Health data model optimization method, system, and program product

Abstract

The application is suitable for the technical field of data privacy and model optimization, and provides a health data model optimization method, system and program product. The method comprises the following steps: performing cross-modal causal invariant representation distillation, extracting stable causal features and generating causal sensitivity information; constructing a privacy budget topology structure based on the information, and generating a topology mask combined with multi-dimensional constraints to screen effective model parameter fragments; performing differential privacy enhancement and verifiable weighted aggregation on the effective fragments to generate a global model; performing consistency verification on the model based on a causal invariant subspace and performing a closed-loop write-back to form feedback data; and performing incremental evolution on the allocation rules based on the feedback data and chaining and solidifying the operation credentials. The application improves the model precision and the privacy protection level, reduces the computing overhead, and meets the compliance audit requirements.

Description

Technical Field

[0001] This application belongs to the field of data privacy and model optimization technology, and in particular relates to a method, system and program product for optimizing a health data model. Background Technology

[0002] With the collaborative development of medical consortia, the growth of the gene testing industry, and the widespread adoption of wearable health devices, medical and health data is exhibiting characteristics of distributed storage, multimodal heterogeneity, and cross-institutional collaborative computing. Existing health data model aggregation technologies, such as traditional federated learning or static differential privacy methods, have many inherent shortcomings in practical applications, making it difficult to meet the industry's demands for high accuracy, high security, and high compliance.

[0003] Specifically, existing technologies suffer from the following problems: 1. Poor modal heterogeneity adaptability: Image, genetic, and wearable device time-series data belong to different systems. Simple normalization cannot effectively remove pseudo-correlation features introduced by factors such as device, environment, and acquisition method, nor can it extract stable causal features across scenarios, resulting in large fluctuations in the accuracy of the aggregated model and poor generalization ability. 2. Coarse privacy protection mechanisms: Static and uniform privacy budget allocation methods are commonly used, failing to combine the causal importance of data features, the difference in parameter contribution, and the iteration rounds for fine amortization, resulting in insufficient protection of high-value parameters and wasting privacy budget on low-value parameters. 3. High aggregation computational overhead: Traditional technologies usually use all local model parameters indiscriminately for aggregation. A large number of invalid parameters with low contribution not only waste computing resources but also reduce the convergence accuracy and final performance of the model due to excessive privacy noise. 4. Lack of a closed-loop consistency verification process: The aggregation update process only focuses on the convergence of parameter values, without verifying the consistency of the model's inherent causal structure. This can easily disrupt the inherent stable causal relationships in the data and even cause model performance degradation during iteration. 5. System updates lead to service interruptions: Traditional rule updates or policy iterations rely on a global system restart, which cannot meet the 24 / 7 continuous service requirements of medical scenarios. 6. Lack of end-to-end compliance auditing: The lack of a systematic system for credential storage, traceability, and verification makes it difficult to meet increasingly stringent medical data security and compliance regulatory requirements. Summary of the Invention

[0004] This application provides a method, system, and program product for optimizing a health data model, aiming to solve the technical problems existing in the prior art, such as poor modality adaptation, crude allocation of privacy budget, large overhead of invalid parameter aggregation, easy degradation of model iteration, system downtime required for updates, and lack of compliant evidence storage throughout the entire process.

[0005] In a first aspect, embodiments of this application provide a method for optimizing a health data model, including: Cross-modal causal invariant characterization distillation was performed on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features; Based on the aforementioned cross-domain stable causal characteristics, causal sensitivity information is generated that is bound to the parameter slices of each local model; Based on the causal sensitivity information, a privacy budget topology is constructed and the budget is allocated. A topology mask is generated by combining multi-dimensional constraints to select effective model parameter slices from the local model parameter slices. Differential privacy enhancement processing and verifiable weighted aggregation are performed on the effective model parameters in segments to generate a global model; Based on a predefined causal invariant subspace, the global model and the local models of each participant are checked for causal consistency, and closed-loop write-back is performed according to the check results to form feedback data. Based on the feedback data, the allocation rules in the privacy budget topology are incrementally evolved, edge heterogeneous operators are hot-swapped and deployed, and the full-process operation credentials are solidified in the form of a graph.

[0006] Secondly, embodiments of this application provide a health data model optimization system, including: The causal characterization module is used to perform cross-modal causal invariant characterization distillation on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features, and, based on the cross-domain stable causal features, generate causal sensitivity information bound to the parameter slices of each local model. The privacy topology management module is used to construct a privacy budget topology structure and allocate the budget based on the causal sensitivity information, and generate a topology mask by combining multi-dimensional constraints, so as to filter out effective model parameter fragments from the local model parameter fragments. The secure aggregation module is used to perform differentiated privacy enhancement processing and verifiable weighted aggregation on the effective model parameters to generate a global model; The consistency verification module is used to perform causal consistency verification between the global model and the local models of each participant based on a predefined causal invariant subspace, and to perform closed-loop write-back based on the verification results to form feedback data. The rule evolution module is used to incrementally evolve the allocation rules in the privacy budget topology based on the feedback data, and to hot-swap deploy edge heterogeneous operators. The voucher management module is used to solidify all operation vouchers in the form of a graph.

[0007] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the health data model optimization method described in any one of the first aspects above.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the health data model optimization method described in any one of the first aspects.

[0009] Fifthly, embodiments of this application provide a computer program product that, when run on a computer device, causes the computer device to execute the health data model optimization method described in any one of the first aspects.

[0010] In this embodiment, by performing cross-modal causal invariant representation distillation, modal differences and domain shifts between data introduced by factors such as devices and environment can be effectively eliminated, stable causal features can be extracted, and accurate alignment of multi-source heterogeneous data such as images, genes, and wearable device data can be achieved, thereby improving the accuracy and generalization ability of the aggregation model. By constructing a multi-level privacy budget topology structure based on causal sensitivity, the privacy budget can be allocated in a differentiated and refined manner according to dimensions such as institution, data domain, parameter importance, and iteration round, avoiding the coarseness and waste of the traditional static allocation mode, and improving model utility while effectively protecting data privacy. By combining multi-dimensional constraints to generate a topology mask, high-value effective model parameter slices can be accurately selected to participate in aggregation, eliminating a large number of low-contribution invalid parameters, thereby significantly reducing the invalid computational overhead in the aggregation process, reducing unnecessary privacy noise accumulation, and optimizing the convergence effect of the model. By establishing a closed-loop causal consistency verification mechanism, the deviation of the causal structure between the global model and the local model is checked before the model update is written back, ensuring that the update of the global model does not destroy the inherent causal structure of the local model, fundamentally solving the model performance degradation problem commonly found in traditional federated learning. Through incremental evolution of allocation rules based on feedback data and the adoption of a double-buffer update mechanism, the updates of system rules and edge operators can be completed without service interruption, achieving zero-downtime self-iterative updates and ensuring the continuity of critical businesses such as healthcare. By solidifying the entire process operation vouchers in a graph format, a full-process compliance audit system is constructed, making all operation nodes, parameter changes, budget consumption, and other behaviors traceable, verifiable, and auditable, meeting the stringent compliance and regulatory requirements for healthcare data.

[0011] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a health data model optimization method according to an embodiment of this application.

[0014] Figure 2 This is a schematic diagram of the structure of a health data model optimization system according to an embodiment of this application.

[0015] Figure 3 This is a schematic diagram illustrating a privacy budget topology according to an embodiment of this application. Detailed Implementation

[0016] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0017] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0018] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0019] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0020] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0022] Please see Figure 1 This application provides a method for optimizing a health data model, which includes the following steps: S101, perform cross-modal causal invariant characterization distillation on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features.

[0023] The multi-source heterogeneous health data refers to multimodal health data entities originating from different medical institutions, testing equipment, or wearable terminals, including at least three categories: image health data, genetic health data, and wearable time-series health data.

[0024] The cross-modal causal invariant characterization distillation refers to extracting stable causal structure information that is common to different modalities and is not affected by external factors by constructing local causal graphs for each modality and removing pseudo-association features introduced by non-essential factors such as equipment and environment.

[0025] The cross-domain stable causal feature refers to a causal feature vector obtained from the distillation process that can maintain consistency across different data domains.

[0026] S101 specifically includes: Step a1: For different data modalities in the multi-source heterogeneous health data, construct their respective local causal graphs and remove pseudo-correlation features introduced by non-essential factors.

[0027] The local causal graph is a directed graph constructed using data variables within a single modality as nodes and causal dependencies between variables as edges. The non-essential factors include equipment model, data acquisition environment, and differences in operators. The purpose of this step is to remove pseudo-associations on the data surface and retain the true causal structure. Different modalities of data follow different causal logics. By constructing local causal graphs separately, the causal relationships within each modality can be fully preserved and analyzed, and pseudo-associations introduced by non-essential factors can be eliminated.

[0028] Step a2: Extract the common causal invariants across modalities from the local causal graph, and use the common causal invariants as the cross-domain stable causal features.

[0029] The shared causal invariants refer to causal subgraphs that exist in local causal graphs across different modalities and have consistent structural relationships. By extracting shared invariants, stable alignment of cross-modal data can be achieved, improving the generalization ability of subsequent models.

[0030] S102, Based on the cross-domain stable causal characteristics, generate causal sensitivity information that is bound to the parameter slices of each local model.

[0031] The local model parameter shards refer to subsets or fragments of model parameters independently trained by each data participant (such as a medical institution) using their local data. In distributed learning scenarios, these shards serve as carriers of knowledge from each participant and are the core objects processed and aggregated by this method.

[0032] The causal sensitivity information is used to quantify the importance of each parameter slice in the cross-domain stable causal feature space.

[0033] S102 specifically includes: Step b1: Encode the cross-domain stable causal features into a fixed-dimensional causal invariant anchor vector.

[0034] The causal invariant anchor vector is a 256-dimensional, numerically normalized vector that encodes stable core causal relationships extracted from multimodal data, unaffected by external factors such as device or environment. It serves as a unified benchmark for heterogeneous data alignment and causal analysis, ensuring consistency and fairness in comparisons and analyses across data from different sources.

[0035] Step b2: Using the causal invariant anchor vector as a reference, map the features of the local model parameter slices to the causal invariant subspace.

[0036] The causal invariant subspace is an abstract mathematical space spanned by the anchor vectors of the causal invariants. Mapping data or models from different modalities to this space eliminates the differences in their original representations, retaining only information about their core causal structure, thus enabling fair comparison and analysis. This application utilizes this subspace for causal consistency verification to prevent model iterative degradation.

[0037] Step b3: Calculate the correlation strength between the representation of each local model parameter slice in the causal invariant subspace and the causal invariant anchor vector to generate the causal sensitivity information.

[0038] The correlation strength is measured using cosine similarity. A stronger correlation indicates a higher contribution of this parameter segment to the core causal relationship, and more privacy budget should be allocated to it subsequently.

[0039] S103, construct a privacy budget topology based on the causal sensitivity information and allocate the budget, and generate a topology mask by combining multi-dimensional constraints to select effective model parameter slices from the local model parameter slices.

[0040] The privacy budget topology refers to a multi-level directed acyclic graph structure used to allocate and manage the global privacy budget (a quantitative indicator for measuring privacy leakage risk) in a refined and differentiated manner according to dimensions such as organization, data domain, parameter importance, and iteration rounds. Specifically, this structure includes four dimensions: organization level, data domain level, parameter sharding level, and iteration round level. The budget allocation refers to distributing the preset global privacy budget from top to bottom along the privacy budget topology, and dynamically adjusting the budget ratio of each topology path based on the causal sensitivity information of each parameter shard.

[0041] The multi-dimensional constraints include three dimensions: causal contribution, budget compliance, and historical utility.

[0042] The topology mask is a binary (0 or 1) vector used to selectively mask or pass through certain fragments from all local model parameter fragments. A value of 1 indicates that the fragment is selected to participate in this round of aggregation, and a value of 0 indicates that it is discarded.

[0043] The effective model parameter slices refer to parameter slices whose global contribution is greater than or equal to the dynamic filtering threshold.

[0044] S103 specifically includes: Step c1: Construct a four-dimensional privacy budget topology that includes an organizational hierarchy, a data domain hierarchy, a parameter sharding hierarchy, and an iteration round hierarchy.

[0045] This topology breaks down macro-level privacy protection requirements layer by layer to micro-level computing units, providing a structured framework for differentiated budget allocation.

[0046] Step c2: Distribute the preset global privacy budget (which can be set to 8.0 in this embodiment) from top to bottom along the privacy budget topology, and adjust the privacy budget allocation ratio of the corresponding path in the privacy budget topology according to the causal sensitivity information of each local model parameter slice.

[0047] Among them, the parameter slice with higher causal sensitivity receives a higher proportion of the budget for its corresponding path. Based on the allocated privacy budget, the noise intensity coefficient of each parameter slice is calculated (the noise intensity is inversely proportional to the privacy budget value), thereby realizing the differentiated allocation of privacy resources.

[0048] Step c3: For each local model parameter segment, calculate the causal contribution score, budget compliance score, and historical utility score for that local model parameter segment.

[0049] Among them, the causal contribution score directly reuses the causal sensitivity information, reflecting the theoretical importance of the parameter; the budget compliance score is calculated based on the remaining privacy budget of the slice, the maximum allocable privacy budget limit for a single parameter slice, and the noise intensity coefficient, assessing the privacy cost required to use the parameter; the historical utility score is calculated based on the accuracy improvement and convergence stability of the slice over the past multiple rounds (e.g., 10 rounds), representing the actual performance of the parameter in past iterations.

[0050] Step c4 involves weighting and fusing the causal contribution score, the budget compliance score, and the historical utility score to obtain the global contribution of each local model parameter segment.

[0051] Step c5: Calculate a dynamic filtering threshold based on the statistical distribution of the global contribution of all local model parameter slices in this round.

[0052] The screening threshold can be set as the mean of the global contribution plus 0.5 times the standard deviation. Using a dynamic threshold adapts to the overall level of parameter contribution in each round, ensuring relative fairness and effectiveness in the screening process.

[0053] Step c6: Compare the global contribution of each local model parameter fragment with the filtering threshold to generate the topology mask, so as to filter out the effective model parameter fragments from the local model parameter fragments.

[0054] In this method, parameter fragments with a global contribution greater than or equal to the screening threshold are assigned a 1 (valid fragment) in the corresponding position of the mask, and otherwise assigned a 0. By applying a topological mask, high-value parameters are accurately selected for aggregation, while low-contribution invalid parameters are removed, thereby reducing computational overhead and minimizing unnecessary privacy noise accumulation. Simultaneously, the mask sparsity (i.e., the proportion of parameters removed) is verified to be within the compliant range of 30% to 60%.

[0055] S104, perform differentiated privacy enhancement processing and verifiable weighted aggregation on the effective model parameter shards to generate a global model.

[0056] The differentiated privacy enhancement processing refers to using different noise perturbation strategies for parameter fragmentation based on different source modalities.

[0057] The verifiable weighted aggregation refers to performing a weighted summation of each valid fragment in a trusted execution environment, and verifying the integrity and correctness of the aggregation process through a homomorphic commitment chain.

[0058] Among them, homomorphic commitment chain is a cryptographic technique that links the commitment values ​​(a hash value that can verify data integrity but cannot reverse the original text) of each parameter fragment in a specific order. It is used to verifiably ensure that all parameters are not tampered with and are used correctly during the aggregation process without exposing the parameter content, thus providing auditability and security guarantees for the aggregation process.

[0059] S104 specifically includes: Step d1 involves performing differential privacy enhancement processing on the effective model parameter slices, including one or a combination of the following three methods: applying Laplacian noise to the local model parameter slices derived from image health data; performing precision truncation perturbation on the local model parameter slices derived from gene health data; and applying Gaussian noise to the local model parameter slices derived from wearable health data.

[0060] Among them, image data features are continuous and spatially correlated, and Laplacian noise can preserve image edge information; gene data is discrete and highly dimensional and sparse, and precision truncation perturbation can protect privacy without introducing external random noise; wearable time series data has autocorrelation, and Gaussian noise can be smoothly incorporated and retain data trends.

[0061] Step d2 involves performing verifiable weighted aggregation on the effective model parameters after enhancement processing to generate a global model.

[0062] During aggregation, the weights of each valid fragment satisfy the normalization condition (the sum of all weights is 1). Aggregation is performed in a trusted execution environment, and the consistency of the aggregation process is verified through a homomorphic commitment chain to ensure that all parameters have not been tampered with and are used correctly. The final output is the global model parameters.

[0063] S105, based on a predefined causal invariant subspace, performs causal consistency verification on the global model and the local models of each participant, and performs closed-loop write-back based on the verification results to form feedback data.

[0064] The causal invariant subspace is an abstract mathematical space spanned by causal invariant anchor vectors.

[0065] The causal consistency check refers to calculating the structural deviation between the global model and the local model in the subspace and comparing the deviation with a preset threshold.

[0066] The closed-loop writeback refers to writing back the global model update parameters that have passed the consistency check to the local models of each participant.

[0067] S105 specifically includes: Step e1: Map the global model and the local model to the causal invariant subspace to generate a global causal representation vector and a local causal representation vector, respectively.

[0068] This mapping places models from different sources in the same coordinate system, making their causal structures comparable.

[0069] Step e2: Calculate the structural deviation between the global causal representation vector and the local causal representation vector.

[0070] The structural deviation can be calculated using vector distance or similarity metrics (e.g., Euclidean distance, Manhattan distance, cosine similarity). In this embodiment, the mean deviation distance algorithm is used to quantify the degree to which the global model update deviates from the inherent causal laws of the local model.

[0071] Step e3: The structural deviation is compared with a preset deviation threshold (set to 0.08 in this embodiment) to determine whether the update of the global model maintains causal consistency.

[0072] If the structural deviation is less than the threshold, the global model update is considered to have maintained causal consistency and is a healthy update, and a closed-loop write-back is performed; otherwise, it is considered that the update may introduce false associations or destroy the original structure, and should be rejected or trigger re-aggregation (i.e., re-execute aggregation).

[0073] Step e4: Perform closed-loop write-back based on the verification results to generate feedback data.

[0074] Specifically, for global models that pass the consistency check, a hierarchical asynchronous write-back is performed, sending update parameters to each participant to update their local models, and the receipts are collected into iterative feedback data; for global models that fail the check, the aggregation results of this round are discarded, and no write-back is performed.

[0075] S106, Based on the feedback data, the allocation rules in the privacy budget topology are incrementally evolved, the edge heterogeneous operators are hot-swapped and deployed, and the full-process operation credentials are solidified in the form of a graph.

[0076] The incremental evolution refers to locating abnormal topological edges and adjusting allocation rules based on feedback data; the hot-swappable deployment refers to dynamically loading or replacing edge operators without interrupting system services; the full-process operation credentials include nine core credentials: data access, causal extraction, budget allocation, mask filtering, privacy enhancement, aggregation verification, consistency verification, rule evolution, and on-chain solidification, which are stored in the blockchain in the form of a knowledge graph.

[0077] S106 specifically includes: Step f1: Based on the feedback data, locate the abnormal topological edges in the privacy budget topology.

[0078] The abnormal topological edge refers to a topological path that continuously generates high deviation or low contribution parameters, which is identified by analyzing the iteration deviation, budget consumption fluctuation and causal verification failure rate in the feedback data.

[0079] Step f2: Quantify the abnormal score of the abnormal topological edge, and determine whether to trigger incremental evolution based on the preset abnormal threshold.

[0080] The anomaly score is calculated by weighting and fusing iteration bias, budget consumption volatility, and causal verification failure rate. If the anomaly score is greater than a preset anomaly threshold (set to 0.12 in this embodiment), the incremental evolution process is triggered.

[0081] Step f3: Under the condition of triggering incremental evolution, the weights and budget ratios corresponding to the abnormal topology edges are optimized through a double-buffer hot update mechanism to achieve zero-downtime update of the allocation rules.

[0082] The dual-buffer mechanism maintains two sets of rule parameters, old and new, allowing for instantaneous switching without interrupting ongoing services, thus meeting the 24 / 7 continuous service requirements of medical scenarios. The optimization process can employ reinforcement learning or gradient descent to maximize model performance and minimize privacy budget consumption by adjusting the weights and budget allocations corresponding to anomalous topological edges.

[0083] Step f4: Hot-swap deployment of edge heterogeneous operators.

[0084] The system automatically segments heterogeneous operators and dynamically loads or replaces them based on the differences in computing power among different edge devices, enabling collaborative work and unified model updates among edge devices without interrupting services during deployment.

[0085] Step f5: Solidify the entire process operation voucher in the form of a graph.

[0086] Among them, the above nine types of core credentials are constructed into a knowledge graph and stored in the blockchain to ensure that all operation nodes, parameter changes, budget consumption and other behaviors are traceable, verifiable and auditable, meeting the strict compliance and regulatory requirements for medical and health data.

[0087] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0088] In a specific embodiment of this application, the technical solution of this application is described using a scenario of distributed collaborative aggregation of health data from a multi-institutional medical consortium. The standardized medical and health data processed in this embodiment includes, but is not limited to, lung CT images, tumor gene mutation detection data, and wearable time-series data of human dynamic heart rate. All data have undergone pre-processing for anonymization and compliance. In this embodiment, all core quantitative parameters are set to globally unified fixed values ​​to ensure reproducibility. For example, the global total privacy budget is set to 8.0, the dimension of the causal invariant anchor vector is 256, the mask sparsity compliance range is 30% to 60%, the causal consistency deviation threshold is 0.08, the topological anomaly quantification threshold is 0.12, and the global contribution weights w1 (causal contribution) are equal to 0.5, w2 (budget compliance) are equal to 0.3, and w3 (historical utility) are equal to 0.2.

[0089] First, a cross-modal causal invariant representation distillation process is performed. For three types of data—CT images, gene sequences, and heart rate time series—local causal graphs for each unimodal data type are constructed, and spurious associations introduced by factors such as device model and ambient lighting are removed. Then, common cross-modal causal invariants representing inherent health patterns are extracted from these three data types and encoded into a 256-dimensional, numerically normalized causal invariant anchor vector. Using this anchor vector as a reference, features from all local model parameter slices are uniformly mapped to this 256-dimensional causal invariant subspace. Finally, by calculating the cosine similarity between the feature vectors of each parameter slice in the subspace and the anchor vector, a unique causal sensitivity score is generated and assigned.

[0090] Next, based on the above causal sensitivity scores, a four-dimensional hierarchical privacy budget topology diagram is constructed (e.g., Figure 3(As shown). The four levels of this topology are, in order, global organization nodes, subdivided data domain nodes, parameter sharding nodes, and iterative aggregation round nodes. The global privacy budget (total value of 8.0) is allocated from top to bottom. During the allocation process, the topology path corresponding to the parameter shard with a higher causal sensitivity score will receive a higher proportion of the privacy budget. Based on the allocated privacy budget ε, the noise intensity coefficient σ of each parameter shard is calculated using the formula:

[0091] Where σ is the parameter sharding noise intensity coefficient; ε is the privacy budget value allocated to the parameter sharding node.

[0092] Then, the topology mask is generated. For each parameter slice, its three core evaluation scores are calculated. The causal contribution score S directly reuses the causal sensitivity score. The budget compliance score B is calculated as follows: first, the remaining privacy budget for the slice is calculated. Then substitute it into the formula:

[0093] in, Remaining privacy budget for sharding; A privacy budget has been allocated to the current parameter shard node; Privacy budget has been used up for this iteration.

[0094]

[0095] in, For budget compliance score; Remaining privacy budget for sharding; The maximum allocatable privacy budget limit for a single-parameter shard; This is the noise penalty factor; This corresponds to the noise intensity coefficient of the segment. The historical utility score U is calculated based on the accuracy improvement and convergence stability of the segment over the past 10 rounds.

[0096] Subsequently, the global contribution G is calculated using a weighted fusion formula:

[0097] Where G represents the global contribution of each local model parameter segment; w1 equals 0.5; S represents the causal contribution score; w2 equals 0.3; B represents the budget compliance score; w3 equals 0.2; and U represents the historical utility score. Next, the mean μ and standard deviation σ of the global contributions of all parameter segments are calculated, and a screening threshold T is generated.

[0098] Where μ is the mean global contribution; This represents the global contribution value of all local model parameter fragments in this round; n is the total number of local model parameter fragments participating in the calculation in this round.

[0099]

[0100] in, The standard deviation of the global contribution; The global contribution value for all local model parameter slices in this round; is the global average contribution; n is the total number of local model parameter slices participating in the calculation in this round.

[0101]

[0102] Where T is the topology mask selection threshold for this round; μ is the global average contribution. The standard deviation of the global contribution is given. Parameter fragments with a global contribution G greater than or equal to T are marked as valid fragments and a binary topological mask is generated. Simultaneously, the mask sparsity is verified to be within the compliance range of 30% to 60% using the following formula:

[0103] in, For mask sparsity; This represents the total number of local model parameter shards participating in this aggregation round. Let be an element of a one-dimensional binary mask array M. If the i-th parameter slice is a valid slice, then... If it is an invalid fragment, then .

[0104] Subsequently, valid slices are selected based on topological masks, and homomorphic commitment chains are constructed for these slices in a trusted execution environment. Next, Laplacian noise with a scale parameter of 0.012 is applied to image slices, 6-digit truncation is performed on gene slices, and Gaussian noise with a mean of 0 and a variance of 0.008 is applied to wearable slices. Then, while ensuring the aggregation weights are normalized, weighted aggregation calculations are performed to generate the initial global model parameter draft θ_init.

[0105] in, This represents the number of valid fragments after mask filtering. The topology aggregation weights are for all valid shards.

[0106]

[0107] in, This is a draft of the global model parameters; This represents the number of valid fragments after mask filtering. For topological aggregation weights; These are compliant and valid fragmentation parameters after specific privacy enhancement, noise perturbation, and filtering calibration processes for image, genetic, and wearable applications. Finally, the consistency of the aggregation process is verified through a homomorphic commitment chain, and the final global model parameters are output.

[0108] Subsequently, the global model parameters and local model parameters are uniformly mapped to a 256-dimensional causal invariant subspace to generate a global causal representation vector. and local causal representation vector Then, the structural deviation D is calculated using the mean-deviation distance algorithm:

[0109] Where D is the structural bias; and N is the fixed dimension of the causal invariant subspace, with a value of 256. Global causal representation vector The i-th element; Local causal representation vector The i-th element. If the calculated D is less than the preset threshold of 0.08, then the causal consistency is deemed compliant, and a hierarchical asynchronous write-back is performed, and the receipts are aggregated into iterative feedback data; otherwise, the aggregation result of this round is discarded.

[0110] Finally, based on the feedback data, the topological edges in the privacy budget topology graph are evaluated. The anomaly quantification score E for the topological edges is calculated by weighted fusion iteration bias, budget consumption fluctuation, and causality verification failure rate.

[0111] Where E is the topological edge anomaly quantization score; For fixed weights; This represents the iteration deviation value; For fixed weights; Consumption volatility for budget purposes; For fixed weights; Let be the failure rate of causal consistency. And satisfy:

[0112] Where α, β, and γ are fixed weights. If E exceeds the anomaly threshold of 0.12, incremental evolution is triggered, and the weights and budget allocations of the topological edge are optimized through a double-buffer hot-update mechanism. Simultaneously, the system automatically segments heterogeneous operators and hot-swaps them to edge hardware. The credentials generated from all the above steps are used to construct a knowledge graph and solidify it into the blockchain, completing the closed loop of the entire process. This complete process can undergo multiple rounds of automated iteration to achieve continuous model optimization.

[0113] Please see Figure 2 This application also provides a health data model optimization system, including: The causal characterization module is used to perform cross-modal causal invariant characterization distillation on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features, and, based on the cross-domain stable causal features, generate causal sensitivity information bound to the parameter slices of each local model. The privacy topology management module is used to construct a privacy budget topology structure and allocate the budget based on the causal sensitivity information, and generate a topology mask by combining multi-dimensional constraints, so as to filter out effective model parameter fragments from the local model parameter fragments. The secure aggregation module is used to perform differentiated privacy enhancement processing and verifiable weighted aggregation on the effective model parameters to generate a global model; The consistency verification module is used to perform causal consistency verification between the global model and the local models of each participant based on a predefined causal invariant subspace, and to perform closed-loop write-back based on the verification results to form feedback data. The rule evolution module is used to incrementally evolve the allocation rules in the privacy budget topology based on the feedback data, and to hot-swap deploy edge heterogeneous operators. The voucher management module is used to solidify all operation vouchers in the form of a graph.

[0114] It should be noted that the information interaction and execution process between the above-mentioned components are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0115] This application also provides a computer device, which includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in the above-described method embodiments.

[0116] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0117] This application provides a computer program product that, when run on a computer device, enables the computer device to execute the steps described in the various method embodiments above.

[0118] An embodiment of this application provides a computer device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in the various health data model optimization method embodiments described above.

[0119] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0120] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for optimizing a health data model, characterized in that, include: Cross-modal causal invariant characterization distillation was performed on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features; Based on the aforementioned cross-domain stable causal characteristics, causal sensitivity information is generated that is bound to the parameter slices of each local model; A privacy budget topology is constructed based on the causal sensitivity information and a budget is allocated. A topology mask is generated by combining multi-dimensional constraints to select effective model parameter slices from the local model parameter slices. Differential privacy enhancement processing and verifiable weighted aggregation are performed on the effective model parameters in segments to generate a global model; Based on a predefined causal invariant subspace, the global model and the local models of each participant are checked for causal consistency, and closed-loop write-back is performed according to the check results to form feedback data. Based on the feedback data, the allocation rules in the privacy budget topology are incrementally evolved, edge heterogeneous operators are hot-swapped and deployed, and the full-process operation credentials are solidified in the form of a graph.

2. The method according to claim 1, characterized in that, The process involves performing cross-modal causal invariant characterization distillation on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features. Based on these cross-domain stable causal features, causal sensitivity information bound to the parameter slices of each local model is generated, including: For different data modalities in the multi-source heterogeneous health data, local causal graphs are constructed for each, and pseudo-association features introduced by non-essential factors are removed. Extract cross-modal common causal invariants from the local causal graph, and use the common causal invariants as the cross-domain stable causal features; The cross-domain stable causal features are encoded as causal invariant anchor vectors with fixed dimensions; Based on the causal invariant anchor vector, the features of the local model parameter slices are mapped to the causal invariant subspace; The correlation strength between the representation of each local model parameter slice in the causal invariant subspace and the anchor vector of the causal invariant is calculated to generate the causal sensitivity information.

3. The method according to claim 1, characterized in that, The process of constructing a privacy budget topology and allocating the budget based on the causal sensitivity information includes: Construct a four-dimensional privacy budget topology that includes institutional hierarchy, data domain hierarchy, parameter sharding hierarchy, and iteration round hierarchy; The preset global privacy budget is allocated from top to bottom along the privacy budget topology, and the privacy budget allocation ratio of the corresponding path in the privacy budget topology is adjusted according to the causal sensitivity information of each local model parameter slice.

4. The method according to claim 1, characterized in that, The step of generating a topology mask by combining multi-dimensional constraints to filter out effective model parameter pieces from the local model parameter pieces includes: For each local model parameter slice, calculate the causal contribution score, budget compliance score, and historical utility score for that local model parameter slice. The causal contribution score, the budget compliance score, and the historical utility score are weighted and fused to obtain the global contribution of each local model parameter slice; Based on the statistical distribution of the global contribution of all local model parameter slices in this round, a screening threshold is calculated. The global contribution of each local model parameter fragment is compared with the filtering threshold to generate the topology mask, so as to filter out the effective model parameter fragments from the local model parameter fragments.

5. The method according to claim 1, characterized in that, The differential privacy enhancement processing performed on the effective model parameter slices includes one or a combination of the following three methods: Laplacian noise is applied to the local model parameter slices derived from image health data; For local model parameter slices derived from genetic health data, precision truncation perturbation is performed. Gaussian noise is applied to the local model parameter slices derived from wearable health data.

6. The method according to claim 1, characterized in that, The step of performing causal consistency verification between the global model and the local models of each participant, based on a predefined causal invariant subspace, includes: The global model and the local model are uniformly mapped to the causal invariant subspace, generating a global causal representation vector and a local causal representation vector, respectively. Calculate the structural deviation between the global causal representation vector and the local causal representation vector; The structural deviation is compared with a preset deviation threshold to determine whether the update of the global model maintains causal consistency.

7. The method according to claim 1, characterized in that, The incremental evolution of the allocation rules in the privacy budget topology includes: Based on the feedback data, locate the abnormal topological edges in the privacy budget topology; The abnormal scores of the abnormal topological edges are quantified, and incremental evolution is determined based on a preset abnormal threshold. Under the condition of triggering incremental evolution, the weights and budget ratios corresponding to the abnormal topology edges are optimized through a double-buffer hot update mechanism to achieve zero-downtime update of the allocation rules.

8. A health data model optimization system, characterized in that, include: The causal characterization module is used to perform cross-modal causal invariant characterization distillation on the acquired multi-source heterogeneous health data to extract cross-domain stable causal features, and, based on the cross-domain stable causal features, generate causal sensitivity information bound to the parameter slices of each local model. The privacy topology management module is used to construct a privacy budget topology structure and allocate the budget based on the causal sensitivity information, and generate a topology mask by combining multi-dimensional constraints, so as to filter out effective model parameter fragments from the local model parameter fragments. The secure aggregation module is used to perform differentiated privacy enhancement processing and verifiable weighted aggregation on the effective model parameters to generate a global model; The consistency verification module is used to perform causal consistency verification between the global model and the local models of each participant based on a predefined causal invariant subspace, and to perform closed-loop write-back based on the verification results to form feedback data. The rule evolution module is used to incrementally evolve the allocation rules in the privacy budget topology based on the feedback data, and to hot-swap deploy edge heterogeneous operators. The voucher management module is used to solidify all operation vouchers in the form of a graph.

9. A computer program product, characterized in that, Includes a computer program that, when run, implements the method as described in any one of claims 1 to 7.

Images