A high-dimensional biomarker screening method and system for processing multi-source heterogeneous medical data

Abstract

This invention relates to the field of medical data mining and statistical inference technology, specifically to a method and system for screening high-dimensional biomarkers from multi-source heterogeneous medical data. The method includes: constructing a heterogeneous distributed high-dimensional dataset at various distributed sites; constructing a regression model containing global parameters and heterogeneity parameters, with each distributed site computing an initial consistent estimator and a locally effective score function in parallel; calculating density ratio weights based on the locally effective score function to construct a globally effective communication score function; introducing an adaptive discontinuous penalty term based on the initial consistent estimator and the globally effective communication score function; updating the global parameters using local quadratic approximation and coordinate descent under fixed heterogeneity parameters; outputting a globally sparse estimator and screening key biomarkers. This invention enables the simultaneous and accurate screening of high-dimensional biomarkers and accurate prediction of disease states without sharing the full original underlying data and with precise calibration of multi-center distribution heterogeneity.

Description

Technical Field

[0001] This invention relates to the field of medical data mining and statistical inference technology, specifically to a method and system for screening high-dimensional biomarkers in multi-source heterogeneous medical data. Background Technology

[0002] With the accelerating pace of medical informatization, multi-center collaborative research models on a global scale have become a crucial engine for driving biomedical discovery. In clinical research, many disease states exhibit binary response variables, making logistic regression models the standard statistical tool for analyzing such medical data. However, genomic and other data collected in modern clinical research often possess significant high-dimensionality. In this context, the sample size of a single medical institution is insufficient to support stable model estimations, making the aggregation of large-scale multi-center data an inevitable trend. However, due to strict privacy regulations, original patient data from different medical institutions... Direct merging is not feasible, requiring distributed modeling without sharing underlying data. Existing technologies have primarily developed distributed processing methods based on "weighted average" or "communication-efficient substitution likelihood." However, in real-world high-dimensional, multi-center medical scenarios, existing technologies face the following insurmountable challenges: Due to differences in geography, equipment, and clinical practice, data distribution across different medical sites often exhibits significant heterogeneity. Most existing distributed likelihood methods are based on the assumption of independent and identically distributed data. Ignoring heterogeneity parameters reflecting baseline risk differences and directly transmitting or aggregating first-order gradients within the network can lead to severe estimation bias. As the number of collaborating sites increases, their asymptotic performance deteriorates significantly. To address heterogeneity, some recent techniques have introduced site-specific redundant parameters and constructed effective communication score functions by calculating density-to-skew weights. However, this framework is strictly limited to low-dimensional inference. In high-dimensional scenarios with tens of thousands of gene features, the equation does not impose sparsity constraints on the global parameter β, resulting in infinitely many solutions and making it impossible to accurately identify key variables. In centralized high-dimensional scenarios, discontinuous penalty methods such as adaptive LASSO are widely used. However, their penalty terms are highly dependent on a high-quality initial consistent estimator. In distributed high-dimensional scenarios, ordinary least squares estimation does not exist; if standard LASSO is forcibly used as the initial estimate, it is very easy to compress the weight of the weak real signal to zero, causing the penalty weight to tend to infinity, thereby losing key biomarkers. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for screening high-dimensional biomarkers in processing multi-source heterogeneous medical data in order to solve the above-mentioned technical problems. This method and system can simultaneously achieve accurate screening of high-dimensional biomarkers and accurate prediction of disease status without sharing all the original underlying data and accurately calibrating the heterogeneity of multi-center distribution.

[0004] The objective of this invention can be achieved through the following technical solutions: A method for screening high-dimensional biomarkers in processing multi-source heterogeneous medical data includes: S1. Construct heterogeneous distributed high-dimensional datasets at each distributed site; S2. Construct a regression model containing global parameters and heterogeneous parameters based on the heterogeneous distributed high-dimensional dataset. Each distributed site calculates the initial consistent estimator and the local effective score function in parallel, and sends the local effective score to the central site. S3. Calculate the density ratio weight at the central station based on the local effective score function of each distributed station, and construct the global effective communication score function based on the density ratio weight. S4. Based on the initial consensus estimator and the global communication effective score function, an adaptive discontinuous penalty term is introduced to construct the penalty communication effective estimation equation; the global parameters are iteratively updated through local quadratic approximation and coordinate descent method, and iterative optimization is performed until the convergence threshold is met. S5. Output the optimal global sparse estimator and select key biomarkers based on the global sparse estimator. Predict the target disease status using key biomarkers.

[0005] Furthermore, step S1 includes: Includes K distributed sites, and the dataset of the k-th distributed site. Including binary response variables and global covariate characteristics ,in The distributed site-specific covariate features are d, which represents the number of genes retained according to variance sorting, and n represents the number of samples. The global covariate features are high-dimensional data features shared across sites. The global covariate features are processed locally at the distributed site using Z-score standardization.

[0006] Furthermore, step S2 includes: The regression model is a logistic regression model, with high-dimensional sparse global parameters and heterogeneity parameters representing the data distribution differences at the k-th distributed site. Initial consistent estimators are calculated for each distributed site using distributed Ridge regression. The local effective score function of the k-th distributed site is calculated using orthogonal projection techniques. : Where β is the global regression coefficient vector corresponding to d genes. To capture parameters of baseline risk and pathological stage heterogeneity across distributed sites, For the log-likelihood function of the distributed sites, and The information matrix corresponding to each distributed site is divided into blocks, and each distributed site will receive a local effective score. A single transmission to the central station.

[0007] Furthermore, step S3 specifically includes: S31. At the central station, calculate the density ratio weight of remote station k relative to the central station. : S32. Constructing an effective global communication score function using density ratio weights. : in, This represents the valid score item for the i-th sample at the central site. This indicates the local effective score.

[0008] Furthermore, step S4 includes, S41. By determining the penalty parameters through the extended Bayesian information criterion, a non-smooth adaptive penalty term is introduced to construct an effective estimation equation for penalized communication. ; S42. Set the iteration number t=0. In the current t-th iteration: With the global parameter β(t) fixed, the heterogeneity parameter is calculated by maximizing the local log-likelihood of the central station; By utilizing a local quadratic approximation to handle the penalty term, one-dimensional coordinate descent is used to update each component of the global parameter β(t). ; S43. Determine whether the global parameters have reached the convergence threshold. If the convergence threshold has not been reached, update the iteration count t=t+1, return to step S42 and re-execute until the convergence threshold is reached.

[0009] Furthermore, step S41 specifically includes: Construct an effective estimation equation for penalized communication: in, Let be the first derivative of the penalty function; the penalty function employs an adaptive LASSO strategy, and its derivative form is defined as follows: Adaptive weights ; Let ν be the initial consistent estimator, and ν be a normal constant that controls the sparsity penalty.

[0010] Furthermore, step S42 specifically includes: Initialization is performed at the central station. And start iterative solution: S421. Calculate heterogeneity parameters by maximizing the local log-likelihood of the central site. : S422, For each component of the global parameter β Calculate the local quadratic approximation coefficients of the penalty term. ,in The constant is a small constant; simultaneously, the approximate values ​​of the diagonal elements of the weighted Hessian matrix for the central stations are calculated. ; S423. Using local quadratic approximation for global parameters Perform coordinate descent update: .

[0011] Furthermore, step S43 specifically includes: The convergence of the algorithm is determined by calculating the Euclidean norm of the difference between the global parameters in two adjacent iterations. If the above equation is not satisfied, the iteration count is updated to t=t+1, and step S42 is executed again; if the above equation is satisfied, the iteration stops, and the current iteration count is set to... As a global sparse estimator Output.

[0012] Furthermore, step S5 specifically includes: Output the optimal global sparse estimator obtained in step S4. Gene feature variables with non-zero coefficients are extracted as key biomarkers; High-dimensional gene expression data and pathological stage-specific covariates of the target patient are extracted, substituted into the regression model, and the clinical disease status prediction results of the patient are output.

[0013] A high-dimensional biomarker screening system for processing multi-source heterogeneous medical data includes: Memory and processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the method as claimed in any one of claims 1-9; the processor is configured to execute the following modules: The data acquisition and preprocessing module is used to collect patient observation data from various distributed sites and construct heterogeneous distributed high-dimensional datasets at each distributed site. The distributed local computing module is used to construct a regression model containing global parameters and heterogeneous parameters based on a heterogeneous distributed high-dimensional dataset. Each distributed site computes the initial consistent estimator and the local effective score function in parallel and sends the local effective score to the central site. The penalty equation construction module is used to calculate the density ratio weight at the central station based on the local effective score function of each distributed station, construct the global communication effective score function based on the density ratio weight, and introduce an adaptive discontinuous penalty term on the basis of the initial consistent estimator and the global communication effective score function to construct the penalty communication effective estimation equation. The iterative update module is used to update the global parameters through local quadratic approximation and coordinate descent, iteratively optimizing until the convergence threshold is met, and outputting the optimal global sparse estimate. The biomarker screening and prediction output module is used to screen key biomarkers based on global sparse estimates and predict the target disease status through key biomarkers.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention calculates an initial consistent estimator and a locally effective score function at each distributed site, and sends the locally effective score function to the central site. Based on the locally effective score function, a global communication effective score function is constructed, so that each distributed site can participate in model construction without transmitting the original patient observation data, thereby reducing the amount of data transmission and communication overhead.

[0015] 2. This invention calculates density ratio weights based on local effective score functions at the central site and uses these density ratio weights to calibrate the data at each distributed site, thereby enabling the global effective score function to reflect the heterogeneous distribution differences between different distributed sites and thus improving the accuracy of global parameter estimation.

[0016] 3. This invention introduces an adaptive discontinuous penalty term on the basis of the global communication effective score function, and constructs a penalty communication effective estimation equation. Under the condition of fixed initial values ​​of heterogeneous parameters, the global parameters are iteratively updated through local quadratic approximation and coordinate descent method, so that the global parameters exhibit sparsity, thereby reducing the impact of high-dimensional variables on model calculation, realizing consistent selection of key pathogenic genes that are truly related to the disease state, and improving the estimation efficiency of model parameters and the robustness of prediction of unknown clinical data.

[0017] 4. This invention iteratively updates global parameters and outputs a global sparse estimate. Key biomarkers are then selected based on the global sparse estimate, making the set of variables obtained from high-dimensional data more representative. Attached Figure Description

[0018] Figure 1 This is a flowchart of a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data according to the present invention; Figure 2 This is a flowchart of the core algorithm for a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data according to the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0020] like Figure 1 and Figure 2 The method shown is a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data, comprising: S1. Collect patient observation data at each distributed site and perform standardized preprocessing on the patient observation data locally at each distributed site to construct a heterogeneous distributed high-dimensional dataset. This is used to complete data preprocessing without sharing the original data, unify the data scale of different distributed sites, and preserve the distribution differences between sites.

[0021] S2. Construct a regression model containing global parameters and heterogeneous parameters based on the heterogeneous distributed high-dimensional dataset. Each distributed site computes the initial consistent estimator and the local effective score function in parallel, and sends the local effective scores to the central site. Extract key information from each site in a distributed manner, and transmit only local statistics instead of the original data, thereby reducing the risk of data leakage. At the same time, the initial consistent estimator provides an initial foundation for subsequent high-dimensional parameter optimization.

[0022] S3. At the central site, the density ratio weight is calculated based on the local effective score function of each distributed site. A global effective score function for communication is constructed based on the density ratio weight. This is used to eliminate the impact of data distribution heterogeneity between different distributed sites on model estimation, so that multi-center data can be jointly modeled under a unified statistical framework, thereby improving the accuracy of global parameter estimation.

[0023] S4. Based on the initial consistent estimator and the global communication effective score function, an adaptive discontinuous penalty term is introduced to construct a penalized communication effective estimation equation. Global parameters are updated through local quadratic approximation and coordinate descent, iteratively optimizing until the convergence threshold is met. Specifically, the adaptive discontinuous penalty term employs an adaptive LASSO strategy, applying differentiated penalties to different parameter components to achieve high-dimensional variable selection. In each iteration, the global parameters are first fixed to calculate heterogeneous parameters, and then the global parameters are updated through dimension-wise coordinate descent based on the local quadratic approximation. This achieves stable parameter estimation and variable selection in high-dimensional space, effectively eliminating noisy variables by introducing sparsity constraints, and improving model solution efficiency and reducing computational complexity through iterative optimization.

[0024] S5 outputs the optimal global sparse estimator and selects key biomarkers based on the global sparse estimator. These key biomarkers are then used to predict the target disease state. This process is used to select key variables related to the target disease state from high-dimensional candidate features, providing a basis for disease analysis and clinical decision-making.

[0025] Specifically, step S1 includes: Clinical data and transcriptome sequencing (RNA-Seq) data of patients from K=10 different tissue-derived sites were collected from the Cancer Genome Atlas (TCGA) database, with a sample size of n=647. The patients' estrogen receptor status (ERStatus, Y ∈ {0,1}, 1 indicating positive, 0 indicating negative) was used as a binary response variable. Gene expression features were extracted as global covariate features. The top d=1000 genes were retained based on variance sorting; tumor pathological stage was extracted as a distributed site-specific covariate feature. The global covariate features are high-dimensional data features shared across sites. These features are standardized locally at each distributed site using Z-scores to eliminate the influence of dimensions and meet the numerical stability requirements of penalized regression. The final dataset for the k-th site is denoted as... ,in Let k be the sample size at site k.

[0026] Specifically, step S2 includes: The regression model is a logistic regression model, expressed as: Where β is the global regression coefficient vector corresponding to d genes. To capture parameters of baseline risk and pathological stage heterogeneity across distributed sites, the global parameters are high-dimensional and sparse, while the heterogeneity parameters are the data distribution differences at the k-th distributed site. To prevent the weights of weak signals from being set to infinity due to direct use of standard LASSO, an initial consensus estimator is calculated for each distributed site using distributed Ridge regression. The local effective score function of the k-th distributed site is calculated using orthogonal projection techniques. : in, For the log-likelihood function of the distributed sites, and Divide the Hessian matrix corresponding to each distributed site into blocks, and each distributed site will receive a locally effective score. The data is transmitted to a central site in a single transmission, with the central site being the master node with the largest sample size. This process does not transmit raw gene expression or clinical data, thus protecting patient privacy.

[0027] Specifically, step S3 includes: S31. At the central station, calculate the density ratio weight of remote station k relative to the central station. : in, This represents the joint probability density function corresponding to the k-th distributed site; This represents the joint probability density function corresponding to the central station; Let be the heterogeneity parameter of the k-th distributed site; The heterogeneity parameter for the central site.

[0028] S32. Constructing an effective global communication score function using density ratio weights. : in, This represents the valid score item for the i-th sample at the central site. This indicates the local effective score. This represents the set of heterogeneity parameters for all distributed sites. This represents the effective score of the i-th sample at the central site. By constructing density ratio weights, data from different distributed sites are mapped to a unified reference distribution, reducing the impact of distribution differences on model estimation. A unified global communication effective score function is formed by weighting and integrating the local effective score functions of each distributed site, achieving information fusion from multiple data sources. The global communication effective score function depends only on the local effective score functions and density ratio weights, without transmitting the original data, thus meeting the requirements for distributed privacy protection.

[0029] Specifically, step S4 includes, S41. By determining the penalty parameter through the extended Bayesian information criterion, the optimal penalty parameter λ=0.0007 is determined. A non-smooth adaptive penalty term (ALASSO) is introduced to construct the penalty communication effective estimation equation. Specifically, the effective estimation equation is: in, Let be the first derivative of the penalty function; the penalty function employs an adaptive LASSO strategy, and its derivative form is defined as follows: Adaptive weights ; Let ν be the initial consistent estimator, and ν be the normal constant controlling the sparsity penalty. By introducing an adaptive discontinuous penalty term, differential constraints are imposed on different features, causing the coefficients of unimportant variables to tend to zero.

[0030] S42. Set the iteration number t=0. In the current t-th iteration: With the global parameter β(t) fixed, the heterogeneity parameter is calculated by maximizing the local log-likelihood of the central site. By utilizing a local quadratic approximation to handle the penalty term, one-dimensional coordinate descent is used to update each component of the global parameter β(t). Specifically, it includes: Initialization is performed at the central station. And start iterative solution: S421. Update heterogeneity parameters by maximizing the local log-likelihood of the central site. : S422, For each component of the global parameter β Calculate the local quadratic approximation coefficients of the penalty term. ,in The constant is a small constant; simultaneously, the approximate values ​​of the diagonal elements of the weighted Hessian matrix for the central stations are calculated. ; S423. Using local quadratic approximation for global parameters Perform coordinate descent update: .

[0031] By using local quadratic approximation and coordinate descent, the complex high-dimensional optimization problem is transformed into a dimension-by-dimensional update problem, thus reducing computational complexity.

[0032] S43. Determine if the global parameters have reached the convergence threshold; if not, update the iteration count t = t + 1, return to step S42 and re-execute until the convergence threshold is reached; specifically including: The convergence of the algorithm is determined by calculating the Euclidean norm of the difference between the global parameters in two adjacent iterations. If the above equation is not satisfied, the iteration count is updated to t=t+1, and step S42 is executed again; if the above equation is satisfied, the iteration stops, and the current iteration count is set to... As a global sparse estimator Output. Through iterative updates and convergence determination mechanisms, the global parameters and heterogeneous parameters gradually approach a stable solution.

[0033] Specifically, step S5 includes: Output the optimal global sparse estimator obtained in step S4. Gene feature variables with non-zero coefficients are extracted as key biomarkers; High-dimensional gene expression data and pathological stage-specific covariates of the target patient are extracted, substituted into the regression model, and the clinical disease status prediction results of the patient are output.

[0034] To verify the effectiveness of the method in this embodiment, simulation experiments were designed with different sample sizes n and different dimensions d. Mean squared error (MSE) was used as the parameter estimation accuracy index, and true positive rate (TPR), false positive rate (FPR), the number of correctly identified zero coefficients C, and the number of incorrectly identified zero coefficients I were used as variable selection accuracy indexes. The true global parameter vector was set as follows: The first five covariates had significant effects, while the rest were noise variables. The model incorporates site-specific interference parameters. Simulate distribution heterogeneity.

[0035] Table 1. Comparison of variable selection performance for continuous high-dimensional features in distributed heterogeneous environments.

[0036] Table 2. Comparison of variable selection performance for hybrid high-dimensional features in distributed heterogeneous environments.

[0037] The results in Tables 1 and 2 show that the adaptive discontinuous penalty method ALASSO used in this embodiment of the invention exhibits optimal statistical performance in a distributed heterogeneous environment. Specifically: First, in terms of parameter estimation accuracy, whether in continuous covariate scenarios or complex mixed covariate scenarios, the mean squared error (MSE) of the method in this invention is significantly lower than that of the traditional LASSO and SCAD methods. Moreover, as the sample size or number of sites increases, its performance highly approaches that of centralized Oracle benchmark estimators without privacy barriers. Second, in terms of high-dimensional variable selection accuracy, the method in this invention maintains an extremely high true positive rate (TPR) close to 1 and strictly controls the false positive rate (FPR) to a level close to 0, achieving accurate removal of high-dimensional noisy variables. This verifies the advantages of the "adaptive density ratio weighting" combined with the "local quadratic approximation" strategy in this invention in overcoming the heterogeneity of multi-center data and achieving consistent selection of high-dimensional sparse features.

[0038] Furthermore, the above algorithm was applied to a real-world multicenter invasive breast cancer dataset from a cancer genome map, aiming to accurately screen core pathogenic genes highly correlated with estrogen receptor status from a massive pool of candidate genes. The results of key gene variable selection using the method of this invention compared to traditional methods are shown in Table 3 below: Table 3: Comparison of key gene variable selection methods in distributed heterogeneous breast cancer data

[0039] As shown in Table 3, compared with traditional LASSO, which is prone to selecting noise variables that have not been biologically validated, and SCAD, which has excessive penalty and causes the loss of key signals, the method proposed in this invention, through weight adjustment of distributed Ridge initialization, not only successfully reconstructs the local gene regulatory network of "receptor ESR1-downstream target AGR3 / CA12", but also effectively suppresses noise variables.

[0040] Finally, based on these three validated key biomarker sets and their corresponding non-zero regression parameters, the high-dimensional gene expression data and pathological stage data of the patient to be predicted are substituted into the Logistic regression model constructed in step S2. The probability prediction result of the patient's estrogen receptor status (ER positive / negative) can be calculated and output, providing a reliable tool for biomarker discovery and clinical precision diagnosis under multi-center privacy protection.

[0041] Specifically, a high-dimensional biomarker screening system for processing multi-source heterogeneous medical data includes: Memory and processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the method as claimed in any one of claims 1-9; the processor is configured to execute the following modules: The data acquisition and preprocessing module is used to collect patient observation data from various distributed sites and construct heterogeneous distributed high-dimensional datasets at each distributed site. The distributed local computing module is used to construct a regression model containing global parameters and heterogeneous parameters based on a heterogeneous distributed high-dimensional dataset. Each distributed site computes the initial consistent estimator and the local effective score function in parallel and sends the local effective score to the central site. The penalty equation construction module is used to calculate the density ratio weight at the central station based on the local effective score function of each distributed station, construct the global communication effective score function based on the density ratio weight, and introduce an adaptive discontinuous penalty term on the basis of the initial consistent estimator and the global communication effective score function to construct the penalty communication effective estimation equation. The iterative update module is used to iteratively update the global parameters and heterogeneity parameters through local quadratic approximation and coordinate descent method, and iteratively optimizes until the convergence threshold is met, and outputs the optimal global sparse estimate. The biomarker screening and prediction output module is used to screen key biomarkers based on global sparse estimates and predict the target disease status through key biomarkers.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0044] The present invention has been further described above with reference to specific embodiments. However, it should be understood that the specific description herein should not be construed as limiting the nature and scope of the present invention. Various modifications made to the above embodiments by those skilled in the art after reading this specification are all within the scope of protection of the present invention.

Claims

1. A high-dimensional biomarker screening method for processing multi-source heterogeneous medical data, characterized in that, The method includes: S1. Construct heterogeneous distributed high-dimensional datasets at each distributed site; S2. Construct a regression model containing global parameters and heterogeneous parameters based on the heterogeneous distributed high-dimensional dataset. Each of the distributed sites calculates the initial consistent estimator and the local effective score function in parallel, and sends the local effective score to the central site. S3. At the central station, the density ratio weight is calculated based on the local effective score function of each distributed station, and a global communication effective score function is constructed based on the density ratio weight. S4. Based on the initial consensus estimate and the global communication effective score function, an adaptive discontinuous penalty term is introduced to construct the penalty communication effective estimation equation; the global parameters are iteratively updated through local quadratic approximation and coordinate descent method, and iterative optimization is performed until the convergence threshold is met. S5. Output the optimal global sparse estimate and select key biomarkers based on the global sparse estimate, and use the key biomarkers to predict the target disease state.

2. The method of claim 1, wherein the method is a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data. Step S1 includes: Including the K distributed sites, the dataset of the k-th distributed site Including binary response variables and global covariate characteristics ,in The distributed site-specific covariate features are d, which represents the number of genes retained according to variance sorting, and n represents the number of samples. The global covariate features are high-dimensional data features shared across sites, and the global covariate features are processed locally at the distributed site using Z-score standardization.

3. The method and system for screening high-dimensional biomarkers for processing multi-source heterogeneous medical data according to claim 2, characterized in that, Step S2 includes: The regression model is a Logistic regression model, the global parameter is a high-dimensional sparse parameter, and the heterogeneity parameter is a data distribution difference parameter of the kth distributed site , and a local effective score function of the kth distributed site is calculated through an orthogonal projection technique : Where β is the global regression coefficient vector corresponding to d genes. To capture parameters of baseline risk and pathological stage heterogeneity across distributed sites, For the log-likelihood function of the distributed sites, and Divide the Hessian matrix corresponding to each distributed site into blocks, and each distributed site will assign the local effective score. A single transmission to the central station.

4. The method of claim 1, wherein the method is a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data. Step S3 specifically includes: S31, at the central site, compute a density ratio weight of the remote site k relative to the central site : S32, constructing a global communication effective score function using the density ratio weight : wherein, represents the effective score item for the i-th sample of the central site, denotes the local effective score.

5. The method of claim 1, wherein the method is a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data. Step S4 includes, S41, determine a penalty parameter by extending a Bayesian information criterion, introduce a non-smooth adaptive penalty term, and construct the penalized communication effective estimation equation ; S42. Set the iteration number t=0. In the current t-th iteration: Keeping the heterogeneity parameter at its initial fixed value, we update each component of the global parameter β(t) by performing one-dimensional coordinate descent after processing the penalty term using a local quadratic approximation. ; S43. Determine whether the global parameters have reached the convergence threshold; if the convergence threshold has not been reached, update the iteration count t=t+1, return to step S42 and re-execute until the convergence threshold is reached.

6. The method of claim 5, wherein the method further comprises: Step S41 specifically includes: Construct the effective estimation equation for the penalty communication: in, Let be the first derivative of the penalty function; the penalty function employs an adaptive LASSO strategy, and the derivative form of the penalty function is set as follows: Adaptive weights ; Let ν be the initial consistent estimator, and ν be a normal constant that controls the sparsity penalty.

7. The method of claim 5, wherein the method further comprises: Step S42 specifically includes: Initialization at the central site and start the iterative solution: S421. updating the initial heterogeneity parameter by maximizing the local log-likelihood at the central site : S422, for each component of the global parameter β , compute a local quadratic approximation coefficient of the penalty term , where is a small constant; and simultaneously compute an approximation of the diagonal element of the weighted Hessian matrix of the central site ; S423, using local quadratic approximation, update global parameters performing a coordinate descent update: 。 8. The method of claim 5, wherein the method further comprises: Step S43 specifically includes: The convergence of the algorithm is determined by calculating the Euclidean norm of the difference between the global parameters in two adjacent iterations. If the above equation is not satisfied, the iteration number is updated as t = t + 1, and step S42 is re-executed; if the above equation is satisfied, the iteration is stopped, and the current As a global sparse estimator Output.

9. The method of claim 1, wherein the method is a high-dimensional biomarker screening method for processing multi-source heterogeneous medical data. Step S5 specifically includes: The output step S4 obtains the optimal global sparse estimator The gene feature variables with non-zero coefficient values are extracted as key biomarkers. High-dimensional gene expression data and pathological stage-specific covariates of the target patient are extracted, substituted into the regression model, and the clinical disease status prediction results of the patient are output.

10. A high-dimensional biomarker screening system for processing multi-source heterogeneous medical data, characterized in that, include: Memory and processor; The memory is used to store a computer program; the processor is used to execute the computer program to implement the method as described in any one of claims 1-9; the processor is configured to execute the following modules: The data acquisition and preprocessing module is used to collect patient observation data from various distributed sites and construct heterogeneous distributed high-dimensional datasets at each distributed site. The distributed local computing module is used to construct a regression model containing global parameters and heterogeneous parameters based on the heterogeneous distributed high-dimensional dataset. Each of the distributed sites computes the initial consistent estimator and the local effective score function in parallel, and sends the local effective score to the central site. The penalty equation construction module is used to calculate the density ratio weight at the central station based on the local effective score function of each distributed station, construct the global communication effective score function based on the density ratio weight, introduce an adaptive discontinuous penalty term on the basis of the initial consistent estimator and the global communication effective score function, and construct a penalty communication effective estimation equation. The parameter iteration update module, under the condition of fixed initial values ​​of heterogeneous parameters, iteratively updates the global parameters through local quadratic approximation and coordinate descent method, and iterates and optimizes until the convergence threshold is met, and outputs the optimal global sparse estimate. The biomarker screening and prediction output module is used to screen key biomarkers based on global sparse estimates and predict the target disease status through the key biomarkers.

Images