A prostate cancer life cycle management system based on early screening databases

Abstract

This invention relates to the field of artificial intelligence technology and discloses a prostate cancer lifecycle management system based on an early screening database. The system includes modules for the early screening database, data processing, privacy compliance, risk assessment, and lifecycle management. The privacy compliance module utilizes a multi-mechanism privacy-preserving WGAN model, introducing progressive gradient pruning, progressive noise injection, and a dual-channel accounting mechanism based on Taylor expansion RDP and Gaussian difference privacy to achieve compliant data generation. The risk assessment module employs a look-ahead decision tree, combined with local look-ahead search and impurity aggregation, to improve the stability and consistency of risk stratification. The lifecycle management module formulates dynamic monitoring plans based on risk levels and triggers warnings and alerts when data anomalies occur. This invention enables the integrated utilization and dynamic management of multi-source data throughout its entire lifecycle while ensuring data privacy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a prostate cancer lifecycle management system based on an early screening database. Background Technology

[0002] Prostate cancer is one of the most common malignant tumors among men worldwide. Early detection and continuous management are crucial for reducing mortality and improving quality of life. Current clinical techniques often involve screening and risk assessment using serological indicators (such as PSA, fPSA, p2PSA), imaging examinations (such as ultrasound, MRI / PIRADS, CT), pathological results (such as Gleason scores), and follow-up records. However, these methods have several shortcomings: Firstly, multi-source screening data suffers from heterogeneous formats, high missing rates, and inconsistent standards, making efficient integration and modeling difficult. Secondly, databases involving sensitive health data often rely on single differential privacy or noise injection mechanisms for privacy protection, which can lead to a conflict between privacy protection and system availability, potentially posing a risk of data leakage and reducing the accuracy of risk prediction. Furthermore, most existing risk assessment models adopt traditional decision trees with greedy partitioning, which lack forward-looking consideration of the potential splitting capacity of sub-layers, are prone to getting stuck in local optima, resulting in unstable results and making it difficult to support long-term risk monitoring and dynamic adjustment. At the same time, life cycle management often remains at a one-time or phased assessment, lacking differentiated monitoring plans and real-time early warning mechanisms based on risk levels, and cannot achieve dynamic management of the entire life cycle of an individual. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in current prostate cancer screening and management technologies, such as difficulty in integrating multi-source data, insufficient privacy protection, lack of forward-looking risk assessment, and imperfect lifecycle management. It proposes a prostate cancer lifecycle management system based on an early screening database. The system's privacy compliance module constructs a multi-mechanism privacy-preserving WGAN model. During training, it introduces progressive gradient pruning, progressive noise injection, and a dual-channel privacy accounting mechanism combining adaptive Taylor expansion-remainder control (RDP) and Gaussian difference privacy to dynamically control privacy loss and ensure the compliance of generated data. The risk assessment module uses a next-layer lookahead decision tree, introducing local lookahead search and impurity aggregation on top of real-time impurity calculation, optimizing node partitioning strategies, improving the stability and global consistency of risk stratification results, and outputting personalized risk reports. The lifecycle management module formulates differentiated dynamic monitoring plans for different objects based on risk levels, and triggers early warnings and management prompts when preset anomalies are detected during data evolution, achieving dynamic monitoring and continuous management throughout the entire lifecycle. This invention addresses the challenges of heterogeneous data integration, insufficient privacy protection, and unstable risk stratification by combining multi-mechanism privacy protection modeling with forward-looking decision tree evaluation. It overcomes the limitations of single or phased evaluations and can dynamically generate follow-up and management plans based on risk changes, thereby achieving accurate monitoring and intelligent management throughout the entire lifecycle.

[0004] This invention provides a prostate cancer lifecycle management system based on an early screening database. The system includes an early screening database module, a data processing module, a privacy compliance module, a risk assessment module, and a lifecycle management module.

[0005] The early screening database module acquires early screening data from the target population;

[0006] The data processing module communicates with the early screening database module and is used to clean, integrate and standardize the early screening data to obtain standardized early screening data.

[0007] The privacy compliance module establishes a multi-mechanism privacy-preserving WGAN model. This model processes standardized early screening data to obtain compliant early screening data. The construction method of the multi-mechanism privacy-preserving WGAN model is as follows: a WGAN model (Wasserstein GAN) is established, and a dynamically decaying pruning threshold mechanism, a dynamically decaying Gaussian noise injection mechanism, and a dual-channel privacy accounting mechanism of adaptive Taylor expansion-remainder term control RDP and Gaussian difference privacy are introduced to optimize the gradient update process and privacy loss control process of the WGAN model, thus constructing the multi-mechanism privacy-preserving WGAN model.

[0008] The risk assessment module communicates and connects with the privacy compliance module to establish the next-level forward-looking decision tree. It processes early compliance screening data through the next-level forward-looking decision tree, generates risk stratification results, and generates personalized risk assessment reports based on the risk stratification results. The construction method of the next-level forward-looking decision tree is as follows: establish a decision tree, introduce first-order local forward-looking search and forward-looking impurity aggregation, optimize the node partitioning strategy and impurity assessment method of the decision tree, and obtain the next-level forward-looking decision tree.

[0009] The lifecycle management module communicates with the risk assessment module and is used to develop differentiated dynamic monitoring plans for individuals with different risk levels based on risk stratification results. When preset abnormal changes are detected in the screening, follow-up and health data of individuals over time, risk warnings are triggered, and corresponding health management measures are automatically generated and recommended, so as to realize dynamic monitoring and continuous management of individuals throughout their entire life cycle.

[0010] Furthermore, the process of processing standardized early screening data through a multi-mechanism privacy-preserving WGAN model to obtain compliant early screening data specifically includes the following steps:

[0011] Step S1: Data Access: Set the batch size and sampling rate to perform random subsampling on the standardized early screening data to obtain the training batch dataset;

[0012] Step S2: Model Building: Initialize the multi-mechanism privacy-preserving WGAN model, which includes a generator and a critic. Initialize the parameters of the multi-mechanism privacy-preserving WGAN model, including the initial gradient pruning threshold, initial noise magnitude, momentum coefficient, and asymptotic decay parameter.

[0013] Step S3: Progressive Gradient Clipping: Input the training batch dataset into the multi-mechanism privacy-preserving WGAN model for iterative training. The initial gradient clipping threshold is dynamically adjusted through a dynamically decaying clipping threshold mechanism to obtain a progressively decaying clipping threshold. Under the constraint of the progressively decaying clipping threshold, norm clipping is performed on the gradient vector of each sample in the training batch dataset to limit the influence of a single sample on the overall parameter update, resulting in the clipped batch gradient set.

[0014] Step S4: Progressive noise injection: In each iteration of the iterative training, the batch gradient set after pruning is averaged to obtain the batch average gradient; the initial noise amplitude is dynamically adjusted through a dynamically decaying Gaussian noise injection mechanism to generate a gradually decaying noise amplitude; under the constraint of the gradually decaying noise amplitude, a random noise term following a Gaussian distribution is superimposed on the batch average gradient to obtain the noisy gradient set.

[0015] Step S5: Introduce the momentum mechanism and combine it with the noisy gradient set to perform row-by-row iterative updates on the parameters of the multi-mechanism privacy-preserving WGAN model;

[0016] Step S6: During the parameter iteration update process, a fixed-size subsampling is introduced to replace the sampling probability in Rényi differential privacy (RDP), thus constructing a fixed-size sampling-RDP. Based on the fixed-size sampling-RDP, Taylor expansion, remainder control, and adaptive order selection are introduced to refine the single-step privacy loss bound of the fixed-size sampling-RDP, thereby obtaining an adaptive Taylor expansion-remainer control RDP with non-asymptoticity, controllable accuracy, and dynamic adaptability. A dual-channel privacy accounting mechanism of adaptive Taylor expansion-remainer control RDP and Gaussian differential privacy (GDP) is adopted to measure and accumulate privacy loss, thereby realizing dynamic tracking and constraint of the overall privacy budget.

[0017] Step S7: Set privacy budget constraints. After completing iterative training and meeting the privacy budget constraints, output compliance early screening data through the output layer of the WGAN model with multi-mechanism privacy protection.

[0018] Furthermore, the process of processing early compliance screening data through the next layer of forward-looking decision tree to generate risk stratification results specifically includes the following steps:

[0019] Step B1: Standardize and feature-align the early compliance screening data to form the input sample set;

[0020] Step B2: At the current node of the next look-ahead decision tree, randomly extract a subset of features from all features of the input sample set, and construct multiple candidate split thresholds for each feature based on the quantile to form candidate split pairs.

[0021] Step B3: Based on the candidate split pairs, the input sample set is divided into a left subset and a right subset, and the Gini impurity of the left subset and the right subset is calculated respectively; the Gini impurity is weighted according to the ratio of the number of samples in the left subset and the right subset to obtain the parent layer split impurity index, and several feature sets with the best performance are retained according to the parent layer split impurity index to form the parent layer retained feature pool.

[0022] Step B4: Perform a first-order local look-ahead search on the left and right subsets respectively; within each subset, extract features from the parent layer's retained feature pool, and generate secondary candidate thresholds based on the value distribution of the extracted features in the subset. Using the secondary candidate thresholds as the splitting condition, calculate the feasible splitting impurity of the extracted features in the subset, obtain the minimum impurity value of the left subset and the minimum impurity value of the right subset, and combine them through the lower-level aggregation function to obtain the look-ahead impurity index.

[0023] Step B5: Combining the impurity index of the parent layer and the prospective impurity index with the depth perception weighting factor and stability adjustment parameter, calculate the total score of the candidate split pair, select the candidate split pair that minimizes the total score as the optimal split of the current node, so as to take into account both the immediate discrimination capability and the potential prospective capability in the node partitioning process, and achieve global consistency optimization of risk stratification decision-making.

[0024] Step B6: Preset the maximum depth of the decision tree, recursively execute steps B2 to B5 until the maximum depth of the decision tree is reached, and generate a complete tree structure; based on the complete tree structure, output the risk stratification results of individuals.

[0025] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0026] First, this invention achieves compliant processing of multi-source sensitive data in the early screening database by constructing a multi-mechanism privacy-preserving WGAN model in the privacy compliance module. Through progressive gradient pruning and progressive noise injection mechanisms, the balance between privacy protection and model training is improved, resolving the performance degradation problem caused by over-pruning or noise interference in existing methods. Simultaneously, by combining an adaptive Taylor expansion-remainder control (RDP) and Gaussian difference privacy dual-channel accounting mechanism, dynamic measurement and constraint of privacy loss are achieved, enhancing data security and compliance throughout the system's entire lifecycle.

[0027] Secondly, this invention introduces a next-level prospective decision tree into the risk assessment module, enabling multi-level analysis and prospective assessment of individual early screening data. By adding local prospective search and impurity aggregation on top of real-time impurity calculation, the stability of node partitioning and the consistency of the overall structure are improved, solving the problems of traditional decision trees easily getting trapped in local optima and unstable stratification results. This design not only enhances the accuracy and robustness of risk stratification results but also enables individualized risk reports to better reflect potential future trends, providing a reliable basis for subsequent dynamic management.

[0028] Finally, this invention establishes a differentiated dynamic monitoring mechanism based on risk levels within the lifecycle management module, enabling data tracking and anomaly warnings that evolve over time. By setting targeted monitoring plans for individuals with different risk levels and automatically triggering warnings and alerts when preset anomalies occur in the monitoring data, the system's real-time response capability to risk changes is improved. This overcomes the limitations of traditional methods that are confined to single or phased assessments, enhancing the system's continuity and adaptability throughout the entire lifecycle management process. Overall, this invention organically combines multi-mechanism privacy-preserving modeling with forward-looking decision tree evaluation, achieving simultaneous improvements in data compliance, risk assessment stability, and full-cycle dynamic management capabilities, providing excellent technical support for the intelligent utilization of prostate cancer early screening data. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a module of a prostate cancer life cycle management system based on an early screening database proposed in this invention;

[0030] Figure 2 This is a line graph of PSA dynamic monitoring proposed in Example 6. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] Example 1, according to Figure 1 This invention provides a prostate cancer lifecycle management system based on an early screening database. The system includes an early screening database module, a data processing module, a privacy compliance module, a risk assessment module, and a lifecycle management module.

[0033] The early screening database module acquires early screening data from the target population. The early screening data includes: serological indicators (PSA, fPSA, p2PSA), imaging data (transrectal ultrasound, prostate MRI / PIRADS, CT), pathological results (including Gleason score), medical records and medication records, family history and lifestyle information, gene / molecular marker test results, and follow-up and prognostic outcome information.

[0034] The data processing module communicates with the early screening database module and is used to clean, integrate and standardize the early screening data to obtain standardized early screening data.

[0035] The privacy compliance module establishes a multi-mechanism privacy-preserving WGAN model. This model processes standardized early screening data to obtain compliant early screening data. The construction method of the multi-mechanism privacy-preserving WGAN model is as follows: a WGAN model (Wasserstein GAN) is established, and a dynamically decaying pruning threshold mechanism, a dynamically decaying Gaussian noise injection mechanism, and a dual-channel privacy accounting mechanism of adaptive Taylor expansion-remainder term control RDP and Gaussian difference privacy are introduced to optimize the gradient update process and privacy loss control process of the WGAN model, thus constructing the multi-mechanism privacy-preserving WGAN model.

[0036] The risk assessment module communicates and connects with the privacy compliance module to establish the next-level prospective decision tree. This next-level prospective decision tree processes early compliance screening data, generates risk stratification results, and based on these results, produces a personalized risk assessment report. The report includes the current risk level, predicted future disease probability, and recommended follow-up screening. The next-level prospective decision tree is constructed as follows: a decision tree is established, a first-order local prospective search and prospective impurity aggregation are introduced, and the node partitioning strategy and impurity assessment method of the decision tree are optimized to obtain the next-level prospective decision tree.

[0037] The lifecycle management module communicates with the risk assessment module and is used to develop differentiated dynamic monitoring plans for individuals with different risk levels based on risk stratification results. When preset abnormal changes are detected in the screening, follow-up and health data of individuals over time, risk warnings are triggered, and corresponding health management measures are automatically generated and recommended, so as to realize dynamic monitoring and continuous management of individuals throughout their entire life cycle.

[0038] Example 2, based on Example 1, describes the process of obtaining compliant early screening data by processing standardized early screening data using a multi-mechanism privacy-preserving WGAN model. The specific steps include:

[0039] Step S1: Data Access: Set the batch size and sampling rate to perform random subsampling on the standardized early screening data to obtain the training batch dataset;

[0040] Step S2: Model Building: Initialize the multi-mechanism privacy-preserving WGAN model, which includes a generator and a critic. Initialize the parameters of the multi-mechanism privacy-preserving WGAN model, including the initial gradient pruning threshold, initial noise magnitude, momentum coefficient, and asymptotic decay parameter.

[0041] Step S3: Progressive Gradient Clipping: Input the training batch dataset into the multi-mechanism privacy-preserving WGAN model for iterative training. The initial gradient clipping threshold is dynamically adjusted through a dynamically decaying clipping threshold mechanism to obtain a progressively decaying clipping threshold. Under the constraint of the progressively decaying clipping threshold, norm clipping is performed on the gradient vector of each sample in the training batch dataset to limit the influence of a single sample on the overall parameter update, resulting in the clipped batch gradient set.

[0042] Formula for dynamic decay pruning threshold mechanism:

[0043] ;

[0044] in, Indicates the current iteration number. Indicates the maximum number of iterations. This represents the initial gradient clipping threshold. The exponential factor that controls the decay rate. Indicates the first The gradually decreasing pruning threshold in each iteration;

[0045] This design allows the model to impose stricter constraints on gradients in the early iteration phase to enhance privacy protection and reduce the impact of excessively large gradients; while gradually relaxing the constraints in the later stages of training to improve the model's expressiveness and convergence ability.

[0046] Step S4: Progressive noise injection: In each iteration of the iterative training, the batch gradient set after pruning is averaged to obtain the batch average gradient; the initial noise amplitude is dynamically adjusted through a dynamically decaying Gaussian noise injection mechanism to generate a gradually decaying noise amplitude; under the constraint of the gradually decaying noise amplitude, a random noise term following a Gaussian distribution is superimposed on the batch average gradient to obtain the noisy gradient set.

[0047] Formula for dynamic decay Gaussian noise injection mechanism:

[0048] ;

[0049] in, Indicates the first The noise amplitude gradually decreases during each iteration. Indicates the initial noise amplitude. Indicates the noise attenuation index factor;

[0050] This design enhances the privacy protection of sensitive samples by introducing significant noise in the early stages of training, thereby reducing the risk of data leakage. In the later stages of training, the noise amplitude is gradually reduced to mitigate the interference of noise on the true gradient direction, thus improving the convergence stability of the model and the effectiveness of the generated data.

[0051] Step S5: Introduce a momentum mechanism and combine it with the noisy gradient set to iteratively update the parameters of the multi-mechanism privacy-preserving WGAN model, obtaining the updated model parameter set. The formula used is as follows:

[0052] ;

[0053] in, Indicates the first The momentum term at the next iteration This represents the momentum term (historical cumulative amount) from the previous iteration. Indicates the momentum coefficient. This represents the average gradient after adding noise, which is obtained from the set of gradients after adding noise.

[0054] ;

[0055] in, Indicates the first Model parameters for the next iteration Indicates the first Model parameters for the next iteration Indicates the learning rate;

[0056] Step S6: During the parameter iteration update process, a fixed-size subsampling is introduced to replace the sampling probability in Rényi differential privacy (RDP), thus constructing a fixed-size sampling-RDP. Based on the fixed-size sampling-RDP, Taylor expansion, remainder control, and adaptive order selection are introduced to refine the single-step privacy loss bound of the fixed-size sampling-RDP, thereby obtaining an adaptive Taylor expansion-remainer control RDP with non-asymptoticity, controllable accuracy, and dynamic adaptability. A dual-channel privacy accounting mechanism of adaptive Taylor expansion-remainer control RDP and Gaussian differential privacy (GDP) is adopted to measure and accumulate privacy loss, thereby realizing dynamic tracking and constraint of the overall privacy budget.

[0057] Adaptive Taylor Expansion-Residual Term Control RDP (FS-TE RDP):

[0058] Under fixed-size sampling (FS) conditions, single step The upper bound of the privacy loss in Rényi differential privacy (RDP) depends on the sampling probability. Complex nonlinear functions; due to the sampling probability in the differential privacy stochastic gradient descent algorithm. Typically small, this nonlinear function can be considered as about The analytic function, its in Perform a finite-order Taylor expansion in the vicinity and provide an explicit upper bound for the remainder of the higher-order terms; in this way, a given precision threshold can be satisfied. Under the given conditions, the minimum Taylor expansion order is adaptively selected to obtain a single-step RDP upper bound with non-asymptoticity, controllable error, and high precision.

[0059] The logarithmic inner term of privacy loss in fixed-scale sampling-RDP can be abstracted as a function of the sampling rate. The analytic function is defined as a one-step kernel function:

[0060] ;

[0061] And satisfy the remaining constraints:

[0062] ;

[0063] in, This indicates the order of the RDP, used to control the Rényi divergence order used when measuring privacy loss. This represents the standard deviation of Gaussian noise. This represents the kernel function in one step, i.e., the logarithmic term of the fixed-size sample-RDP that varies with the sampling rate. The changing nonlinear dependency structure; This represents the index of the order of the Taylor expansion. Indicates the Taylor truncation order. To represent factorial; This indicates sensitivity in differential privacy. This indicates a fixed-size sampling - RDP one-step kernel function in The First derivative coefficients; This represents the remainder term of the Taylor expansion of the kernel function in one step, i.e., the remainder term control;

[0064] Step S7: Set privacy budget constraints. After completing iterative training and meeting the privacy budget constraints, output compliance early screening data through the output layer of the WGAN model with multi-mechanism privacy protection.

[0065] Example 3: This example is based on Example 1. In this example, the standardized early screening data is processed through the WGAN model to obtain compliant early screening data.

[0066] Example 4, based on Example 2, describes the process of processing early compliance screening data and generating risk stratification results through a next-level forward-looking decision tree. The specific steps include:

[0067] Step B1: Standardize and feature-align the early compliance screening data to form the input sample set;

[0068] Step B2: At the current node of the next look-ahead decision tree, randomly extract a subset of features from all features of the input sample set, and construct multiple candidate split thresholds for each feature based on the quantile to form candidate split pairs.

[0069] Step B3: Based on the candidate split pairs, the input sample set is divided into a left subset and a right subset, and the Gini impurity of the left subset and the right subset is calculated respectively. The Gini impurity is weighted according to the ratio of the number of samples in the left subset and the right subset to obtain the parent layer split impurity index, which is used to characterize the immediate discrimination ability of the candidate split at the current node. Based on the parent layer split impurity index, several feature sets with the best performance are retained to form the parent layer reserved feature pool for use in the next layer look-ahead search.

[0070] Step B4: Perform a first-order local look-ahead search on the left and right subsets respectively; within each subset, extract features from the parent layer's reserved feature pool, and generate a secondary candidate threshold based on the value distribution of the extracted features in the subset. Using the secondary candidate threshold as the splitting condition, calculate the feasible splitting impurity of the extracted features in the subset, obtain the minimum impurity value of the left subset and the minimum impurity value of the right subset, and combine them through the lower-level aggregation function to obtain the look-ahead impurity index, which is used to characterize the discriminative ability of the parent layer candidate splits under the next layer of potential splits;

[0071] Formation of forward-looking impurities through aggregation:

[0072] ;

[0073] in, Indicates a forward-looking impurity index. This represents the minimum impurity value of the left subset. This represents the minimum impurity value of the right subset;

[0074] In the traditional decision tree construction process, node partitioning generally adopts a greedy strategy, that is, the optimal splitting feature and threshold are selected based solely on the immediate impurity index of the current node (such as Gini impurity and information gain). Although this method is computationally simple, it cannot predict the potential splitting ability of the child layer because it relies entirely on the partitioning effect of the parent layer. It is easy to get stuck in local optima, which leads to the instability of the overall tree structure and even reduces the accuracy of classification or risk stratification.

[0075] To address this, this invention proposes a forward-looking impurity evaluation mechanism based on real-time impurity calculation (parent layer discrimination). This mechanism performs a first-order local look-ahead search on the left and right subsets of the candidate split pairs at the parent layer, obtaining the minimum impurity value of the subset under the potential next-layer split, and forming a forward-looking impurity index through an aggregation function. The index simultaneously considers optimal potential (ensuring at least one subset has good splitting ability) and overall balance (balancing the splitting trends of both subsets), overcoming the limitation of conventional decision trees that rely solely on the current splitting effect. This allows node selection to consider both current and future splitting quality, effectively reducing the risk of local optima and improving the stability and prediction accuracy of the overall tree structure.

[0076] Step B5: Combining the impurity index of the parent layer and the prospective impurity index with the depth perception weighting factor and stability adjustment parameter, calculate the total score of the candidate split pair, select the candidate split pair that minimizes the total score as the optimal split of the current node, so as to take into account both the immediate discrimination capability and the potential prospective capability in the node partitioning process, and achieve global consistency optimization of risk stratification decision-making.

[0077] The overall scoring formula is as follows:

[0078] ;

[0079] in, This represents the total score of the candidate split pair. This indicates the impurity index of the parent layer partition. This indicates the tree depth of the current node. This represents the average impurity of the current subtree. This represents the depth-sensing weighting factor. Indicates the stability adjustment parameter. This represents a positive stability factor to prevent weight collapse or a zero denominator in extreme cases.

[0080] Step B6: Preset the maximum depth of the decision tree, recursively execute steps B2 to B5 until the maximum depth of the decision tree is reached, and generate a complete tree structure; based on the complete tree structure, output the risk stratification results of individuals.

[0081] explain:

[0082] In the next-level prospective decision tree construction process of the present invention, step B5 is used to complete the comprehensive scoring of candidate split pairs and the optimal split selection at the current node; when the optimal split pair is determined, the current node is divided into left and right child nodes, and the input sample set is also allocated to the left and right subsets accordingly.

[0083] Based on this, the system does not simply return to step B2, but recursively executes steps B2 to B5 for the newly generated left and right child nodes respectively: that is, it re-extracts feature subsets within each child node, constructs candidate split thresholds, calculates instantaneous impurity and prospective impurity, and selects the optimal split through a comprehensive scoring mechanism; through this recursive call, the decision tree can be expanded layer by layer until the preset maximum depth limit or the minimum number of samples in the leaf node is reached.

[0084] Therefore, step B5 is not only the end point of the current node split, but also the starting point of the subsequent recursive process. The design ensures that the risk stratification tree takes into account both the immediate discrimination effect and the potential future split quality in a global scope, thereby significantly improving the stability and prediction performance of the tree structure. Finally, in step B6, a complete risk stratification result and an individualized risk assessment report are generated.

[0085] Example 5 is based on Example 2. In this example, compliance early screening data is processed using a conventional decision tree to generate risk stratification results.

[0086] Example 6, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment,

[0087] The risk assessment module processes compliance early screening data through the next layer of forward-looking decision tree, generates risk stratification results, and generates personalized risk assessment reports based on the risk stratification results. The reports include the current risk level, future disease probability prediction, and recommended follow-up screening.

[0088] In this embodiment, we take a 50-year-old male individual A as an example. The data entered by this individual in the early screening database is as follows:

[0089] Serological markers: PSA = 6.8 ng / mL, fPSA = 0.75 ng / mL, p2PSA = 24 pg / mL;

[0090] Imaging data: Prostate MRI / PIRADS score = 4 points;

[0091] Pathological findings: No biopsy record available;

[0092] Family history and lifestyle: The father had prostate cancer and had a high-fat diet and lack of exercise.

[0093] Follow-up data: PSA showed a continuous upward trend within one year (5.4→6.2→6.8 ng / mL).

[0094] Risk stratification results:

[0095] Current risk level: Medium to high risk;

[0096] Predicted probability of developing prostate cancer in the next 5 years: 42.7%;

[0097] Personalized risk report: The report indicates that the individual's elevated PSA is accompanied by a high PIRADS score, and the risk is significant due to family history. Further screening should be strengthened.

[0098] The lifecycle management module develops differentiated dynamic monitoring plans for individuals with different risk levels based on risk stratification results. When preset abnormal changes are detected in the screening, follow-up and health data of individuals over time, risk warnings are triggered, and corresponding health management measures are automatically generated and recommended, so as to realize dynamic monitoring and continuous management of individuals throughout their entire lifecycle.

[0099] The lifecycle management module develops a differentiated dynamic monitoring plan for this individual:

[0100] A PSA re-examination and imaging examination should be performed every 6 months;

[0101] Lifestyle survey data and health records are updated every 12 months;

[0102] During the monitoring process, a threshold condition is set: if the PSA level rises by more than 0.75 g / mL twice consecutively or the PIRADS score rises to 5 points, a risk warning is triggered.

[0103] Figure 2 A line chart showing PSA dynamic monitoring over 0-15 months; Figure 2 In the middle, the horizontal axis represents time (months), and the vertical axis represents PSA value (ng / mL); the blue line represents the actual PSA value, the red dashed line represents the warning threshold (7.5 g / mL), and the red solid dot represents the warning trigger point;

[0104] After a year of follow-up, the system detected that PSA had increased to 7.6 g / mL, triggering the abnormal threshold. The system automatically issued an early warning and generated lifestyle management recommendations: increase the frequency of aerobic exercise, improve high-fat diet habits, and control weight.

[0105] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A prostate cancer lifecycle management system based on an early screening database, characterized in that: The system includes: Early screening database module, to obtain early screening data; The data processing module processes the early screening data to obtain standardized early screening data; The privacy compliance module establishes a multi-mechanism privacy protection WGAN model, which processes standardized early screening data to obtain compliant early screening data. The risk assessment module establishes a next-level forward-looking decision tree, processes early compliance screening data through this tree, and generates risk stratification results. The next-level prospective decision tree is constructed as follows: a decision tree is built, a first-order local prospective search and prospective impurity aggregation are introduced, the node partitioning strategy and impurity evaluation method of the decision tree are optimized, and the next-level prospective decision tree is obtained. The process of generating risk stratification results through the next layer of forward-looking decision tree includes the following steps: Step B1: Standardize and feature-align the early compliance screening data to form the input sample set; Step B2: At the current node of the next look-ahead decision tree, extract a subset of features from the input sample set to form candidate split pairs; Step B3: Based on the candidate split pairs, the input sample set is divided into a left subset and a right subset to obtain the parent layer split impurity index. The optimal feature set is retained according to the parent layer split impurity index to form the parent layer retained feature pool. Step B4: Perform a first-order local look-ahead search on the left and right subsets respectively; within each subset, extract features from the parent layer's retained feature pool, and generate secondary candidate thresholds based on the extracted features. Using the secondary candidate thresholds as the splitting condition, calculate the feasible splitting impurity of the extracted features, obtain the minimum impurity value of the left subset and the minimum impurity value of the right subset, and combine them through the lower-level aggregation function to obtain the look-ahead impurity index. Step B5: Combine the parent layer partition impurity index and the prospective impurity index to calculate the total score of the candidate split pairs and select the optimal split for the current node; Step B6: Preset the maximum depth of the decision tree, recursively execute steps B2 to B5 until the maximum depth of the decision tree is reached, generate a complete tree structure, and output the risk stratification results.

2. The prostate cancer lifecycle management system based on an early screening database according to claim 1, characterized in that: The construction method of the multi-mechanism privacy-preserving WGAN model is as follows: establish the WGAN model, introduce a dynamically decaying pruning threshold mechanism, a dynamically decaying Gaussian noise injection mechanism, and a dual-channel privacy accounting mechanism, optimize the gradient update process and privacy loss control process of the WGAN model, and construct the multi-mechanism privacy-preserving WGAN model.

3. A prostate cancer lifecycle management system based on an early screening database according to claim 2, characterized in that: The process of obtaining compliance early screening data through the multi-mechanism privacy-preserving WGAN model includes the following steps: Step S1: Perform random subsampling on the standardized early screening data to obtain the training batch dataset; Step S2: Initialize the parameters of the multi-mechanism privacy-preserving WGAN model, including the initial gradient pruning threshold and the initial noise magnitude; Step S3: Input the training batch dataset into the multi-mechanism privacy-preserving WGAN model for iterative training. The initial gradient clipping threshold is dynamically adjusted through a dynamically decaying clipping threshold mechanism to obtain a progressively decaying clipping threshold. Under the constraint of the progressively decaying clipping threshold, norm clipping is performed on the gradient vector of each sample in the training batch dataset to obtain the clipped batch gradient set. Step S4: In each iteration of the iterative training, the batch gradient set after clipping is averaged to obtain the batch average gradient; the initial noise amplitude is dynamically adjusted through a dynamically decaying Gaussian noise injection mechanism to generate a gradually decaying noise amplitude; under the constraint of the gradually decaying noise amplitude, a random noise term following a Gaussian distribution is superimposed on the batch average gradient to obtain the noisy gradient set. Step S5: Introduce the momentum mechanism and combine it with the noisy gradient set to iteratively update the parameters of the multi-mechanism privacy-preserving WGAN model; Step S6: During the parameter iteration update process, a dual-channel privacy accounting mechanism of adaptive Taylor expansion-remainder control RDP and Gaussian difference privacy is used to measure privacy loss and obtain the privacy budget; Step S7: Set privacy budget constraints. After completing iterative training and meeting the privacy budget constraints, output compliance early screening data through the output layer of the multi-mechanism privacy-preserving WGAN model.

4. A prostate cancer lifecycle management system based on an early screening database according to claim 3, characterized in that: The method for constructing the adaptive Taylor expansion-remainder control RDP is as follows: a fixed-size subsampling is introduced to replace the sampling probability in Rényi differential privacy, thus constructing a fixed-size sampling-RDP; Taylor expansion, remainder control, and adaptive order selection are introduced to refine the single-step privacy loss bound of the fixed-size sampling-RDP, thus obtaining the adaptive Taylor expansion-remainder control RDP.

Images