medical knowledge to introduce inductive bias into the model. This significantly improves the model's training efficiency and generalization ability on limited data, ensuring the reliability of the evaluation results.

[0076] Example 2: This embodiment takes "assessing the associated risk of pancreatic cancer" as an example to illustrate the data processing procedure for periodontitis indicators in this scheme, including the following steps: Acquire users' periodontal index data, periodontal imaging data, and clinical index data; extract features from periodontal imaging data to obtain periodontal imaging feature vectors; Based on the principle of semantic consistency, features reflecting similar information in periodontal index data, periodontal image feature vectors, and clinical index data are fused, and the fused data is mapped to nodes in a heterogeneous graph. Based on statistical correlation, domain knowledge, and biological association, the edges between different nodes are defined to construct the initial heterogeneous graph structure data. The initial heterogeneous graph structure data is aggregated and updated using a heterogeneous graph neural network model. The aggregation and update process includes: projecting the features of different types of nodes to the same feature space using a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node. Based on the updated feature representations of all nodes, classification processing is performed to output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer, or non-cancer population.

[0077] This embodiment is the same as Embodiment 1 in terms of overall process. In terms of data collection, individuals with no prior cancer records are selected from the physical examination hall to exclude those diagnosed with cancer. Data on individuals diagnosed with pancreatic cancer / gastric cancer are obtained through pathological examination. This embodiment selects data from individuals diagnosed with pancreatic cancer; the remaining steps are the same as in Embodiment 1.

[0078] Example 3: like Figure 6 As shown, the gastric / pancreatic cancer risk assessment system based on periodontal data includes: The patient information entry module is configured to: acquire the user's periodontal index data, periodontal imaging data, and clinical index data; The data processing module is configured to: extract features from periodontal image data to obtain periodontal image feature vectors; The data processing module is also configured to: fuse features reflecting similar information in periodontal index data, periodontal image feature vectors and clinical index data according to the principle of semantic consistency, and map the fused data into nodes in the heterogeneous graph; define the edges between different nodes based on statistical correlation, domain knowledge and biological association, and construct the initial heterogeneous graph structure data. The data processing module is also configured to: aggregate and update the initial heterogeneous graph structure data using a heterogeneous graph neural network model; the aggregation and update process includes: projecting the features of different types of nodes to the same feature space through a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node; The data processing module is also configured to: perform classification processing based on the updated feature representations of all nodes, and output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer or non-cancer population; The data storage module is configured to store users' periodontal index data, periodontal imaging data, clinical index data, and risk assessment results.

[0079] The patient information entry module enables interaction between the user and the risk assessment system. Users fill in periodontal and clinical data, and upload periodontal images. After clicking "Submit Data," the submitted data is integrated and transferred to the data processing module. After receiving patient data, the data processing module processes and assesses risks, outputs and displays the assessment results, and stores the patient's periodontal-related data, clinical data, periodontal images, and risk assessment results in the data storage module for easy viewing and retrieval later.

[0080] This method constructs a unified graph model from multimodal periodontal data (clinical indicators, index scores, and imaging). Deep learning is then used to automatically uncover the complex nonlinear relationships between periodontitis characteristics and the risk of gastric / pancreatic cancer, ultimately achieving a quantitative assessment of cancer risk. This approach overcomes the limitations of traditional methods, such as high detection costs and complex procedures, transforming biological qualitative associations into operable quantitative risk scores, providing a new approach for low-cost and convenient early screening.

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing the risk of gastric / pancreatic cancer based on periodontal data, characterized in that, Includes the following steps: Acquire the user's periodontal index data, periodontal imaging data, and clinical index data; extract features from the periodontal imaging data to obtain periodontal imaging feature vectors; Based on the principle of semantic consistency, features reflecting similar information in the periodontal index data, periodontal image feature vectors, and clinical index data are fused, and the fused data is mapped to nodes in a heterogeneous graph. Based on statistical correlation, domain knowledge, and biological association, the edges between different nodes are defined to construct the initial heterogeneous graph structure data. The initial heterogeneous graph structure data is aggregated and updated using a heterogeneous graph neural network model; The aggregation update process includes: projecting the features of different types of nodes to the same feature space using a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node; Based on the updated feature representations of all nodes, classification processing is performed to output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer, or non-cancer population.

2. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 1, characterized in that, Feature extraction is performed on periodontal image data to obtain periodontal image feature vectors. Specifically, a pre-trained residual network model takes periodontal images as input, processes them through multiple residual blocks, and performs feature extraction and nonlinear transformation on the input periodontal images through convolutional layers, normalization layers and activation functions in each residual block to obtain periodontal image feature vectors.

3. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 1, characterized in that, Based on the principle of semantic consistency, features reflecting similar information in periodontal index data, periodontal image feature vectors, and clinical index data are fused. Specifically, information related to the set periodontal index is selected from periodontal image features, and the selected information is weighted and averaged with the periodontal index data to generate a comprehensive feature node. The original feature data is then removed.

4. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 3, characterized in that, Periodontal indicators should include at least the depth of periodontal probing or the extent of attachment loss.

5. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 1, characterized in that, The features of different types of nodes are projected onto the same feature space using a projection matrix, specifically as follows: For periodontal image feature nodes, the first mapping matrix is ​​used to linearly transform the high-dimensional features to the target feature space; For numerical periodontal index data nodes, after Min-Max normalization, they are projected onto the target feature space through the second mapping matrix; For categorical clinical indicator data nodes, after one-hot encoding, they are projected onto the target feature space using a third mapping matrix.

6. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 5, characterized in that, The first mapping matrix, the second mapping matrix, and the third mapping matrix correspond to different types of nodes, respectively.

7. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 1, characterized in that, By using a node-level attention mechanism to learn the weights between different nodes, the feature representation of each node is obtained, as shown in the following equation: ； ； in, Represents the weight matrix. and For nodes and Node features after feature projection Represents a node For nodes attention, This represents the activation function. Indicates a connection operation. For another weight matrix, Represents a node and all its neighboring nodes.

8. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 7, characterized in that, The node-level attention mechanism is extended to a multi-head attention mechanism, which involves repeating the independent attention calculation process multiple times and concatenating the results of the multiple calculations to form the final feature representation of the node.

9. The gastric / pancreatic cancer risk assessment method based on periodontal data as described in claim 1, characterized in that, Classification is performed based on the updated feature representations of all nodes, specifically as follows: Flatten and concatenate the updated feature representations of all nodes to form a one-dimensional feature vector; The one-dimensional feature vector is input into a fully connected layer containing an activation function to obtain the final risk assessment result.

10. A system for implementing the gastric / pancreatic cancer risk assessment method based on periodontal data as described in any one of claims 1-9, characterized in that, include: The patient information entry module is configured to: acquire the user's periodontal index data, periodontal imaging data, and clinical index data; The data processing module is configured to: extract features from periodontal image data to obtain periodontal image feature vectors; The data processing module is also configured to: fuse features reflecting similar information in periodontal index data, periodontal image feature vectors and clinical index data according to the principle of semantic consistency, and map the fused data into nodes in the heterogeneous graph; define the edges between different nodes based on statistical correlation, domain knowledge and biological association, and construct the initial heterogeneous graph structure data. The data processing module is also configured to: aggregate and update the initial heterogeneous graph structure data using a heterogeneous graph neural network model; the aggregation and update process includes: projecting the features of different types of nodes to the same feature space through a projection matrix, and learning the weights between different nodes using a node-level attention mechanism to obtain the feature representation of each node; The data processing module is also configured to: perform classification processing based on the updated feature representations of all nodes, and output the risk assessment result of whether the user belongs to the gastric cancer, pancreatic cancer or non-cancer population; The data storage module is configured to store users' periodontal index data, periodontal imaging data, clinical index data, and risk assessment results.

Images