Multi-gene molecular diagnosis model as well as construction method and application thereof

Abstract

The invention relates to the technical field of bioinformatics and medical data processing, and discloses a polygene molecular diagnosis model and a construction method and application thereof.The construction method comprises the steps that a data acquisition module constructs a genome, clinical phenotype and environmental exposure data matrix, a preprocessing module executes regression filling and standardizes continuous variables, and a data processing module performs data processing; the feature screening module executes double-layer screening by using LASSO and a random forest model to output a core feature subset, the model building module builds a logic regression architecture to calculate a baseline logarithm probability, and the dynamic updating module outputs a real-time risk probability in combination with follow-up visit environment data, a time adjustment coefficient and the baseline logarithm probability. And the interactive output module outputs a risk layering label and a feature contribution degree. According to the method, redundancy is eliminated through double-layer screening, the time dimension is introduced to adjust the real-time correction probability, visual attribution is realized in combination with the SHAP algorithm, and the dynamic monitoring capability and interpretability are improved.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and medical data processing technology, specifically to a multi-gene molecular diagnostic model, its construction method, and its application. Background Technology

[0002] With the popularization of high-throughput sequencing technology and the development of precision medicine, predicting disease risk based on genetic information has become an important means of chronic disease prevention and control. Especially in the field of complex polygenic genetic diseases such as type 2 diabetes and cardiovascular disease, using machine learning algorithms to build diagnostic models can help doctors identify high-risk groups and formulate intervention strategies.

[0003] However, existing molecular diagnostic technologies face technical bottlenecks in clinical applications. Current risk assessment models rely on a single data dimension, calculating static genetic risk based solely on genomic data or assessing current status solely based on clinical biochemical indicators. This approach ignores the interaction between genetic background and environmental exposures such as diet and exercise, leading to biased prediction results. Since germline genomic information is lifelong and unchanging, existing technologies can only provide a one-time baseline risk assessment and cannot respond to risk fluctuations caused by lifestyle changes during follow-up, making it difficult to meet the actual needs of long-term dynamic health management.

[0004] Furthermore, biomedical big data is characterized by high dimensionality, sparsity, and high noise. Directly inputting massive gene loci into the model can easily lead to the curse of dimensionality and model overfitting. Existing feature selection methods are difficult to simultaneously consider the statistical significance of features and nonlinear classification capabilities. More importantly, with the introduction of complex algorithms, diagnostic models become uninterpretable black boxes. Medical personnel cannot intuitively understand the specific contribution direction and weight of certain features to the prediction results. The lack of interpretability makes it difficult for clinicians and patients to trust the diagnostic results, thus limiting the further promotion of multi-gene molecular diagnostic models in precision medicine scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-gene molecular diagnostic model, its construction method, and its application, which solves the problems of existing molecular diagnostic technologies that are based solely on static data evaluation, lack multimodal data fusion, and lack interpretability of black-box models.

[0006] To achieve the above objectives, the first aspect of the present invention provides a multi-gene molecular diagnostic model, including a data acquisition module, a preprocessing module, a feature screening module, a model construction module, a dynamic update module, and an interactive output module.

[0007] The data acquisition module is configured to acquire the raw dataset of the subject population and construct a genomic data matrix, a clinical phenotype data matrix, and an environmental exposure data matrix based on the raw dataset.

[0008] The preprocessing module is connected to the data acquisition module and is configured to perform regression imputation on missing feature values ​​in the data and perform standardization on continuous variables, outputting a fused feature matrix. In the regression imputation logic, for features with missing values, the preprocessing module constructs a linear regression model using other complete features in the original dataset as independent variables and calculates the missing feature values. The preprocessing module accumulates the intercept term of the linear regression model with the weighted terms of each complete feature participating in the prediction, where each weighted term is defined as the product of the regression coefficient of the complete feature and its corresponding observation value.

[0009] The feature selection module is connected to the preprocessing module. The feature selection module includes a first-level selection unit and a second-level selection unit. The first-level selection unit is configured to perform a compression operation on the regression coefficients by introducing an L1 regularization penalty term, retaining features with non-zero coefficients to generate a candidate feature set. The second-level selection unit is configured to quantify the Gini importance score of each feature based on the decrease in Gini impurity during the splitting process of the decision tree node, and extract a subset of core features based on the ranking result of the Gini importance scores.

[0010] The model building module is connected to the feature selection module and is configured to establish a logistic regression operation architecture based on the core feature subset to calculate the baseline log odds of the subjects.

[0011] The dynamic update module is connected to the model building module and is configured to combine follow-up environmental exposure data, time adjustment coefficients, and the baseline log-odds ratio to output a real-time disease risk probability. The logic of the dynamic update module in calculating the real-time disease risk probability is as follows: First, the training weight coefficients of each feature in the environmental exposure feature subset are calculated as a weighted cumulative value of their standardized changes. The weighted cumulative value is then multiplied by the time adjustment coefficient. Next, the resulting product is added to the baseline log-odds to obtain the dynamic log-odds. Finally, the dynamic log-odds is mapped to a probability interval of zero to one using the nonlinear transformation logic of the sigmoid function to obtain the real-time disease risk probability.

[0012] The interactive output module is connected to the dynamic update module and is configured to output the risk stratification label and feature contribution corresponding to the real-time disease risk probability. The logic for calculating the feature contribution of the interactive output module is as follows: Based on the definition of the SHAP algorithm, a feature subset combination that does not contain the target feature is constructed. The marginal contribution value of the addition of the target feature to the model prediction output is evaluated. Based on the positive or negative polarity and numerical magnitude of the marginal contribution value, the target feature is determined to be a positive risk driving factor or a negative risk protection factor.

[0013] A second aspect of this invention provides a method for constructing a multi-gene molecular diagnostic model, comprising the following steps: S1: Construct the genomic data matrix, the clinical phenotype data matrix, and the environmental exposure data matrix using the data acquisition module; S2: Output the fused feature matrix using the preprocessing module. In this step, the normalization process employs Z-score normalization, and the specific process includes: For each continuous variable, the arithmetic mean and standard deviation of the continuous variable in the overall sample space are statistically calculated. For the original value of each sample point, a centering operation is performed to determine the deviation of the original value from the arithmetic mean. A scaling operation is performed, and the data distribution is converted into a standard normal distribution with a mean of zero and a variance of one by calculating the ratio of the deviation to the standard deviation. S3: The feature filtering module is used to perform a two-layer filtering on the fused feature matrix to generate the core feature subset; S4: Calculate the baseline log odds using the model building module; S5: Utilize the dynamic update module to receive follow-up environmental exposure data, calculate the standardized change of environmental characteristics, and output the real-time disease risk probability. In this step, the calculation process of the standardized change specifically includes: The baseline original measurement value of the subject at the time of model construction is established as the reference value. The real-time measurement value obtained at the current follow-up time is compared with the reference value to obtain the absolute difference value reflecting the longitudinal change of the individual. The statistical dispersion index of the subject group on which the model is based in terms of environmental characteristics, i.e., the baseline standard deviation, is retrieved. The ratio of the absolute difference value to the baseline standard deviation is calculated and normalized to eliminate the influence of the dimension, and the standardized change is output. S6: Generate a visual chart using the interactive output module. In this step, the logic for generating risk stratification labels is as follows: The real-time disease risk probability is compared with preset low-risk and high-risk thresholds to determine the numerical range. If the probability value falls into the range less than the low-risk threshold, a low-risk stratification label is generated. If the probability value falls into the range between the low-risk and high-risk thresholds, a medium-risk stratification label is generated. If the probability value falls into the range greater than or equal to the high-risk threshold, a high-risk stratification label is generated.

[0014] The third aspect of this invention provides an application of a multi-gene molecular diagnostic model, which is applied to the risk assessment and dynamic monitoring of type 2 diabetes.

[0015] This invention provides a multi-gene molecular diagnostic model, its construction method, and its application. It has the following beneficial effects: 1. This invention constructs a multimodal data matrix that includes genomes, clinical phenotypes, and environmental exposures, and adopts a two-layer screening strategy combining LASSO regression and random forest. This effectively solves the problems of multicollinearity and noise redundancy in high-dimensional biomedical data, and can lock in the core feature subset with nonlinear classification capabilities, thereby improving the model's generalization ability and risk prediction accuracy in complex clinical scenarios.

[0016] 2. This invention introduces a time adjustment coefficient through a dynamic update module to quantify the standardized differences and weighted cumulative effects of environmental exposure characteristics changing over time. It can adjust the probability of disease in real time according to changes in the subject's lifestyle, providing a quantifiable dynamic tracking method for long-term health management.

[0017] 3. This invention integrates attribution analysis logic based on the SHAP algorithm into the interactive output stage. By calculating the marginal contribution value of features, it intuitively shows whether each biomarker participates in prediction as a positive risk driver or a negative protective factor. Visual attribution not only clarifies the contribution weight of each risk factor, but also enhances the understanding and trust of clinicians and subjects in the diagnostic results. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the module structure of the present invention; Figure 3 This is a block diagram of the overall architecture of the present invention applied to risk assessment of type 2 diabetes.

[0019] The module consists of: 101. Data acquisition module; 102. Preprocessing module; 103. Feature selection module; 104. Model building module; 105. Dynamic update module; and 106. Interactive output module. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 This invention provides a method for constructing a multi-gene molecular diagnostic model, comprising the following steps: S1: Multimodal Data Acquisition and Matrix Construction. Data Acquisition Module 101 acquires the raw dataset of the subject population. The raw dataset contains three independent data dimensions: genomic data, clinical electronic health record data, and environmental exposure data.

[0022] The data acquisition module 101 defines the total sample size in the test group as... Regarding the first 10 subjects, of which The value range is 1 to The data acquisition module 101 constructs the following three data matrices: The first matrix is ​​the genome data matrix. Genomic data matrix This includes single nucleotide polymorphism (SNP) sites identified through genome-wide association studies (GWAS). The total number of SNP sites is set to [value missing]. , No. The first subject A single nucleotide polymorphism characteristic is represented as follows: . The value can be 0, 1, or 2, corresponding to wild-type homozygotes, heterozygotes, and mutant homozygotes, respectively.

[0023] The second matrix is ​​the clinical phenotype data matrix. Clinical phenotype data matrix Physiological and biochemical indicators derived from electronic health records. The total number of clinical phenotypic characteristics was set to [number missing]. , No. The first subject Each clinical phenotypic feature is represented as The third matrix is ​​the environmental exposure data matrix. Environmental exposure data matrix It includes quantified lifestyle and environmental factors. The total number of environmental exposure characteristics is set at [value missing]. , No. The first subject Each environmental exposure characteristic is represented as .

[0024] S2: Multimodal data normalization preprocessing module 102 receives the genome data matrix. Clinical phenotype data matrix and environmental exposure data matrix The preprocessing module 102 performs missing value imputation and Z-score normalization.

[0025] Clinical phenotype data matrix or environmental exposure data matrix For missing feature values ​​present in the data, the preprocessing module 102 uses a regression prediction algorithm to fill them in. (Feature settings are then defined.) If missing values ​​exist, the preprocessing module 102 uses features As the dependent variable, other complete features in the dataset Using these as independent variables, a linear regression model is constructed. Preprocessing module 102 calculates the missing feature values. The formula is as follows: ; in: For the first The first subject was on the padding values ​​on each feature; The regression intercept; The number of complete features used for prediction; For the first The regression coefficients corresponding to each complete feature; For the first The first subject The preprocessing module 102 performs Z-score standardization on all continuous variables after missing value imputation. For any feature... raw observations Eigenvalues ​​after standardization The calculation formula is as follows: ; in: These are the original observations; Features The arithmetic mean in the test group; Features The standard deviation in the subject group. Preprocessing module 102 outputs the standardized fusion feature matrix. Fusion feature matrix The sum of dimensions equal , and sum.

[0026] S3: Dual-layer feature filtering module with 103 pairs of fused feature matrices Perform two-stage dimensionality reduction processing.

[0027] In the first-level screening, the feature screening module 103 utilizes the LASSO regression algorithm. The LASSO regression algorithm, by introducing an L1 regularization penalty term, fuses the feature matrix... The regression coefficients of key features in China and Africa are compressed to zero. Feature selection module 103 retains features with non-zero regression coefficients and generates a candidate feature set. In the second-level screening, the feature screening module 103 selects candidate feature sets. Input a random forest model. The random forest model contains multiple decision trees. Feature selection module 103 calculates the candidate feature set. The Gini importance score for each feature.

[0028] The feature selection module 103 sorts the features in descending order based on the Gini importance score and extracts features whose cumulative importance reaches a preset threshold to form a core feature subset. Core Feature Subset It includes selected genomic features, clinical phenotypic features, and environmental exposure features.

[0029] S4: Multi-gene Molecular Diagnostic Model Construction Module 104 Based on Core Feature Subsets A multi-gene molecular diagnostic model was constructed. The multi-gene molecular diagnostic model employs a logistic regression architecture. Module 104 of the model construction module defines the baseline log-odds ratio. Baseline log odds The calculation formula is as follows: ; in: The baseline logarithmic odds; For model bias terms; A subset of core features; For core feature subset The Middle The training weight coefficients corresponding to each feature; For the first The first subject The standardized value of each feature.

[0030] Multigene molecular diagnostic models output subject data through the sigmoid function. Baseline disease probability .

[0031] S5: Dynamic Update Mechanism Embedded and Executed. The dynamic update module 105 embeds a follow-up data interface into the multi-gene molecular diagnostic model. The follow-up data interface is configured to receive data from subjects over time. Data on changes in lifestyle.

[0032] The dynamic update module 105 defines the dynamic risk score update function. When the follow-up data interface receives new environmental exposure data, the dynamic update module 105 calculates the change in environmental characteristics. Changes in environmental characteristics For time The difference between the environmental characteristic value and the baseline environmental characteristic value.

[0033] Dynamic update module 105 calculation time dynamic logarithmic odds The calculation formula is as follows: ; in: The baseline log odds obtained from step S4; For core feature subset A subset of environmental exposure features; Environmental exposure characteristics The corresponding training weight coefficients; Environmental exposure characteristics In time The change in; This is a time adjustment factor used to adjust the weight of the impact of follow-up data on the overall risk.

[0034] Dynamic update module 105 based on dynamic logarithmic odds Output subject at time Real-time probability of disease .

[0035] S6: Hierarchical assessment and visualization output interactive output module 106 receives real-time disease risk probability. The interactive output module 106 will display the real-time disease risk probability. Compared with the preset low-risk threshold and high-risk threshold Compare them.

[0036] when At that time, the interactive output module 106 generates low-risk hierarchical labels; when At that time, the interactive output module 106 generates medium-risk stratification labels; when At that time, the interactive output module 106 generates high-risk hierarchical labels. The interactive output module 106 is also configured to calculate a subset of core features. The SHAP values ​​of each feature are calculated, and a visualization chart containing the feature contribution is generated.

[0037] See attached document Figure 2 The present invention provides a multi-gene molecular diagnostic model, specifically including: a data acquisition module 101, a preprocessing module 102, a feature screening module 103, a model construction module 104, a dynamic update module 105, and an interactive output module 106.

[0038] The data acquisition module 101 is configured as the input terminal of the model. The data acquisition module 101 includes multi-channel communication interfaces for connecting to a gene sequencing database, a hospital electronic medical record server, and an environmental monitoring terminal. The data acquisition module 101 is configured to receive raw genomic data streams, raw clinical phenotype data streams, and raw environmental exposure data streams in parallel. The data acquisition module 101 internally includes a data parsing unit configured to convert the received data streams into a genomic data matrix. Clinical phenotype data matrix and environmental exposure data matrix The data acquisition module 101 is connected to the preprocessing module 102 via a data bus and is configured to process the genomic data matrix. Clinical phenotype data matrix and environmental exposure data matrix Transmitted to preprocessing module 102.

[0039] The preprocessing module 102 is connected to the data acquisition module 101. The preprocessing module 102 internally includes a regression calculation unit and a standardization calculation unit. The regression calculation unit is configured to execute the aforementioned multiple linear regression algorithm to calculate the clinical phenotype data matrix. and environmental exposure data matrix The missing value imputation amount is determined. The normalization unit is configured to perform Z-score normalization on all continuous feature columns. The preprocessing module 102 is configured to generate a normalized fusion feature matrix without missing values. The preprocessing module 102 is connected to the feature selection module 103 and is configured to process the fused feature matrix. Transmitted to feature filtering module 103.

[0040] Feature selection module 103 is connected to preprocessing module 102. Feature selection module 103 internally includes a first-level selection unit and a second-level selection unit. The first-level selection unit stores LASSO regression algorithm instructions and is configured to calculate the fused feature matrix. The L1 regularized regression coefficients of each feature are calculated, and the candidate feature set is output. The second-level filtering unit stores instructions for the random forest algorithm, configured to calculate candidate feature sets. The Gini importance score of each feature is calculated. The feature selection module 103 is configured to generate a core feature subset based on the Gini importance score. The feature selection module 103 is connected to the model building module 104 and configured to select a subset of core features. Transferred to model building module 104.

[0041] The model building module 104 is connected to the feature selection module 103. The model building module 104 includes a weight storage unit and a logistic regression operation unit. The weight storage unit is configured to store a subset of core features. The training weight coefficients corresponding to each feature The logistic regression computation unit is configured to utilize a subset of core features. and the corresponding training weight coefficients Calculate the baseline log odds of the subjects. and baseline disease probability The model building module 104 is connected to the dynamic update module 105 and configured to update the baseline log odds. and core feature subset The weight parameters are transmitted to the dynamic update module 105.

[0042] The dynamic update module 105 is connected to the model building module 104. The dynamic update module 105 also includes a follow-up data input interface, configured to receive follow-up environmental exposure data of subjects at specific time points. Internally, the dynamic update module 105 includes a difference calculation unit and a dynamic aggregation unit. The difference calculation unit is configured to calculate the standardized change of the follow-up environmental exposure data relative to the baseline data. The dynamic aggregation unit is configured to incorporate a time adjustment factor. Training weight coefficients and baseline log odds Calculation time point dynamic logarithmic odds Furthermore, it generates real-time disease risk probabilities using the Sigmoid function. The dynamic update module 105 is connected to the interactive output module 106 and is configured to update the real-time disease risk probability. Transmitted to interactive output module 106.

[0043] The interactive output module 106 is connected to the dynamic update module 105. The interactive output module 106 internally includes a threshold comparison unit and an attribution analysis unit. The threshold comparison unit stores preset low-risk thresholds. and high risk threshold Configured to display the real-time probability of disease risk The results are compared with the aforementioned thresholds to generate corresponding low, medium, and high risk stratification signals. The attribution analysis unit stores the SHAP value calculation algorithm and is configured to calculate the core feature subset. Each feature corresponds to the real-time probability of disease risk. The contribution values ​​are displayed. The interactive output module 106 is connected to the display device and configured to drive the display device to display a visual chart of the risk stratification signal and the contribution values ​​of the features.

[0044] The multi-gene molecular diagnostic module is configured to operate within an electronic device. The electronic device physically includes a central processing unit (CPU), memory, a network communication interface, and input / output interfaces. The CPU, memory, network communication interface, and input / output interfaces are electrically connected and transmit data signals via a system bus.

[0045] The memory is configured to store computer program instructions and multimodal data required for the operation of the multigene molecular diagnostic module. The memory includes high-speed random access memory and non-volatile memory. The operating system and the multigene molecular diagnostic application program are stored in the memory. The multigene molecular diagnostic application program contains computer-executable code for performing the aforementioned steps S1 to S6.

[0046] The network communication interface is configured to establish communication connections between the electronic device and external data sources. The electronic device connects to the gene sequencing database server, the hospital information management system server, and the follow-up user terminal via the network communication interface.

[0047] The gene sequencing database server is configured to transmit genome-wide association analysis (GWAs) data to electronic devices. The GWAs data is used to construct the aforementioned genome data matrix. The hospital information management system server is configured to transmit electronic medical record data to electronic devices. This electronic medical record data is used to construct the aforementioned clinical phenotype data matrix. The follow-up user terminal is configured to collect lifestyle data entered by the subjects, which is used to construct the aforementioned environmental exposure data matrix. And real-time input for dynamically updating module 105.

[0048] The central processing unit is configured to read the multi-gene molecular diagnostic application from memory and perform the following operations: receive raw data streams from the gene sequencing database server, the hospital information management system server, and follow-up user terminals via a network communication interface; perform regression prediction algorithms on the raw data streams using the internal floating-point arithmetic unit to fill in missing values; and perform Z-score normalization operations using the arithmetic logic unit to generate a fusion feature matrix. .

[0049] The central processing unit (CPU) is further configured to perform matrix operations to achieve L1 regularization in the LASSO regression algorithm and to perform logical judgment operations to construct decision tree node splits in the random forest model. Based on the computation results, the CPU generates the aforementioned real-time disease risk probability. And through input / output interfaces, the real-time probability of disease risk is displayed. Visual charts can be sent to a display device or sent back to the follow-up user's terminal via a network communication interface.

[0050] The electronic device can be a single server, a server cluster, or a cloud computing platform node. When the electronic device is a cloud computing platform node, the LASSO regression operation and random forest model training tasks executed by the central processing unit are distributed to multiple computing nodes for parallel processing to handle the dimensions. The fusion feature matrix is ​​greater than the preset threshold .

[0051] See attached document Figure 1 The preprocessing module 102 first receives the genome data matrix transmitted by the data acquisition module 101. Clinical phenotype data matrix and environmental exposure data matrix The preprocessing module 102 processes the clinical phenotype data matrix. and environmental exposure data matrix The process involves traversing the data to identify feature columns containing null or invalid values. The preprocessing module 102 defines these feature columns as feature vectors to be filled. .

[0052] For each feature vector to be filled The preprocessing module 102 constructs a regression imputation model. The preprocessing module 102 extracts data from the clinical phenotype data matrix. and environmental exposure data matrix Feature columns that do not contain null values ​​are selected from the data to form a predictor matrix. The preprocessing module 102 identifies the feature vector to be filled. The sample row index corresponding to a non-empty element is defined as the observed sample set. Simultaneously identify the feature vector to be filled. The index of the sample row corresponding to the empty value element is defined as the missing sample set. .

[0053] Preprocessing module 102 utilizes the observation sample set The data in the dataset is used to train a multiple linear regression equation. Preprocessing module 102 processes the observed sample set... Corresponding predictor matrix The rows are used as input variables, and the observed sample set is used as input variables. The corresponding feature vector to be filled The elements are used as inputs for the dependent variable, and the regression coefficient vector is solved by the least squares method.

[0054] Preprocessing module 102, based on the obtained regression coefficient vector, processes the missing sample set... The calculation is performed on the null values ​​in the missing sample set. The first in 1 sample, feature vector to be filled Fill value in Calculate according to the following formula: ; in: Indicates the first The predicted values ​​of each sample in the feature column to be filled; This represents the intercept term of the regression model; Represents the predictor matrix The total number of feature columns; Represents the predictor matrix The Middle The regression weight coefficients corresponding to each feature column; Indicates the first Each sample in the predictor matrix The Middle The observed values ​​on each feature column.

[0055] Preprocessing module 102 calculates the fill values Write the feature vector to be filled The corresponding null value positions, up to the clinical phenotype data matrix. and environmental exposure data matrix All feature columns do not contain null values.

[0056] After completing all missing value imputation operations, the preprocessing module 102 processes the genomic data matrix. Clinical phenotype data matrix and environmental exposure data matrix All continuous variable features are standardized using Z-score. Preprocessing module 102 calculates the arithmetic mean and population standard deviation for each feature column.

[0057] For any feature column The preprocessing module 102 will process the feature columns The Middle The original values ​​of each sample Convert to standardized values Standardized values Calculate according to the following formula: ; in: Indicates the first Each sample in the feature column Standardized values; Indicates the first Each sample in the feature column The original value on; Characteristic columns The arithmetic mean of all sample values ​​in the dataset; Indicates the total number of samples; Indicates the first Each sample in the feature column The original values ​​on the denominator represent the characteristic columns. The sample standard deviation.

[0058] Preprocessing module 102 processes the genomic data matrix after Z-score normalization. Clinical phenotype data matrix and environmental exposure data matrix The features are concatenated along the column dimensions to generate a standardized fusion feature matrix. Fusion feature matrix It is configured as the input data source for the subsequent feature filtering module 103.

[0059] See attached document Figure 1 The feature selection module 103 receives the fused feature matrix output by the preprocessing module 102. The feature selection module 103 first performs a first-level feature screening based on the LASSO regression algorithm. The feature selection module 103 constructs a linear regression objective function with an L1 regularization term. The feature selection module 103 then solves for the regression coefficient vector that minimizes the objective function using either coordinate descent or least angle regression. The objective function used by feature selection module 103 Constructed based on the following formula: ; in: Indicates the total number of subject samples; Indicates the first The disease status label for each subject, with a value of 0 or 1; This represents the intercept term of the regression model; Represents the fusion feature matrix The total number of features; Indicates the first The first subject was on the Standardized values ​​for each feature; Indicates the first The regression coefficients corresponding to each feature; This represents the regularization parameter, used to control the sparsity of the model; Regression coefficients The absolute value of.

[0060] The feature selection module 103 uses the K-fold cross-validation method to determine the optimal value of the regularization parameter. The feature selection module 103 iterates through the obtained regression coefficient vector. Feature selection module 103 identifies regression coefficients. Features that are not equal to zero are identified, and these identified features are extracted into a candidate feature set. .

[0061] After completing the first-level feature screening, the feature selection module 103 checks the candidate feature set. Perform second-level feature selection based on the random forest algorithm. Feature selection module 103 is constructed and includes... A random forest classifier with decision trees. The feature selection module 103 selects candidate feature sets. The disease state label is used as the input variable and the target variable to train the random forest classifier.

[0062] The feature selection module 103 calculates the candidate feature set during the training process. The Gini importance score for each feature in the random forest classifier. Feature filtering module 103 computing nodes impurity of the gin The calculation formula is as follows: ; in: This represents the total number of decision trees in the random forest classifier; Indicates the first All features used in a decision tree The set of nodes to be split; Representation of features At the node The decrease in impurity of the ginni caused by the location.

[0063] Feature selection module 103 scores based on global importance. For candidate feature sets The features are sorted in descending order. Feature filtering module 103 extracts the features with the highest sorted position. The core feature subset is formed by selecting features that have accumulated importance scores that reach a preset percentage threshold. Core Feature Subset It is configured to be transferred to the model building module 104 for building a multi-gene molecular diagnostic model.

[0064] See attached document Figure 1 The dynamic update module 105 establishes a communication connection with the follow-up data interface. The dynamic update module 105 receives data from the subject at the current time point through the follow-up data interface. The resulting follow-up observation dataset contains numerical values ​​of the environmental exposure characteristics of the subjects at different time points.

[0065] The dynamic update module 105 updates the core feature subset. The system identifies features belonging to environmental exposure types and defines the identified features as a variable environmental feature set. The dynamic update module 105 retrieves the baseline statistical parameters stored in the preprocessing module 102. These baseline statistical parameters include a variable environmental feature set. The arithmetic mean and standard deviation of each feature in the baseline subject population.

[0066] For variable environment feature sets Each feature in The dynamic update module 105 calculates the subjects' time points. Standardized variation relative to baseline Standardized variation Calculate according to the following formula: ; in: Representation of features At the point of time Standardized variation; Indicates the subject at time point Regarding features The measured values; Indicates the subject's characteristics at baseline. The original measured values; Representation of features Standard deviation in the baseline subject group.

[0067] The dynamic update module 105 retrieves the baseline log odds calculated by the model building module 104. and features Corresponding training weight coefficients The dynamic update module 105, combined with a time adjustment function, calculates the subject's time points. dynamic logarithmic odds Dynamic logarithmic odds Calculate according to the following formula: ; in: Indicates a point in time The dynamic logarithmic odds; This represents the subject's static risk score at baseline; This represents the time decay factor, used to adjust the weight of the impact of follow-up data on the overall risk; Represents a variable set of environmental features; This represents the regression weight coefficient of feature k in a multi-gene molecular diagnostic model; Representation of features At the point of time The standardized variation.

[0068] In the above formula, the time decay factor The configuration is set to either decay exponentially over time or remain constant. When configured for exponential decay, ,in This is a preset decay constant used to assign higher weights to recent data compared to older data.

[0069] The dynamic update module 105 will calculate the dynamic logarithmic odds. Mapped to probability values, outputting the subject's performance at each time point. Real-time probability of disease Real-time probability of disease Calculate according to the following formula: ; in: This represents the predicted probability of a subject developing the disease at time point t. is the base of the natural logarithm; Let be the dynamic logarithmic probability at time point t.

[0070] The dynamic update module 105 will calculate the generated real-time disease risk probability. The data is transmitted to the interactive output module 106. The interactive output module 106 is configured to output data based on the real-time disease risk probability. Generate a risk trajectory curve that changes over time and display the risk trajectory curve through a display device.

[0071] See attached document Figure 3 This invention provides a specific implementation method for a multi-gene molecular diagnostic model, which is configured to apply the aforementioned construction method to the risk assessment and dynamic monitoring of type 2 diabetes mellitus (T2MD), specifically including: The data acquisition module 101 acquires type 2 diabetes-related multimodal data from the subjects. The data acquisition module 101 constructs a genomic data matrix. genomic data matrix This includes rs7903146 ​​locus data for the transcription factor 7 analog 2 gene (TCF7L2) and rs2237892 locus data for the potassium voltage-gated channel subfamily Q member 1 gene (KCNQ1). Data acquisition module 101 constructs a clinical phenotype data matrix. Clinical phenotype data matrix This includes fasting blood glucose levels, glycated hemoglobin levels, and body mass index. Data acquisition module 101 constructs an environmental exposure data matrix. Environmental exposure data matrix It includes weekly aerobic exercise duration and daily dietary calorie intake.

[0072] The preprocessing module 102 receives the aforementioned type 2 diabetes-related multimodal data. The preprocessing module 102 identifies missing data in the glycated hemoglobin (HbA1c) values ​​and uses fasting blood glucose and body mass index (BMI) as predictor variables to calculate the imputed HbA1c values ​​using the aforementioned multiple linear regression algorithm. The preprocessing module 102 performs the aforementioned Z-score standardization on the fasting blood glucose, HbA1c, BMI, weekly aerobic exercise duration, and daily dietary calorie intake to generate a standardized type 2 diabetes fusion feature matrix.

[0073] The feature selection module 103 receives the standardized type 2 diabetes fusion feature matrix. It performs LASSO regression to select candidate features with non-zero regression coefficients. The module then inputs these candidate features into a random forest model to calculate the Gini importance of each feature for type 2 diabetes prevalence. Based on the Gini importance ranking, the module determines a core feature subset. Core Feature Subset Specifically, it includes the rs7903146 ​​locus variant status, body mass index, and weekly aerobic exercise duration.

[0074] Model building module 104 is based on a subset of core features Construct a logistic regression model. Model construction module 104 stores a subset of core features. The training weights corresponding to each feature are calculated. Model building module 104 calculates the baseline log-odds ratio of the subjects. Baseline log odds The result is obtained by adding the weighted values ​​of the rs7903146 ​​locus variant status, the body mass index, and the weekly aerobic exercise duration. Model building module 104 outputs the baseline prevalence of type 2 diabetes for the subjects.

[0075] The dynamic update module 105 is configured to receive exercise intervention data from subjects during the follow-up phase. When the follow-up data interface receives data from subjects at specific time points... When updating the weekly aerobic exercise duration, the dynamic update module 105 calculates the standardized change of the updated weekly aerobic exercise duration relative to the baseline weekly aerobic exercise duration. The dynamic update module 105 calculates the time point according to the following formula. dynamic logarithmic odds : ; in: Indicates a point in time The dynamic logarithmic odds; Indicates the baseline logarithmic odds; This represents the weighting coefficient of weekly aerobic exercise duration in the logistic regression model; This indicates the duration of aerobic exercise per week at specific times. The standardized variation.

[0076] The dynamic update module 105 will dynamically update the logarithmic odds. The probability of developing type 2 diabetes is converted to a real-time probability using the Sigmoid function. The interactive output module 106 receives this real-time probability. The interactive output module 106 then uses the SHAP algorithm to calculate the contribution of the rs7903146 ​​locus variant state, body mass index, and weekly aerobic exercise duration to the real-time probability of developing type 2 diabetes. The interactive output module 106 generates a risk attribution map on the display device, which visually demonstrates the contribution of the increase in weekly aerobic exercise duration to reducing the probability of developing type 2 diabetes.

[0077] See attached document Figure 3 The interactive output module 106 receives the real-time disease risk probability output by the dynamic update module 105. and the multi-gene molecular diagnostic model functions stored in the model building module 104. The interactive output module 106 is configured to use the SHAP algorithm to calculate a subset of core features. Each feature corresponds to the real-time probability of disease risk. The independent contribution value.

[0078] Interactive output module 106 defines the set of features of the entire set. The set of features of the whole set Includes a subset of core features All feature indices in the set. For the entire feature set. Any target feature in The interactive output module 106 does not contain target features. Feature subset Feature subset It is the set of features of the entire set. Exclude target features Any subset thereafter.

[0079] Interactive output module 106 calculates target features SHAP value For the first Sample data of each subject SHAP value Calculate according to the following formula: ; in: Representing target features For sample data The marginal contribution value of the model's prediction results; Represents the set of features of the entire set; Indicates that the target feature is not included. A subset of features; This represents the total number of features in the feature set of the entire set; Indicates the number of features in the feature subset; This indicates that the model uses only a subset of features. The conditional expected value of the predicted value when the features are in the middle; The model indicates that it uses a subset of features. With target features The conditional expected value of the union of the given set.

[0080] The interactive output module 106 calculates the SHAP value. Applying multi-gene molecular diagnostic models to sample data Predicted output The decomposition is the sum of the baseline value and the characteristic contribution value. The decomposition logic is performed according to the following formula: ; in: The model represents the sample data. The predicted output value corresponds to the aforementioned logarithmic odds; This represents the average expected value of the multi-gene molecular diagnostic model on the training dataset; This represents the sum of the SHAP values ​​of all input features.

[0081] The interactive output module 106 calculates the SHAP value. The sign and absolute value of the SHAP value are used to generate a list of feature attribution data. At that time, the interactive output module 106 will output the target features Marked as a positive risk driver; when the SHAP value At that time, the interactive output module 106 will output the target features Marked as a negative risk protection factor.

[0082] Interactive output module 106 is based on real-time disease risk probability Perform a risk stratification assessment. The interactive output module 106 stores preset low-risk thresholds. and high risk threshold ,in .

[0083] Interactive output module 106 will display the real-time disease risk probability. With low risk threshold and high-risk threshold Perform numerical comparison.

[0084] Real-time probability of disease risk Less than the low risk threshold At that time, the interactive output module 106 generates a low-risk level signal and maps the low-risk level signal to green identification data.

[0085] Real-time probability of disease risk Greater than or equal to the low-risk threshold And less than the high-risk threshold At that time, the interactive output module 106 generates a medium-risk level signal and maps the medium-risk level signal to yellow marker data.

[0086] Real-time probability of disease risk Greater than or equal to the high-risk threshold At that time, the interactive output module 106 generates a high-risk level signal and maps the high-risk level signal to red marker data.

[0087] The interactive output module 106 is further configured to generate visualized waterfall chart data. The interactive output module 106 will use the baseline value... Set as the starting baseline for the waterfall plot. The interactive output module 106 outputs the SHAP value of each feature. This is mapped to the height of the bars in the waterfall chart. The interactive output module 106 sorts the bars in descending order according to the absolute value of the SHAP value. The interactive output module 106 configures the bars of positive risk drivers as the first color and the bars of negative risk protection factors as the second color.

[0088] The interactive output module 106 transmits the generated feature attribution data list, risk stratification signal, and visual waterfall plot data to the display device.

[0089] The display device is configured to render and display an interactive interface that includes the subject's current risk level, major risk drivers, and protective factors.

Claims

1. A multi-gene molecular diagnostic model, characterized in that, include: The data acquisition module is configured to acquire the raw dataset of the subject population and construct a genomic data matrix, a clinical phenotype data matrix, and an environmental exposure data matrix based on the raw dataset. The preprocessing module, connected to the data acquisition module, is configured to perform regression imputation on missing feature values ​​in the data, perform standardization on continuous variables, and output a fused feature matrix. The feature filtering module is connected to the preprocessing module and is configured to perform a two-layer filtering on the fused feature matrix using the LASSO regression algorithm and the random forest model to output a core feature subset. The model building module is connected to the feature selection module and is configured to establish a logistic regression operation architecture based on the core feature subset to calculate the baseline log odds of the subjects. The dynamic update module is connected to the model building module and is configured to combine follow-up environmental exposure data, time adjustment coefficient and the baseline log odds to output the real-time disease risk probability. The interactive output module, connected to the dynamic update module, is configured to output the risk stratification label and feature contribution corresponding to the real-time disease risk probability.

2. The multi-gene molecular diagnostic model according to claim 1, characterized in that, The preprocessing module is configured to perform regression imputation according to the following logic: For features with missing values, a linear regression model is constructed using other complete features in the original dataset as independent variables. The missing feature values ​​are calculated, and the intercept term of the linear regression model is summed with the weighted terms of each complete feature involved in the prediction. Each weighted term is defined as the product of the regression coefficient of the complete feature and its corresponding observation value.

3. The multi-gene molecular diagnostic model according to claim 1, characterized in that, The feature filtering module includes a first-level filtering unit and a second-level filtering unit; The first-level screening unit is configured to perform a compression operation on the regression coefficients by introducing an L1 regularization penalty term, retaining features with non-zero coefficients to generate a candidate feature set. The second-level screening unit is configured to quantify the Gini importance score of each feature based on the decrease in Gini impurity during the decision tree node splitting process, and to extract the core feature subset based on the ranking result of the Gini importance score.

4. The multi-gene molecular diagnostic model according to claim 1, characterized in that, The dynamic update module is configured to calculate the real-time disease risk probability according to the following logic: First, the training weight coefficients of each feature in the environmental exposure feature subset are calculated and their standardized changes are weighted and accumulated. Then, the weighted accumulated value is multiplied by the time adjustment coefficient. Next, the obtained product is added to the baseline logarithmic odds to obtain the dynamic logarithmic odds. Finally, through the nonlinear transformation logic of the sigmoid function, the dynamic logarithmic odds are mapped to a probability interval of zero to one to obtain the real-time disease risk probability.

5. A multi-gene molecular diagnostic model according to claim 1, characterized in that, The interactive output module is configured to calculate feature contribution according to the following logic: Based on the definition of the SHAP algorithm, a feature subset combination that does not contain the target feature is constructed. The marginal contribution value of the addition of the target feature to the model prediction output is evaluated. Based on the positive or negative polarity and numerical magnitude of the marginal contribution value, the target feature is determined to be a positive risk driving factor or a negative risk protection factor.

6. The multi-gene molecular diagnostic model according to claim 1, characterized in that, The interactive output module is configured to generate risk stratification labels according to the following logic, and to determine the numerical range of the real-time disease risk probability with the preset low-risk threshold and high-risk threshold respectively. If the probability value falls within the range of less than the low-risk threshold, a low-risk stratified label is generated; If the probability value falls within the range between the low-risk threshold and the high-risk threshold, a medium-risk stratified label is generated. If the probability value falls within the range that is greater than or equal to the high-risk threshold, a high-risk stratification label is generated.

7. A method for constructing a multi-gene molecular diagnostic model, characterized in that, The multi-gene molecular diagnostic model applied to any one of claims 1 to 6 includes the following steps: S1: Construct the genomic data matrix, the clinical phenotype data matrix, and the environmental exposure data matrix using the data acquisition module; S2: Output the fused feature matrix using the preprocessing module; S3: The feature filtering module is used to perform a two-layer filtering on the fused feature matrix to generate the core feature subset; S4: Calculate the baseline log odds using the model building module; S5: Receive follow-up environmental exposure data using the dynamic update module, calculate the standardized change in environmental characteristics, and output the real-time disease risk probability. S6: Use the interactive output module to generate a visual chart.

8. The method for constructing a multi-gene molecular diagnostic model according to claim 7, characterized in that, In step S5, the calculation process for the standardized variation specifically includes: The baseline measurements of the subjects during the model building phase were established as the reference benchmark. The real-time measurements obtained at the current follow-up time were compared with the reference benchmark to obtain the absolute difference value reflecting the longitudinal change of the individual. The statistical dispersion index of the subject group on which the model was based, namely the baseline standard deviation, is retrieved. The normalization operation is performed by calculating the ratio of the absolute difference value to the baseline standard deviation to eliminate the influence of dimensions, and the standardized change is output.

9. The method for constructing a multi-gene molecular diagnostic model according to claim 7, characterized in that, In step S2, the standardization process using Z-score standardization specifically includes: For each continuous variable, calculate the arithmetic mean and standard deviation of the continuous variable in the overall sample space; For each sample point, perform a centering operation on the original feature value to determine the deviation of the original value from the arithmetic mean. A scaling operation is performed to transform the data distribution into a standard normal distribution with a mean of zero and a variance of one by calculating the ratio of the deviation distance to the standard deviation.

10. An application of a multi-gene molecular diagnostic model, characterized in that, The application of a multi-gene molecular diagnostic model according to any one of claims 1 to 6 in the risk assessment and dynamic monitoring of type 2 diabetes.

Images