Method and system for constructing cancer risk prediction model based on multi-modal data

Abstract

The application relates to the technical field of medical information processing, and discloses a cancer risk prediction model construction method and system based on multi-modal data. The method comprises the following steps: acquiring cancer risk related data and converting the data into numerical values; according to the collection difficulty, data correlation and disease correlation degree, configuring difficulty attribute values, correlation attribute values and association attribute values for each data category and assigning weights, dividing each data category into multiple data layers and taking the data categories in each data layer as data nodes, obtaining a numerical function relationship of the numerical values between the data nodes through neural network training, forming an intermediate relationship model, assigning weight values to each data layer and substituting the numerical values of each case into the intermediate relationship model to calculate and generate a cancer prediction value of each case; adjusting the weight values of each data layer until the numerical distribution concentration degree of the prediction value exceeds a set value, obtaining a target prediction model by combining the weight values of the current data layer with the intermediate relationship model, and improving the accuracy of the model output result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-source medical data information fusion processing, in particular to a cancer risk prediction model construction method and system based on multi-modal data. BACKGROUND

[0002] Cancer, as a serious malignant disease, poses a great threat to people's physical health. In the "Health China Action-Cancer Prevention and Control Action Implementation Plan (2023-2030)", it is clearly stated that cancer prevention, screening, early diagnosis and treatment, and scientific research should be strengthened to effectively reduce the harm of cancer.

[0003] In current practice, in order to improve the level of cancer prevention and screening, many cancer risk prediction models have emerged. The basic principle of these models is usually to collect the health data of residents, and then compare them with the evaluation parameters built in the model to determine the cancer risk of the residents. Obviously, the accuracy of the prediction results of the above-mentioned prediction models is closely related to the type and quantity of the collected health data. The more the type and quantity of the collected health data, the more accurate the prediction results. However, in practical application, it is not easy to collect comprehensive health data of residents, especially data closely related to cancer diagnosis. Taking prostate cancer as an example, the risk of cancer is initially determined according to the patient's clinical manifestations and PSA concentration in the blood, and then further diagnosed and screened through biopsy, MRI, rectal examination and other means. Obviously, in the above-mentioned prostate cancer screening and diagnosis process, the collection of data indicators other than clinical manifestations is difficult, and residents need to go to hospitals or medical institutions, which means that residents cannot collect the above-mentioned data regularly. As a result, when residents find that the above-mentioned data indicators are abnormal through physical examination and other means, they may have already been suffering from cancer, ultimately making the cancer prediction model lose its original significance.

[0004] In summary, in order to improve the accuracy of the output results of the cancer risk prediction model, on the one hand, a large amount of case data needs to be collected during model construction to meet the data requirements of model training; on the other hand, the health data of the individual to be tested needs to be collected regularly and comprehensively, and then input into the prediction model for prediction to output the risk prediction results in a timely and accurate manner. To achieve the above goal, although a large amount of case data can be obtained from the large database of medical or health management institutions for model training during model construction, due to the differences between individual residents, the prediction model generated by the above-mentioned medical big data cannot accurately predict the cancer risk of the current individual. In addition, it is not practical to collect the health data of a certain individual regularly and comprehensively.

[0005] In summary, how to effectively use limited individual health data combined with medical case big data to build a cancer risk prediction model, so that only a small amount of individual health data can be used to accurately predict the risk of cancer in the later stage, is a problem that needs to be solved at present. SUMMARY

[0006] In view of the problem that the prediction accuracy of the cancer risk prediction model in actual application is not high, and a large amount of user health data is needed for prediction, the first object of the present application is to provide a cancer risk prediction model construction method based on multi-modal data, which constructs a special model data architecture, and at the same time uses neural network to train medical case big data, so that the prediction result output by the model is more accurate, and a large amount of health data difficult to collect does not need to be provided by the user in the prediction process, thereby making the entire cancer risk prediction model more easy to popularize and use, and the user can regularly obtain the prediction result of the cancer risk, thereby truly realizing the effect of early prediction. In order to implement the cancer risk prediction model construction method described above, the second object of the present application is to propose a cancer risk prediction model construction system based on multi-modal data, and finally propose a computer readable storage medium for protecting the method of the present application. The specific scheme is as follows:

[0007] A cancer risk prediction model construction method based on multi-modal data, comprising,

[0008] Obtaining each data category and data related to cancer risk from the diagnosed case database, and converting the above data into numerical values and storing them in association with the corresponding data category and case ID;

[0009] According to the difficulty of data collection, the correlation between other categories of data, and the degree of association with the disease, the difficulty attribute value, the correlation attribute value and the association attribute value are configured for each data category respectively;

[0010] The difficulty attribute value, the correlation attribute value and the association attribute value are assigned weights, and the hierarchical reference value corresponding to each data category is calculated and generated;

[0011] According to the numerical interval where the hierarchical reference value is located, each data category is divided into a plurality of data layers;

[0012] Taking the data categories in each data layer as data nodes and based on the numerical values corresponding to each data node, the numerical function relationship between the data nodes of adjacent two data layers is obtained through neural network training, and an intermediate relationship model is formed in combination with the above data layer structure;

[0013] Different weight values are assigned to each data layer, and the numerical values associated with each case ID are substituted into the intermediate relationship model to calculate the cancer prediction value of each case ID;

[0014] adjusting the weight values of each data layer until the numerical distribution concentration degree of the pre-judgment value exceeds a set value, to obtain a target prediction model by combining the weight values of the current data layer with the intermediate relationship model;

[0015] wherein the hierarchical reference value L = W1·S P +W2·C O +W3·R E +k;

[0016] the cancer pre-judgment value ;

[0017] S P is a difficulty attribute value, C O is a correlation attribute value, R E is a correlation attribute value, W1, W2, and W3 are weights of the difficulty attribute value, the correlation attribute value, and the correlation attribute value, respectively, and k is an adjustment coefficient;

[0018] In the cancer pre-judgment value calculation formula, N is the number of data layers, α1, α2, and α N are weight values corresponding to the 1st-N data layers, respectively, m, n, and l are the number of data nodes contained in the 1st-N data layers, respectively, L1 k , L2 k , and LN k respectively represent the hierarchical reference values corresponding to the kth data node in the 1st-N data layers.

[0019] Through the above technical solution, first, the data categories related to the cancer risk are divided into multiple data layers according to the difficulty of data collection and other parameters, forming a preliminary data framework, and then the medical case big data is trained based on the above data framework using a neural network to obtain an intermediate relationship model, and finally the weight values of each data layer are adjusted to make the final cancer judgment value converge to a set interval. The construction method of the above prediction model fully considers the difficulty of collecting various data when the model is applied later, and by constructing an intermediate relationship model, some data that are difficult to collect can be indirectly generated as accurately as possible based on limited health data of the user, and the cancer pre-judgment value is output, so that the result of the cancer risk prediction is more accurate. In addition, in the construction process of the prediction model, not only the numerical relationship between different categories of data is fully considered and integrated, but also the influence of the weight of each data layer on the cancer pre-judgment value is fully considered, that is, the influence of the combination of different data categories on the pre-judgment result, reducing the influence of the numerical amount difference of the same category data in different case data on the overall pre-judgment result, improving the robustness of the prediction model, and improving the accuracy of the output result of the prediction model.

[0020] Further, the numerical function relationship obtained by neural network training includes:

[0021] A first type of function relationship: a numerical function relationship between a single data node in an upper data layer and a single data node in a lower data layer; and

[0022] A second type of function relationship: a numerical function relationship between a combination of multiple data nodes in an upper data layer and a single data node in a lower data layer;

[0023] The second type of function relationship is obtained based on a neural network.

[0024] At least two set data nodes in the same data layer are selected and combined to form a combined data node.

[0025] Each combined data node is configured with a combination algorithm.

[0026] Based on the numerical quantities corresponding to each data node in the combined data node, the combination algorithm is used to calculate the combined numerical quantity of the combined data node.

[0027] The second type of function relationship is obtained based on the combined data node and the combined numerical quantity.

[0028] The above technical solution can reduce the amount of data used for training the intermediate relationship model, and by configuring weights for different categories of data through the combination algorithm, the interference of unreliable data on the training result can be reduced.

[0029] Further, the cancer risk prediction model construction method further includes a model optimization step based on user's own health data, including:

[0030] Each data category and data related to the risk of cancer in the current user data are obtained, and the above data are converted into numerical quantities and stored in association with their corresponding data categories and user IDs.

[0031] The numerical quantities corresponding to the user data are substituted into the intermediate relationship model, and the theoretical values of the current numerical quantities are calculated based on the numerical function relationship between different data nodes.

[0032] The difference between the actual value of each numerical quantity and its corresponding theoretical value is calculated.

[0033] If the difference is within the set range, the current intermediate relationship model is continued to be used.

[0034] If the difference exceeds the set range, case data close to the current user data is matched and obtained from the case database as reinforcement data, and a new intermediate relationship model is trained based on the existing numerical function relationship and the reinforcement data.

[0035] The numerical values corresponding to the reinforcement data are substituted into the new intermediate relationship model to calculate a plurality of cancer prediction values, and the weight values of the data layers are adjusted until the distribution concentration of the plurality of cancer prediction values exceeds a set value, to obtain an optimized target prediction model.

[0036] By the above technical solution, when the intermediate relationship model contained in the target prediction model is not suitable for the current user data, the data similar to the current user data can be selected as reinforcement data for further optimization training by means of medical case big data, and the existing numerical function relationship can be further trained, which can greatly shorten the time required for optimization training, and also make the optimized target prediction model more suitable for the needs of the current user, and improve the accuracy of the cancer risk prediction result.

[0037] Further, the model optimization step is further configured with a model verification step, comprising:

[0038] Obtaining each data category and data related to cancer risk in the current user data, and converting the above data into numerical values and storing them in association with their corresponding data categories and user IDs;

[0039] Finding the data categories stored in the user data and the positions of the above data categories in the intermediate relationship model, and generating at least one verification path composed of a plurality of data nodes extending from the upper data layer to the lower data layer;

[0040] Based on the existing numerical function relationship between adjacent data nodes, a path function is fitted and generated to represent the overall numerical relationship between the data nodes on the verification path;

[0041] The numerical values corresponding to each data node on the verification path are retrieved from the user data and substituted into the path function to calculate the difference between the actual value and the corresponding theoretical value of each numerical value:

[0042] If the difference is within the set range, the current intermediate relationship model is continued to be used;

[0043] If the difference exceeds the set range, the model optimization step is implemented.

[0044] By the above technical solution, the adaptation degree between the target prediction model and the user data can be quickly verified, and then it is decided whether to optimize the target prediction model, which shortens the time required for risk prediction while ensuring prediction accuracy.

[0045] Further, before the neural network training obtains the numerical function relationship between the data nodes of the adjacent two data layers, an optimization screening step of the relationship between the data nodes of the adjacent two data layers is further included, comprising:

[0046] acquire one data category or a combination of multiple data categories in the adjacent data layer, extract the corresponding numerical value and temporarily store it as a relationship judgment database;

[0047] Determine the correlation coefficient between the numerical values of each data node in the adjacent data layer using correlation analysis method;

[0048] If the correlation coefficient between a data node and each data node in the upper and lower data layers is lower than the set threshold, the data node is determined to be a weakly associated data node and is marked;

[0049] Use neural network training to obtain the numerical function relationship between data nodes in the adjacent data layer, excluding the weakly associated data nodes;

[0050] Wherein, the set threshold is manually adjusted or maintained at the initial value before training begins.

[0051] Through the above technical solution, the amount of data participating in training can be reduced, the model training speed can be improved, and the interference of irrelevant data on the target prediction model can be excluded, thereby improving the accuracy of the output results of the target prediction model.

[0052] Further, the weights corresponding to the difficulty attribute value, the correlation attribute value and the association attribute value decrease in turn;

[0053] The number of data nodes contained in the lower data layer is not more than the number of data nodes contained in the upper data layer.

[0054] Through the above technical solution, some data categories that are easy to collect and have strong data correlation are placed in the upper data layer, which is conducive to data collection and can indirectly infer related data through easy-to-collect data, which is conducive to the training of the target prediction model.

[0055] Further, the case data includes one or a combination of multiple types of text data, numerical data, time series data, image data and audio / video data;

[0056] The above data is converted into a numerical value, including:

[0057] According to the data type, configure a corresponding data preprocessing algorithm for each data;

[0058] Extract data features through feature extraction algorithms suitable for each data type;

[0059] Convert the data features corresponding to each data into a numerical vector to generate the numerical value.

[0060] Through the above technical solution, different modal data in the case database can be converted into a numerical value, which is convenient for subsequent construction of a target prediction model.

[0061] A cancer risk prediction model construction system based on multi-modal data, comprising:

[0062] A data transformation storage unit configured to be connected with a case database, for obtaining various data categories and data related to cancer risk from the confirmed case database, and transforming the data into numerical values and storing them in association with the corresponding data categories and case IDs;

[0063] A data layer construction unit configured to be connected with the data transformation storage unit, and configured to configure difficulty attribute values, correlation attribute values and association attribute values for each data category according to the difficulty of data collection, the correlation between data categories and the association between diseases, and assign appropriate weights to each attribute value, calculate the hierarchical reference value corresponding to each data category, and divide each data category into multiple data layers according to the numerical interval of the hierarchical reference value;

[0064] An intermediate model construction unit configured to take the data categories in each data layer as data nodes and based on the numerical values corresponding to each data node, to obtain the numerical function relationship between the data nodes of adjacent two data layers through neural network training, and to form an intermediate relationship model in combination with the data layer structure;

[0065] A target model construction unit configured to assign different weight values to each data layer, to substitute the numerical values associated with each case ID into the intermediate relationship model, to calculate the cancer prediction value of each case ID, to adjust the weight values of each data layer until the numerical distribution concentration degree of the prediction value exceeds a set value, and to generate a target prediction model in combination with the intermediate relationship model using the weight values of the current data layer.

[0066] Further, the cancer risk prediction model construction system further comprises a model optimization module, comprising:

[0067] A user data acquisition unit configured to obtain various data categories and data related to cancer risk in current user data, and to transform the data into numerical values and store them in association with the corresponding data categories and user IDs;

[0068] A model adaptation detection unit configured to be connected with the user data acquisition unit, for substituting the numerical values corresponding to the user data into the intermediate relationship model, calculating the theoretical values of the current numerical values based on the numerical function relationship between different data nodes, and calculating the difference between the actual values of the numerical values and the corresponding theoretical values;

[0069] The model optimization adjustment unit is configured to be in data connection with the model fitness detection unit, judge whether the difference is within the set range, if the difference is within the set range, continue to use the current intermediate relationship model; if the difference exceeds the set range, match and obtain case data close to the current user data from the case database as reinforcement data, train a new intermediate relationship model based on the existing numerical function relationship combined with the above reinforcement data, substitute the numerical quantity corresponding to each reinforcement data into the new intermediate relationship model, calculate and generate a plurality of cancer prediction values, adjust the weight value of each data layer until the distribution concentration degree of the plurality of cancer prediction values exceeds the set value, and obtain the optimized target prediction model.

[0070] A computer-readable storage medium having loaded thereon program modules for implementing the cancer risk prediction model construction method based on multi-modal data as described above.

[0071] In summary, the present application includes at least one of the following beneficial technical effects:

[0072] (1) In the construction of the target prediction model, the difficulty of collecting various data in the later application of the model is fully considered. Through the construction of the intermediate relationship model, the data with high collection difficulty can be indirectly generated as accurately as possible based on the limited health data of the user, and the cancer prediction value is outputted, so that the result of cancer risk prediction is more accurate.

[0073] (2) In the construction process of the prediction model, not only the numerical relationship between different categories of data is fully considered and integrated, but also the influence of the weight of each data layer on the cancer prediction value is fully considered, that is, the influence of the combination of different data categories on the prediction result. The influence of the numerical quantity difference of the same category data in different case data on the overall prediction result is reduced, the robustness of the prediction model is improved, and the accuracy of the output result of the prediction model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is the overall method schematic diagram of the cancer risk prediction model construction method of the present application;

[0075] Figure 2 is a data layer structure schematic diagram;

[0076] Figure 3 is a schematic diagram of the model optimization step based on user data;

[0077] Figure 4 is a schematic diagram of the model verification step;

[0078] Figure 5 is a functional module schematic diagram of the cancer risk prediction model construction system of the present application.

[0079] Reference signs: 1, data conversion storage unit; 2, data layer construction unit; 3, intermediate model construction unit; 4, target model construction unit; 5, user data acquisition unit; 6, model adaptation degree detection unit; 7, model optimization adjustment unit. DETAILED DESCRIPTION

[0080] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0081] In the description of the present specification, the description of the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0082] The embodiment of the present application discloses a cancer risk prediction model construction method based on multi-modal data, as shown in Figure 1 The method mainly includes the following steps:

[0083] S100, obtaining each data category and data related to cancer risk from the confirmed case database, converting the above data into numerical values and storing them in association with their corresponding data categories and case IDs;

[0084] S200, according to the difficulty of data collection, the correlation between other categories of data and the degree of association between the disease, respectively, for each data category, configure the difficulty attribute value, the correlation attribute value and the association attribute value;

[0085] S300, assign weights to the difficulty attribute value, the correlation attribute value and the association attribute value, and calculate the corresponding hierarchical reference value of each data category. Wherein, the hierarchical reference value L=W1·S P +W2·C O +W3·R E +k, S P is the difficulty attribute value, C O is the correlation attribute value, R E is the correlation attribute value, W1, W2, W3 are the weights of the difficulty attribute value, the correlation attribute value and the association attribute value, respectively, and k is the adjustment coefficient, which can be directly specified or adjusted by the staff;

[0086] S400, according to the numerical interval where the hierarchical reference value is located, divide each data category into multiple data layers;

[0087] S500, taking the data categories in each data layer as data nodes and based on the corresponding numerical values of the data nodes, obtaining a numerical function relationship between the data nodes of adjacent two data layers through neural network training, and combining the data layer structure to form an intermediate relationship model;

[0088] S600, assigning different weight values to each data layer, substituting the numerical values associated with each case ID into the intermediate relationship model, and calculating to generate a cancer prediction value for each case ID. In the embodiment of the application, the calculation formula of the cancer prediction value is:

[0089] ;

[0090] wherein N is the number of data layers, a1, a2, a N are the weight values corresponding to the 1st-N data layers respectively, m, n, and l are the number of data nodes contained in the 1st-N data layers respectively, L1 k , L2 k , and LN k respectively represent the hierarchical reference values corresponding to the kth data node in the 1st-N data layers.

[0091] S700, adjusting the weight values of each data layer until the numerical distribution concentration degree of the prediction value M exceeds a set value, and obtaining a target prediction model by combining the weight values of the current data layer with the intermediate relationship model.

[0092] In the above step S100, the case database preferably uses the case data of diagnosed patients collected by each medical institution. The data categories related to the risk of cancer mainly include the basic information of the patient, the daily physiological parameter information of the patient, the daily illness information, the laboratory test data, and the targeted detection result data, etc.

[0093] The basic information of the patient includes the patient's age, living environment (including daily life address, environmental pollution degree, and main pollutants, etc.), daily life habits (such as exercise amount, sleep duration, smoking, drinking, and their frequency, etc.), past medical history, family medical history, TCM four diagnostic information, etc. The daily physiological parameter information of the patient includes the patient's blood pressure, blood sugar, heart rate, and respiratory rate, etc. The above daily physiological parameter information can be collected and stored by wearable devices such as smart bracelets, etc. The daily illness information includes information related to a certain type of illness, such as frequent urination, pain, pelvic discomfort, etc. related to prostate cancer. The laboratory test data includes the detection and test data collected for different diseases, such as PSA / VEGF concentration parameters, etc. The targeted detection result data includes relevant examination data for a certain type of disease, such as puncture biopsy data, rectal examination data, imaging examination data, etc. in the diagnosis process of prostate cancer examination.

[0094] In the embodiments of the present application, the case data required for constructing the target prediction model includes multiple modalities, mainly including one or more combinations of text data, numerical data, time series data, image data, and audio and video data. In the model construction process, the above-mentioned multi-modal data needs to be converted into numerical quantities, and the specific process includes:

[0095] S110, configuring a corresponding data preprocessing algorithm for each data according to the data type. For numerical data, the preprocessing algorithm is to scale the data to a unified range according to the set scaling ratio, and for image data, it includes adjusting the image size, graying, and normalizing, etc., and for text data, it includes data integration, synonym replacement, and audio and video data includes frame division, resolution adjustment, etc.

[0096] S111, extracting data features through feature extraction algorithms applicable to each data type. For image data, edge detection, convolutional neural network, etc. are used to extract data features, and for text data, existing medical large language models can be used for data feature extraction, etc.

[0097] S112, converting the data features corresponding to each of the data into numerical vectors to generate the numerical quantities.

[0098] The above technical solution can convert different modal data in the case database into numerical quantities, which is convenient for the construction of subsequent target prediction models. In actual application, in order to facilitate the calling of the above-mentioned case data by each training model, the above-mentioned case data is converted into numerical quantities and associated with the case ID for storage when stored in the case database.

[0099] In the above step S200, the difficulty of collecting each category of data can be specified by artificial design, for example, the collection of blood pressure and heart rate can be easily completed by wearable detection equipment, while the data such as PSA concentration needs to be detected by blood test, and biopsy data needs to be completed by puncture sampling. Obviously, the collection difficulty of the latter is much greater than that of the former, which means that the above-mentioned data cannot be collected frequently.

[0100] The correlation degree between each category of data is obtained based on medical big data correlation analysis methods such as correlation coefficient analysis, regression analysis, principal component analysis, and factor analysis. Generally, the correlation of each data category in the case data is constant, for example, in the common symptoms of prostate cancer, the numerical quantity used to represent dysuria is often associated with the numerical quantity used to represent urinary weakness. Through correlation analysis methods, the potential correlation between each category of data can be found.

[0101] The degree of association between each type of data and the disease can be manually specified or obtained by big data analysis. For example, PSA concentration data is associated with both prostatitis and prostate cancer. Different PSA concentrations and their changes have specific association probabilities with diseases, while puncture biopsy data has strong association with diseases, i.e., the disease diagnosis result can usually be obtained directly through puncture biopsy.

[0102] In the embodiments of the present application, the attribute values of each type of data in step S200 are configured in the same way. For example, the difficulty attribute values include values 0, 1, and 2, where value 2 represents easy collection, value 1 represents general difficulty, and value 0 represents difficult collection.

[0103] In steps S300-S400, the weights corresponding to the difficulty attribute values, the related attribute values, and the association attribute values are preferably decreased in turn, i.e., w1 > w2 > w3, and the number of data nodes contained in the lower data layer is not more than the number of data nodes contained in the upper data layer.

[0104] Based on the above technical solution, some data categories that are easy to collect are placed in the upper data layer, making the input in the upper data layer more sufficient. Data categories with higher related attribute values are placed in the upper data layer, so that in the case of missing case data, the value quantity of the lower data category can be inferred from the value quantity of the data category in the upper data layer, which is beneficial to data collection and can indirectly infer related data from easy-to-collect data, which is beneficial to the training of the target prediction model and the output of accurate risk prediction based on user data.

[0105] In combination with Figure 2 illustrated, Figure 2 The middle data layer is N layers, and the dashed arrows represent the value function relationship between two data nodes (to avoid unclear illustration, the association relationship between part of the data nodes is illustrated with arrows in the figure). In step S500, based on the diagnosed case big data, the value function relationship between adjacent two data layers of data nodes is obtained by convolutional neural network training. The above function can be a linear function or a non-linear function with multiple activation functions embedded.

[0106] In the embodiments of the present application, the value function relationship obtained by neural network training includes a type of function relationship and a type of function relationship. The type of function relationship is defined as the value function relationship between a single data node in the upper data layer and a single data node in the lower data layer, and the type of function relationship is defined as the value function relationship between a combination of multiple data nodes in the upper data layer and a single data node in the lower data layer, such as Figure 2 The data node L11 and the data node L21 are a type of function relationship, and the combination of the data nodes L13 and L14 and the data node L23 are a type of function relationship.

[0107] In detail, the training of the two-class function relationship based on the convolutional neural network specifically comprises:

[0108] S510, selecting at least two set data nodes in the same data layer for combination to form a combined data node. In actual application, the combination of the data nodes can be manually specified, or the combination can be automatically completed by the system, and then the training effect of each combination is evaluated, and finally a data node combination mode is selected.

[0109] S511, configuring a combination algorithm for each combined data node. In the embodiment of the application, the combination algorithm preferably adopts a weighted algorithm, and different weight values are configured for different data nodes according to the reliability of each data in the combined data node.

[0110] S512, calculating the combined numerical quantity of the combined data node based on the numerical quantity corresponding to each data node in the combined data node and combining the combination algorithm. For example, the combined data node includes data node L13 and data node L14, and the numerical quantity and weight value of the two are 2 / 0.6 and 3 / 0.4 respectively, and the output combined numerical quantity is L1 34 =2*0.6+3*0.4=2.4.

[0111] S513, training the two-class function relationship according to the combined data node generated by recombination and each combined numerical quantity.

[0112] Through the above technical solution, the amount of data used for training of the intermediate relationship model can be reduced, and the weight of different category data is configured through the combination algorithm, so that the interference of unreliable data on the training result can be reduced.

[0113] In actual application, when the cancer risk prediction model is used to predict the cancer risk of different users, the accuracy of the prediction result will fluctuate with the user, and the specific reason is that there are differences in the constitution between different user individuals, for example, for different individuals, the correlation between PSA concentration and prostate cancer is different. Based on the above reason, in order to further improve the accuracy of the target prediction model in predicting the cancer risk of the user, the cancer risk prediction model construction method in the embodiment of the application further comprises a model optimization step based on the user's own health data, as shown in Figure 3 , specifically comprising:

[0114] S520, obtaining each data category and data related to the cancer risk in the current user data, and converting the data into a numerical quantity and storing it in association with the corresponding data category and user ID.

[0115] S521, substituting the numerical quantities corresponding to the user data into the intermediate relationship model, and calculating the theoretical values of the current numerical quantities based on the numerical function relationship between different data nodes.

[0116] S522, calculating the difference between the actual value of each numerical quantity and the theoretical value corresponding thereto.

[0117] S523, comparing the above difference with a set range:

[0118] S5231, if the difference is within the set range, the current intermediate relationship model is continued to be used;

[0119] S5232, if the difference exceeds the set range, the case data close to the current user data is matched from the case database as reinforcement data, and a new intermediate relationship model is trained based on the existing numerical function relationship combined with the above reinforcement data;

[0120] Substitute the numerical quantities corresponding to each reinforcement data into the new intermediate relationship model, calculate a plurality of cancer prediction values, adjust the weight values of each data layer until the distribution concentration degree of the plurality of cancer prediction values exceeds a set value, and obtain an optimized target prediction model.

[0121] In the above step S5232, matching the case data close to the current user data from the case database includes: sorting each data node in the current data layer according to the hierarchical reference value of the data category, forming a data queue, matching the case data from the case database whose numerical quantity deviation of the first N data categories in the above data queue is within the set range, if the obtained data quantity is lower than the preset value, the data quantity used for comparison and matching is modified to N-1, and the above steps are repeated until the data quantity that can be matched and obtained meets the preset requirement, so that the data quantity of the reinforcement data is sufficient to support the subsequent convolutional neural network training.

[0122] In the above step S5232, the new intermediate relationship model trained based on the existing numerical function relationship combined with the above reinforcement data refers to adjusting the bias coefficient or weight in the existing numerical function relationship, and generating a new numerical function relationship that can represent the relationship between the reinforcement data.

[0123] In the above technical solution, when the intermediate relationship model contained in the target prediction model is not suitable for the current user data, the data close to the current user data can be selected as reinforcement data for further optimization and training by means of medical case big data, and the existing numerical function relationship can be further trained, which can greatly shorten the time required for optimization and training, and make the optimized target prediction model more suitable for the needs of the current user, and improve the accuracy of the cancer risk prediction result.

[0124] In order to shorten the time required for determining the degree of adaptation between the target prediction model and the current user data, the model verification step is further configured before the model optimization step in the embodiments of the present application, as shown in the following figure: Figure 4

[0125] S530, obtaining each data category and data related to the risk of cancer in the current user data, and converting the above data into numerical values and storing them in association with the corresponding data category and user ID.

[0126] S531, finding the data category stored in the user data and the position of the above data category in the intermediate relationship model, and generating at least one verification path composed of multiple data nodes extending from the upper data layer to the lower data layer. As shown in the following figure: Figure 2

[0127] S532, based on the existing numerical function relationship between adjacent data nodes, fitting to generate a path function for representing the overall numerical relationship between each data node on the verification path. For example, the numerical function relationship between two adjacent data nodes is: y=f(x)=2x+2, x=h(a)=5a+1, then the integrated y=f(a)=10a+4, that is, the value of each data node on the verification path can be expressed as a relationship based on the value of a.

[0128] S533, retrieving the numerical value corresponding to each data node on the verification path from the user data and substituting it into the path function, and calculating the difference between the actual value and the corresponding theoretical value of each numerical value:

[0129] S534, comparing the above difference with the set range:

[0130] S5341, if the difference is within the set range, the current intermediate relationship model is continued to be used;

[0131] S5342, if the difference exceeds the set range, the model optimization step is implemented.

[0132] In practical applications, the number of verification paths selected above is not less than 3.

[0133] The above technical solution can quickly verify the degree of adaptation between the target prediction model and the user data, and then decide whether to optimize the target prediction model, thereby shortening the time required for risk prediction while ensuring prediction accuracy.

[0134] ​​In step S500, in order to reduce the amount of data involved in training, improve the speed of model training, and exclude irrelevant data from interfering with the target prediction model, in the embodiments of the present application, before obtaining the numerical function relationship between the data nodes of the adjacent two data layers through neural network training, an optimization and screening step of the relationship between the data nodes of the adjacent two data layers is further included, which specifically includes:

[0135] S540, one or more data categories in the adjacent data layers are obtained respectively, and the corresponding numerical quantities are extracted and temporarily stored as a relationship determination database;

[0136] S541, the correlation analysis method is used to determine the correlation coefficient between the numerical quantities corresponding to each data node in the adjacent data layers;

[0137] S542, if the correlation coefficient between a data node and each data node in the upper and lower data layers is lower than a set threshold, the data node is determined to be a weakly associated data node and is marked, such as the L12 data node and the L31 data node shown in Figure 2 . The set threshold is manually adjusted or maintained at the initial value before the training starts.

[0138] S543, the neural network training is used to obtain the numerical function relationship between the data nodes in the adjacent data layers, excluding the weakly associated data nodes.

[0139] By marking the weakly associated data nodes, it is convenient to remind the staff to pay attention to some data categories with strong independence, and to improve the accuracy of the output results of the target prediction model.

[0140] In order to implement the construction method of the above-mentioned target prediction model, the embodiments of the present application also propose a cancer risk prediction model construction system based on multi-modal data, as shown in Figure 5 , mainly including a data transformation storage unit 1, a data layer construction unit 2, an intermediate model construction unit 3, and a target model construction unit 4.

[0141] The data transformation storage unit 1 is configured to be connected with a case database, and is used to obtain each data category and data related to the cancer risk from the confirmed case database, and to transform the above-mentioned data into numerical quantities and store them in association with the corresponding data categories and case IDs. The above-mentioned case database is set in the cloud, which is convenient for different medical institutions to upload different case data to the cloud and store them. Preferably, the above-mentioned case database uses a confirmed case database and each case data includes the same data categories. The algorithm modules for transforming each type of data into numerical quantities are all configured in the cloud server, and different modal data information is processed based on a unified preprocessing algorithm, and the corresponding data features are extracted and transformed into numerical quantities.

[0142] The data layer construction unit 2 is configured to be in data connection with the data transformation storage unit 1, and to configure an easy attribute value, a correlation attribute value and an association attribute value for each data category according to the data collection difficulty, the correlation between the data categories and the association degree between the diseases, and to assign a corresponding weight to each attribute value, to calculate a hierarchical reference value corresponding to each data category, and to divide each data category into a plurality of data layers according to the numerical interval of the hierarchical reference value. In specific practice, the data layer construction unit 2 also has an interactive module for interaction with external personnel, which is mainly used to receive input information from external staff in specific situations, and to determine the easy attribute value or the association attribute value corresponding to each data category by manual operation.

[0143] The intermediate model construction unit 3 is configured to take the data categories in each data layer as data nodes and to obtain a numerical function relationship between the data nodes of adjacent two data layers based on the numerical amount corresponding to each data node through neural network training, and to form an intermediate relationship model in combination with the data layer structure. In the embodiments of the present application, the intermediate model construction unit 3 preferably uses an external convolutional neural network training platform to train the model.

[0144] The target model construction unit 4 is configured to assign different weight values to each data layer, to substitute the numerical amount associated with each case ID into the intermediate relationship model, to calculate a cancer prediction value of each case ID, to adjust the weight values of each data layer until the numerical distribution concentration degree of the prediction value exceeds a set value, and to generate a target prediction model in combination with the intermediate relationship model using the weight value of the current data layer.

[0145] In order to improve the adaptability of the target prediction model to user data, the cancer risk prediction model construction system further comprises a model optimization module, mainly including a user data acquisition unit 5, a model adaptation detection unit 6 and a model optimization adjustment unit 7.

[0146] The user data acquisition unit 5 is configured to acquire each data category and data related to the cancer risk in the current user data, and to convert the above data into a numerical amount and store it in association with the corresponding data category and user ID. In detail, the user data acquisition unit 5 has a data input module and a screening module, which are respectively used to receive external input user data such as clinical manifestation data or user basic information, and the screening module is used to preliminarily filter the above user data to remove obviously incorrect or abnormal data and data categories that do not match the data categories in the model, such as patient tongue fur color.

[0147] The model fitness detection unit 6 is configured in data connection with the user data acquisition unit 5, for substituting the numerical quantities corresponding to the user data into the intermediate relationship model, calculating the theoretical values of the current numerical quantities based on the numerical function relationship between different data nodes, and calculating the difference between the actual values of the numerical quantities and the theoretical values corresponding thereto.

[0148] The model optimization adjustment unit 7 is configured in data connection with the model fitness detection unit 6, for judging whether the difference is within the set range, and if the difference is within the set range, continuing to use the current intermediate relationship model. If the difference is beyond the set range, matching and acquiring case data close to the current user data from the case database as reinforcement data, training a new intermediate relationship model based on the existing numerical function relationship combined with the above reinforcement data, substituting the numerical quantities corresponding to the reinforcement data into the new intermediate relationship model, calculating to generate a plurality of cancer prediction values, adjusting the weight values of each data layer until the distribution concentration degree of the plurality of cancer prediction values exceeds a set value, and obtaining an optimized target prediction model.

[0149] The cancer risk prediction model construction system body described in the embodiments of the present application is preferably configured in a cloud server, facilitating data analysis, and the user data acquisition unit 5 mainly cooperates with offline health data acquisition equipment or input equipment to acquire user data.

[0150] In order to facilitate the popularization and use of the model construction method described in the embodiments of the present application, the present application also proposes to protect a computer readable storage medium having a program module for implementing the cancer risk prediction model construction method based on multi-modal data as described above loaded thereon.

[0151] The above computer readable storage medium includes but is not limited to disk memory, CD-ROM, optical memory, etc.

[0152] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for constructing a cancer risk prediction model based on multimodal data, characterized in that, include, Retrieve data categories and data related to cancer risk from the confirmed case database, convert the above data into numerical values ​​and store them in association with their corresponding data categories and case IDs; Based on the ease of data collection, the correlation with other data categories, and the degree of association with the disease, difficulty attribute values, relevance attribute values, and association attribute values ​​are configured for each data category. Weights are assigned to the difficulty attribute values, related attribute values, and associated attribute values, and hierarchical reference values ​​are calculated and generated for each data category. Each data category is divided into multiple data layers based on the numerical range of the hierarchical reference value. Using data categories in each data layer as data nodes and based on the numerical values ​​corresponding to each data node, the numerical function relationship between data nodes in two adjacent data layers is obtained through neural network training, and an intermediate relationship model is formed by combining the above data layer structure. Different weight values ​​are assigned to each data layer, and the numerical values ​​associated with each case ID are substituted into the intermediate relationship model to calculate and generate the cancer prediction value for each case ID. Adjust the weight values ​​of each data layer until the concentration of the numerical distribution of the predicted value exceeds a set value, and combine the weight values ​​of the current data layer with the intermediate relationship model to obtain the target prediction model; Wherein, the layered reference value L = W1·S P +W2·C O +W3·R E +k; Cancer prediction value ; S P For the difficulty attribute value, C O For the relevant attribute values, R E For the relevant attribute values, W1, W2, and W3 are the weights of the difficulty attribute value, the relevant attribute value, and the associated attribute value, respectively, and k is the adjustment coefficient; In the formula for calculating the cancer prediction value, N is the number of data layers, α1, α2, α... N These are the weight values ​​corresponding to the first to Nth data layers, respectively, where m, n, and l are the number of data nodes contained in the first to Nth data layers, respectively. k L2 k LN k These represent the hierarchical reference values ​​corresponding to the k-th data node in the 1st to Nth data layers, respectively.

2. The method for constructing a cancer risk prediction model based on multimodal data according to claim 1, characterized in that, Numerical function relationships obtained through neural network training include: One type of functional relationship: the numerical functional relationship between a single data node in the upper data layer and a single data node in the lower data layer; and Type II functional relationships: numerical functional relationships between combinations of multiple data nodes in the upper data layer and single data nodes in the lower data layer; The two types of functional relationships obtained based on neural network training include: Select at least two defined data nodes from the same data layer and combine them to form a combined data node; Configure a combination algorithm for each combined data node; Based on the numerical values ​​of each data node within the combined data node, the combined numerical value of the combined data node is calculated and output using the combined algorithm. Based on the combined data nodes generated by recombination and the numerical values ​​of each combination, the two types of function relationships are trained and generated.

3. The method for constructing a cancer risk prediction model based on multimodal data according to claim 2, characterized in that, The method for constructing the cancer risk prediction model also includes a model optimization step based on the user's own health data, including: Retrieve the data categories and data related to cancer risk from the current user data, convert the above data into numerical values ​​and store them in association with their corresponding data categories and user IDs; Substitute the numerical values ​​corresponding to the user data into the intermediate relationship model, and calculate the theoretical values ​​of each current numerical value based on the numerical function relationship between different data nodes; Calculate the difference between the actual value and the corresponding theoretical value of each numerical quantity: If the difference is within the set range, the current intermediate relationship model will continue to be used; If the difference exceeds the set range, then case data that is close to the current user data is matched from the case database as reinforcement data, and a new intermediate relationship model is trained based on the existing numerical function relationship and the above reinforcement data. Substitute the numerical values ​​corresponding to each reinforcement data into the new intermediate relationship model to calculate and generate multiple cancer prediction values. Adjust the weight values ​​of each data layer until the distribution concentration of the multiple cancer prediction values ​​exceeds the set value to obtain the optimized target prediction model.

4. The method for constructing a cancer risk prediction model based on multimodal data according to claim 3, characterized in that, The model optimization step is preceded by a model validation step, including: Retrieve the data categories and data related to cancer risk from the current user data, convert the above data into numerical values ​​and store them in association with their corresponding data categories and user IDs; Find the data categories stored in the user data and their positions in the intermediate relationship model, and generate at least one verification path consisting of multiple data nodes extending from the upper data layer to the lower data layer; Based on the existing numerical function relationships between adjacent data nodes, a path function is fitted and generated to characterize the overall numerical relationship between each data node on the verification path. The numerical values ​​corresponding to each data node on the verification path are retrieved from the user data and substituted into the path function to calculate the difference between the actual value and its corresponding theoretical value: If the difference is within the set range, the current intermediate relationship model will continue to be used; If the difference exceeds the set range, then the model optimization step is implemented.

5. The method for constructing a cancer risk prediction model based on multimodal data according to claim 1, characterized in that, Before obtaining the numerical function relationship between data nodes in two adjacent data layers through neural network training, there is also an optimization and filtering step for the data node relationship between two adjacent data layers, including: Each data category or combination of data categories in an adjacent data layer is obtained, and its corresponding numerical value is extracted and temporarily stored in a relation determination database. The correlation coefficient between the corresponding numerical values ​​of each data node in adjacent data layers is determined by using correlation analysis. If the correlation coefficient between a data node and each data node in the upper and lower data layers is lower than a set threshold, the data node is determined to be a weakly correlated data node and marked. The weakly correlated data nodes are excluded by using neural network training to obtain numerical function relationships between data nodes in adjacent data layers. The threshold value is manually adjusted or maintained at its initial value before training begins.

6. The method for constructing a cancer risk prediction model based on multimodal data according to claim 1, characterized in that, The weights corresponding to the difficulty attribute value, the related attribute value, and the associated attribute value decrease sequentially. The number of data nodes in the lower data layer is no more than the number of data nodes in the upper data layer.

7. The method for constructing a cancer risk prediction model based on multimodal data according to claim 1, characterized in that, The case data includes one or more combinations of text data, numerical data, time series data, image data, and audio / video data; Convert the above data into numerical values, including: Configure corresponding data preprocessing algorithms for each data type; Extract data features using feature extraction algorithms applicable to each data type; The data features corresponding to each of the data are converted into numerical vectors to generate the numerical quantity.

8. A system for constructing a cancer risk prediction model based on multimodal data, characterized in that, A method for constructing a cancer risk prediction model based on multimodal data as described in any one of claims 1-7, comprising: The data conversion and storage unit (1) is configured to connect to the case database and is used to obtain various data categories and data related to cancer risk from the confirmed case database, convert the above data into numerical values ​​and store them in association with their corresponding data categories and case IDs. The data layer construction unit (2) is configured to connect with the data transformation and storage unit (1) according to the difficulty of data collection, the correlation between the data and other categories, and the degree of association with the disease. It configures the difficulty attribute value, the related attribute value and the associated attribute value for each data category and assigns corresponding weights to each attribute value. It calculates and generates the hierarchical reference value corresponding to each data category and divides each data category into multiple data layers according to the numerical range of the hierarchical reference value. The intermediate model construction unit (3) is configured to take the data categories in each data layer as data nodes and, based on the numerical values ​​corresponding to each data node, obtain the numerical function relationship between the data nodes of two adjacent data layers through neural network training, and form an intermediate relationship model in combination with the data layer structure. The target model construction unit (4) is configured to assign different weight values ​​to each data layer, substitute the numerical values ​​associated with each case ID into the intermediate relationship model, calculate and generate the cancer prediction value of each case ID, adjust the weight values ​​of each data layer until the numerical distribution concentration of the prediction value exceeds the set value, and generate the target prediction model by combining the weight values ​​of the current data layer with the intermediate relationship model.

9. The cancer risk prediction model construction system based on multimodal data according to claim 8, characterized in that, The cancer risk prediction model construction system also includes a model optimization module, including: The user data acquisition unit (5) is configured to acquire the data categories and data related to cancer risk in the current user data, convert the above data into numerical values ​​and store them in association with their corresponding data categories and user IDs. The model fit detection unit (6) is configured to connect with the user data acquisition unit (5) to substitute the numerical values ​​corresponding to the user data into the intermediate relationship model, calculate the theoretical values ​​of each current numerical value based on the numerical function relationship between different data nodes, and calculate the difference between the actual value of each numerical value and its corresponding theoretical value. The model optimization and adjustment unit (7) is configured to be connected to the model fit detection unit (6) to determine whether the difference is within a set range. If the difference is within the set range, the current intermediate relationship model is continued to be used. If the difference exceeds the set range, case data that is close to the current user data is matched from the case database as reinforcement data. A new intermediate relationship model is trained based on the existing numerical function relationship and the reinforcement data. The numerical values ​​corresponding to each reinforcement data are substituted into the new intermediate relationship model to calculate and generate multiple cancer prediction values. The weight values ​​of each data layer are adjusted until the distribution concentration of multiple cancer prediction values ​​exceeds the set value, and the optimized target prediction model is obtained.

10. A computer-readable storage medium, characterized in that, It is loaded with a program module for implementing the cancer risk prediction model construction method based on multimodal data as described in any one of claims 1-7.

Images