A cross-hospital medical data privacy protection sharing method and system

Abstract

The application discloses a cross-hospital medical data privacy protection sharing method and system, relates to the technical field of data processing, and comprises the following steps: a six-level progressive processing link is constructed, first, multi-source heterogeneous records are converted into standardized semantic representations and an initial attribute space is established, then multi-dimensional associated privacy risks are quantified to adaptively determine generalization levels and noise intensities, cross-hospital local alignment is performed to eliminate inter-hospital heterogeneity, and finally, a safe aggregation result is output through a limited query response layer, so that the privacy protection strength dynamically adapts to the multi-dimensional associated risks, the multi-dimensional attribute combination associated inference attack is effectively resisted, the stability and cross-hospital consistency of the fused data on the global statistical characteristics are ensured, and the collaborative guarantee of the privacy protection capability and the analysis availability is realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for cross-hospital medical data privacy protection and sharing. Background Technology

[0002] In cross-institutional collaborative analysis and joint research scenarios of medical data, to achieve compliant sharing of medical record data across multiple institutions, existing technologies generally employ desensitization techniques such as k-anonymity, l-diversity, or differential privacy to process sensitive fields in electronic medical records. These solutions typically treat each quasi-identifier attribute as an independent object, applying generalization, suppression, or noise addition separately. Their design is based on the assumption that the discriminative power of a single attribute can represent the overall privacy risk. The processing flow is static and fragmented, lacking systematic modeling and dynamic response capabilities for the combined effects of attributes.

[0003] However, real medical data exhibits strong statistical dependencies among its multidimensional attributes, such as demographic characteristics, diagnostic codes, laboratory indicators, and consultation times. Multiple low-sensitivity attributes, when combined, can form highly discriminative joint fingerprints. This allows attackers to leverage publicly available auxiliary information to significantly narrow down candidate records or even uniquely identify an individual through cross-matching. Existing anonymization methods fail to quantify the risks of multidimensional associations, cannot adaptively adjust protection strength based on combined discriminative power, and lack cross-hospital statistical consistency calibration mechanisms after parallel processing in multiple hospitals. Consequently, they either provide insufficient protection, exposing vulnerabilities in association inference, or excessively perturb data, compromising data usability. This makes it difficult to balance privacy and analytical utility. Summary of the Invention

[0004] This invention aims to provide a method and system for sharing cross-hospital medical data with privacy protection, which effectively resists attacks that combine multi-dimensional attributes for association and inference, and ensures the stability and cross-hospital consistency of fused data in terms of global statistical characteristics, thus achieving synergistic protection of privacy and usability of analysis.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for cross-hospital medical data privacy protection and sharing, comprising:

[0006] Standardize and semantically align multi-source heterogeneous medical records to construct an initial attribute space vector;

[0007] A multidimensional correlation privacy risk index is calculated based on the initial attribute space vector, and the global disturbance intensity coefficient is determined accordingly.

[0008] Based on the global perturbation intensity coefficient, an adaptive attribute generalization transformation is performed on the initial attribute space vector to generate a confirmed and valid set of generalized attributes and metadata describing the generalization transformation.

[0009] The generalization transformation is used to adjust the noise parameters of the metadata description, and controlled noise is injected into the confirmed effective generalized attribute set to generate a noisy attribute set that has been statistically verified.

[0010] Based on the statistically validated set of noisy attributes, cross-departmental local alignment and consistency verification are performed to generate a cross-departmental fusion dataset and data quality and privacy protection credentials.

[0011] Based on the cross-institute fusion dataset and the data quality and privacy protection certificate, a shared output package is generated through a restricted query response.

[0012] Preferably, standardized semantic alignment is performed on multi-source heterogeneous medical records to construct an initial attribute space vector, including:

[0013] Parse the original electronic medical record dataset to extract all field values ​​and form the original attribute set;

[0014] The cross-institutional general medical terminology ontology is loaded, and semantic matching is performed on the original attribute set to convert it into a semantically standardized intermediate attribute set.

[0015] Traverse the semantically standardized intermediate attribute set to count the frequency and value distribution of each field, and establish a global attribute dimension index table;

[0016] The attribute fields that meet the preset conditions are selected based on the global attribute dimension index table to form an initial attribute space vector.

[0017] Preferably, a multidimensional correlation privacy risk index is calculated based on the initial attribute space vector, and a global perturbation strength coefficient is determined accordingly, including:

[0018] Select attribute combinations from the initial attribute space vector and calculate the joint value frequency distribution to form a multidimensional correlation frequency matrix;

[0019] The comprehensive equivalence class size of each record is calculated based on the multidimensional association frequency matrix, and a risk indication sequence is generated.

[0020] A multidimensional associated privacy risk index is calculated using the aforementioned risk indication sequence and security equivalence class threshold;

[0021] The global disturbance intensity coefficient is determined based on the multidimensional associated privacy risk index and the preset risk classification mapping rule.

[0022] Preferably, an adaptive attribute generalization transformation is performed on the initial attribute space vector based on the global perturbation intensity coefficient to generate a confirmed and valid set of generalized attributes and metadata describing the generalization transformation, including:

[0023] The target generalization level is determined for each attribute in the initial attribute space vector by combining the global perturbation intensity coefficient, thus forming a hierarchical configuration vector;

[0024] Based on the hierarchical configuration vector, a generalized mapping is performed on the semantically standardized intermediate attribute set to generate a generalized attribute set;

[0025] The generalized multidimensional association privacy risk index of the generalized attribute set is recalculated, and the generalized attribute set is confirmed to be valid after being judged to be qualified.

[0026] Extract the final effective hierarchical configuration vector and package it with the global attribute dimension index table and the multidimensional associated privacy risk index to generate generalized transformation description metadata.

[0027] Preferably, the noise parameters are adjusted using the generalization transformation to describe the metadata, and controlled noise is injected into the confirmed effective generalized attribute set to generate a statistically validated noisy attribute set, including:

[0028] The risk difference is parsed from the generalized transformation description metadata, and the noise scaling factor is calculated.

[0029] The noise standard deviation of each attribute is determined based on the noise scaling factor and the interquartile range of each attribute, thus forming a noise parameter set;

[0030] According to the noise parameter set, controlled noise is injected into the confirmed effective generalized attribute set one by one to generate a noisy attribute set;

[0031] The deviation of the statistical moments between the noisy attribute set and the data before generalization is verified. After the deviation is deemed acceptable, the noisy attribute set and the final noise parameter set that have passed statistical verification are obtained.

[0032] Preferably, based on the statistically validated set of noisy attributes, cross-departmental local alignment and consistency verification are performed to generate a cross-departmental fused dataset and data quality and privacy protection credentials, including:

[0033] Extract the sample mean of the noisy attribute set that has passed statistical verification from each institute, and calculate the inter-institute calibration offset to form a calibration matrix;

[0034] Based on the calibration matrix, mean calibration and classification label unification are performed on the data of each institute to generate a locally aligned set of noisy attributes;

[0035] Perform a consistency check between the noise variance and the associated structure on the noisy attribute set after local alignment of each institute, and generate a consistency check report;

[0036] The cross-institute fusion dataset is generated by splicing data from each institute based on the consistency verification report, and the final noise parameter set, generalization transformation description metadata, and consistency verification report are packaged into a data quality and privacy protection certificate.

[0037] Preferably, based on the cross-institute fusion dataset and the data quality and privacy protection certificate, a shared output package is generated through a restricted query response, including:

[0038] Load the data quality and privacy protection credentials to initialize the query parser and convert the external aggregate query into an internal query plan;

[0039] Based on the internal query plan, an aggregation operation is performed on the cross-institute fusion dataset, and output noise is added according to the size of the aggregation group to generate preliminary query results;

[0040] The validity of the minimum publication threshold and confidence interval width is validated on the preliminary query results to obtain safe query results;

[0041] The secure query results, along with the metadata summary and usage instructions of the data quality and privacy protection certificate, are encapsulated into a shared output package.

[0042] Preferably, the generalized multidimensional association privacy risk index of the generalized attribute set is recalculated, and a valid generalized attribute set is obtained after passing the judgment, including:

[0043] Recalculate the new comprehensive equivalence class size for each record on the generalized attribute set;

[0044] The generalized multidimensional correlation privacy risk index is calculated using the size of the new comprehensive equivalence class and the threshold of the security equivalence class.

[0045] When the generalized multidimensional association privacy risk index is higher than the security threshold, the hierarchical configuration vector is adjusted upward and the generalization mapping is re-executed.

[0046] When the generalized multidimensional association privacy risk index is not higher than the security threshold, the current generalization result is used as the confirmed valid set of generalized attributes.

[0047] Preferably, a consistency check is performed on the noise variance and associated structure of the noisy attribute set after local alignment of each institute, and a consistency check report is generated, including:

[0048] Check the deviation between the actual noise variance of each institute on the same attribute and the theoretical variance declared in the final noise parameter set;

[0049] Calculate the distance metric for the joint distribution of each college under the same attribute combination;

[0050] When the deviation or distance metric exceeds the preset tolerance, the corresponding mismatch item is marked and the severity is recorded.

[0051] Generate a consistency verification report by summarizing all mismatches and their severity.

[0052] On the other hand, this invention proposes a cross-hospital medical data privacy protection and sharing system, comprising:

[0053] Semantic alignment unit is used to standardize semantic alignment of multi-source heterogeneous medical records and construct an initial attribute space vector.

[0054] The risk quantification unit is used to calculate a multidimensional associated privacy risk index based on the initial attribute space vector, and thereby determine the global disturbance intensity coefficient.

[0055] The generalization transformation unit is used to perform adaptive attribute generalization transformation on the initial attribute space vector according to the global perturbation intensity coefficient, and generate a confirmed valid set of generalized attributes and metadata describing the generalization transformation.

[0056] The noise injection unit is used to adjust the noise parameters of the metadata description using the generalization transformation, inject controlled noise into the confirmed effective generalized attribute set, and generate a noisy attribute set that has been statistically verified to be qualified.

[0057] The alignment verification unit is used to perform cross-departmental local alignment and consistency verification based on the statistically verified set of noisy attributes, and generate cross-departmental fusion dataset and data quality and privacy protection certificate;

[0058] The secure output unit is used to generate a shared output package based on the cross-institute fusion dataset and the data quality and privacy protection certificate through a restricted query response.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] This invention constructs a six-level progressive processing chain. First, it transforms multi-source heterogeneous records into standardized semantic representations and establishes an initial attribute space. Then, it quantifies multi-dimensional association privacy risks to adaptively determine the generalization level and noise intensity. Next, it eliminates inter-institutional heterogeneity through cross-institutional local alignment. Finally, it outputs secure aggregation results through a restricted query response layer. This allows the strength of privacy protection to dynamically adapt to multi-dimensional association risks, effectively resisting multi-dimensional attribute combination association inference attacks while ensuring the stability of fused data in global statistical characteristics and cross-institutional consistency. This achieves synergistic protection of privacy capabilities and analytical usability. Attached Figure Description

[0061] Figure 1 This is a flowchart of the cross-hospital medical data privacy protection and sharing method of the present invention;

[0062] Figure 2 This is a block diagram of the cross-hospital medical data privacy protection and sharing system of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0064] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0065] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0066] Please see Figure 1 This invention provides a method for cross-hospital medical data privacy protection and sharing, comprising:

[0067] Standardize and semantically align multi-source heterogeneous medical records to construct an initial attribute space vector; specifically including:

[0068] The original electronic medical record dataset is parsed to extract all field values ​​to form an original attribute set; the cross-institutional general medical terminology ontology is loaded to perform semantic matching on the original attribute set, converting it into a semantically standardized intermediate attribute set; the semantically standardized intermediate attribute set is traversed to count the frequency and value distribution of each field, and a global attribute dimension index table is established; attribute fields that meet the preset conditions are selected based on the global attribute dimension index table to form an initial attribute space vector.

[0069] By uniformly mapping multi-source heterogeneous medical records to a cross-institutional general medical terminology ontology and filtering effective fields based on a global attribute dimension index table, semantic ambiguity and structural differences in field naming, coding systems, and value granularity among different hospitals are eliminated. This provides a semantically consistent and dimensionally regular data foundation for subsequent multi-dimensional correlation privacy risk quantification, avoiding misjudgment of generalization levels and mismatch of noise parameters caused by semantic misalignment.

[0070] A multidimensional correlation privacy risk index is calculated based on the initial attribute space vector, and the global perturbation strength coefficient is determined accordingly; specifically including:

[0071] Attribute combinations are selected from the initial attribute space vector, and the joint value frequency distribution is calculated to form a multidimensional association frequency matrix. The comprehensive equivalence class size of each record is calculated based on the multidimensional association frequency matrix to generate a risk indication sequence. The multidimensional association privacy risk index is calculated using the risk indication sequence and the security equivalence class threshold. The global disturbance intensity coefficient is determined according to the multidimensional association privacy risk index and the preset risk classification mapping rule.

[0072] By quantifying the joint distinguishing ability of attribute combinations and dynamically determining the perturbation intensity accordingly, the level of privacy protection is precisely matched with the actual multidimensional correlation risk level of the data. This avoids excessive perturbation in low-risk scenarios from damaging data usability, and prevents high-risk combinations from being cross-inferred and identified due to insufficient protection, thus achieving an adaptive balance between privacy security and analytical utility.

[0073] An adaptive attribute generalization transformation is performed on the initial attribute space vector based on the global perturbation intensity coefficient, generating a confirmed set of generalized attributes and metadata describing the generalization transformation; specifically including:

[0074] The target generalization level is determined for each attribute in the initial attribute space vector by combining the global perturbation intensity coefficient, forming a hierarchical configuration vector. Generalization mapping is performed on the semantically standardized intermediate attribute set based on the hierarchical configuration vector to generate a generalized attribute set. The generalized multidimensional association privacy risk index of the generalized attribute set is recalculated, and the confirmed valid generalized attribute set is obtained after passing the judgment. The final effective hierarchical configuration vector is extracted and packaged with the global attribute dimension index table and the multidimensional association privacy risk index to generate generalization transformation description metadata.

[0075] The process involves recalculating the generalized multidimensional association privacy risk index of the generalized attribute set. After passing the assessment, a valid generalized attribute set is obtained. This includes: recalculating the new comprehensive equivalence class size for each record in the generalized attribute set; calculating the generalized multidimensional association privacy risk index using the new comprehensive equivalence class size and the safety equivalence class threshold; adjusting the hierarchical configuration vector and re-executing the generalization mapping when the generalized multidimensional association privacy risk index is higher than the safety threshold; and using the current generalization result as the valid generalized attribute set when the generalized multidimensional association privacy risk index is not higher than the safety threshold.

[0076] By introducing a closed-loop verification and dynamic hierarchical adjustment mechanism for the risk index after generalization, the attribute generalization results are ensured to strictly meet the security equivalence class threshold requirements, avoiding the privacy protection issues or over-generalization problems that may be caused by single static generalization. At the same time, the complete transformation description metadata is retained to provide accurate parameter basis for subsequent noise injection, ensuring the certainty and traceability of privacy protection.

[0077] The method utilizes generalization transformation to describe the adjustment of noise parameters in metadata, injects controlled noise into the confirmed effective generalized attribute set, and generates a statistically validated noisy attribute set; specifically including:

[0078] The risk difference is parsed from the metadata description of the generalization transformation, and the noise scaling factor is calculated. Based on the noise scaling factor and the interquartile range of each attribute, the noise standard deviation of each attribute is determined, forming a noise parameter set. Controlled noise is injected into each of the confirmed effective generalized attribute sets according to the noise parameter set to generate a noisy attribute set. The statistical moment deviation between the noisy attribute set and the data before generalization is verified. After passing the verification, the statistically verified noisy attribute set and the final noise parameter set are obtained.

[0079] By adjusting the noise intensity in conjunction with the risk difference and data distribution characteristics during the generalization stage, and supplementing it with statistical moment deviation verification, the injected noise can retain the statistical characteristics of the original data to the maximum extent while meeting privacy protection requirements. This avoids the risk of utility loss or privacy leakage caused by fixed noise parameters, and ensures that the noisy data has reliable statistical validity in subsequent analysis.

[0080] Based on a statistically validated set of noisy attributes, cross-departmental local alignment and consistency checks are performed to generate a cross-departmental fusion dataset and data quality and privacy protection credentials; specifically including:

[0081] Extract the sample mean of the noisy attribute set that has passed statistical verification from each institute, and calculate the inter-institute calibration offset to form a calibration matrix; perform mean calibration and classification label unification on the data of each institute based on the calibration matrix to generate a locally aligned noisy attribute set; perform consistency verification of noise variance and association structure on the locally aligned noisy attribute set of each institute to generate a consistency verification report; splice the data of each institute based on the consistency verification report to generate a cross-institute fusion dataset, and package the final noise parameter set, generalization transformation description metadata, and consistency verification report into a data quality and privacy protection certificate.

[0082] Specifically, the process involves performing a consistency check on the noise variance and associated structure of the noisy attribute set after local alignment of each institute, generating a consistency check report. This includes: checking the deviation between the actual noise variance of each institute on the same attribute and the theoretical variance declared in the final noise parameter set; calculating the distance metric value of the joint distribution of each institute under the same attribute combination; marking the corresponding mismatch item and recording its severity when the deviation or distance metric value exceeds the preset tolerance; and summarizing all mismatch items and their severity to generate a consistency check report.

[0083] By calibrating the inter-institute mean and verifying the consistency of noise variance and correlation structure, statistical distortion and privacy protection gaps caused by distribution offset or privacy parameter execution deviations in the fusion process of multi-source noisy data are eliminated. This ensures that the cross-institute fused dataset maintains analytical usability while having verifiable privacy compliance, and provides auditable evidence of data quality and privacy protection in the form of structured credentials.

[0084] Based on the cross-institute integrated dataset and data quality and privacy protection credentials, a shared output package is generated through restricted query responses; specifically including:

[0085] The query parser is initialized by loading the data quality and privacy protection credentials, converting external aggregate queries into internal query plans. Aggregation operations are performed on the cross-institute fusion dataset according to the internal query plan, and output noise is added according to the size of the aggregation group to generate preliminary query results. The validity of the preliminary query results is verified by the minimum publication threshold and confidence interval width to obtain secure query results. The secure query results, along with the metadata summary and usage instructions of the data quality and privacy protection credentials, are encapsulated into a shared output package.

[0086] By dynamically transforming the query plan based on credentials and superimposing output noise during the query response phase, and combining the minimum publication threshold and confidence interval width for dual validity verification, the risk of small sample aggregation inference attacks and excessive information leakage is effectively blocked. This ensures that the shared output package provides statistically reliable analysis results while meeting privacy and security constraints, and at the same time, the encapsulated metadata digest ensures the compliance and interpretability of data use.

[0087] On the other hand, this invention proposes a cross-hospital medical data privacy protection and sharing system, such as... Figure 2 As shown, it includes:

[0088] Semantic alignment unit is used to standardize semantic alignment of multi-source heterogeneous medical records and construct an initial attribute space vector.

[0089] The risk quantification unit is used to calculate a multidimensional associated privacy risk index based on the initial attribute space vector, and thereby determine the global disturbance intensity coefficient.

[0090] The generalization transformation unit is used to perform adaptive attribute generalization transformation on the initial attribute space vector according to the global perturbation intensity coefficient, and generate a confirmed valid set of generalized attributes and metadata describing the generalization transformation.

[0091] The noise injection unit is used to adjust the noise parameters of the metadata description using the generalization transformation, inject controlled noise into the confirmed effective generalized attribute set, and generate a noisy attribute set that has been statistically verified to be qualified.

[0092] The alignment verification unit is used to perform cross-departmental local alignment and consistency verification based on the statistically verified set of noisy attributes, and generate cross-departmental fusion dataset and data quality and privacy protection certificate;

[0093] The secure output unit is used to generate a shared output package based on the cross-institute fusion dataset and the data quality and privacy protection certificate through a restricted query response.

[0094] Specifically, the aforementioned units, during execution, also implement other steps of the aforementioned method for sharing cross-hospital medical data with privacy protection, as follows:

[0095] Step 1: Standardized semantic alignment and initial attribute space construction of multi-source heterogeneous medical records;

[0096] In the initial stage of cross-institutional data exchange, the primary task is to transform raw medical records from different medical institutions, using different coding systems and storage structures, into standardized semantic representations that can be processed uniformly later. This step is not a simple field mapping, but rather the establishment of an initial attribute space covering the data characteristics of all participants, providing a consistent benchmark for subsequent privacy risk quantification. Specifically, this includes:

[0097] Step 1.1: Receive raw electronic medical record datasets from multiple medical institutions, and perform structured parsing on each record in each dataset to extract all field values, including patient demographic characteristics, diagnostic codes, laboratory indicators, medication records, and visit timestamps, forming a raw attribute set without any processing. The original set of attributes Each element in the database corresponds to a complete patient record. It retains the original field names, original encoded values, and original data types from the source database and has not undergone any form of normalization or semantic conversion.

[0098] Step 1.2: Based on the original attribute set obtained in Step 1.1 Load the pre-built cross-institutional general medical terminology ontology and perform semantic matching operations on each field value of each record in the set.

[0099] Specifically, for diagnostic coding fields, they are mapped to the International Classification of Diseases (ICD) standard code; for drug names, they are mapped to the National Reimbursement Drug List (NRDL) standard code; for laboratory test items, they are mapped to the Uniform Identifier for Clinical Laboratory Tests (UID); and for unstructured text descriptions in demographic features, they are converted into enumerated standard values ​​using a predefined rule table. After the above mapping operations, the original attribute set... Transformed into a semantically standardized set of intermediate attributes This intermediate attribute set All fields are expressed using a unified terminology system, eliminating synonyms or homonyms caused by differences in coding habits among different departments.

[0100] Step 1.3: Using the semantically normalized intermediate attribute set generated in Step 1.2 For the operation object, iterate through all fields of all records, count the frequency of each standard field in the entire dataset and its value distribution range, and build a global attribute dimension index table based on this. This index table It records the name, data type, cardinality, missing rate, and co-occurrence frequency with other fields for each standard attribute field.

[0101] Specifically, for continuous numerical fields, the minimum, maximum, and quantile points are also recorded; for discrete categorical fields, a list of all valid values ​​and the sample count for each value are recorded. This global attribute dimension index table This not only reflects the structural characteristics of the current dataset, but also provides a quantitative basis for subsequent identification of highly discriminative attributes.

[0102] Step 1.4: Based on the global attribute dimension index table established in Step 1.3 From this, all attribute fields with a cardinality greater than a preset minimum threshold and a missing rate lower than a preset upper limit are selected. These fields are then arranged in order of their position in the index table to form an ordered initial attribute space vector. The initial attribute space vector This defines the scope of all subsequent privacy-processing operations, with each component corresponding to a valid attribute dimension for cross-departmental sharing. Fields not included in this vector (such as fields with excessively sparse values ​​or almost entirely empty fields) will be marked as inactive attributes and will not participate in any transformations or perturbations in subsequent processes, but will still be retained in the data record to maintain the integrity of the original record. This initial attribute space vector This serves as a benchmark framework for the quantitative assessment of privacy risks.

[0103] Step 2: Quantitative assessment of multidimensional correlation privacy risks based on the initial attribute space;

[0104] After obtaining the initial attribute space vector constructed in step one. The next step is to quantitatively characterize its inherent privacy leakage risk. Traditional methods often focus only on the identifiability of a single attribute, neglecting the high discriminative power that may arise from combinations of multiple low-sensitivity attributes. This step aims to calculate, from the perspective of combinatorial correlation, the probability that any record in the current attribute space can be uniquely locked or narrowed down by an external attacker using publicly available auxiliary information. Specifically, this includes:

[0105] Step 2.1: Using the initial attribute space vector output in Step 1 Given the input, select all pairwise and triadic attribute combinations, and calculate the joint value frequency distribution of each combination in the current dataset.

[0106] The specific steps are as follows: For a selected attribute combination, iterate through all data records, count the absolute frequency of each joint value, and divide by the total number of records to obtain a joint probability estimate. Summarize the joint probability estimates of all combinations to form a multidimensional association frequency matrix. The matrix The rows correspond to different attribute combinations, the columns correspond to the specific joint values ​​under each combination, and the cell values ​​are the empirical probabilities of that value. This is a multidimensional association frequency matrix. It reveals the implicit statistical dependencies between attributes and serves as a fundamental data carrier for measuring the discriminative power of combinations.

[0107] Step 2.2: Based on the multidimensional correlation frequency matrix obtained in Step 2.1 For the initial attribute space vector For each data record covered, calculate its equivalence class size under the current attribute combination. An equivalence class is a subset of all records that have exactly the same joint value for the selected attribute combination. For the i-th record, its equivalence class size under a certain attribute combination c is denoted as . This value can be directly obtained from the matrix. The frequency of the corresponding joint value is obtained by multiplying it by the total number of records.

[0108] Subsequently, the minimum equivalence class size is taken for all calculated attribute combinations, and this minimum value is used as the comprehensive equivalence class size for that record. The size of this comprehensive equivalence class This reflects the maximum population capacity within which the record can be hidden under the most unfavorable combination. The smaller the value, the easier the record is to distinguish. (The remaining text appears to be incomplete and possibly contains errors.) The values ​​constitute a risk indication sequence , where N is the total number of records.

[0109] Step 2.3: Using the risk indication sequence S generated in Step 2.2, calculate the multidimensional correlation privacy risk index of the entire dataset in the current initial attribute space according to the following formula. :

[0110]

[0111] In the formula, N represents the total number of records in the dataset; Indicates the size of the comprehensive equivalence class of the i-th record; The preset security equivalence class threshold is determined by the minimum anonymity guarantee level agreed upon in the data usage agreement. The calculation ensures that a record's equivalence class size is only counted as contributing to the risk index if it is below a security threshold; records with a size above or equal to the threshold have no positive contribution to the risk index. This multidimensional privacy risk index... The higher the value, the more records in the dataset are in a high-risk state and the more serious the deviation from the safety baseline.

[0112] Step 2.4: Calculate the multidimensional correlation privacy risk index based on Step 2.3. Based on the preset risk classification mapping rules, a global disturbance intensity coefficient is determined. The mapping rule is a piecewise linear relationship: when hour, ;when hour, Follow Linear growth; when hour, .in and This is the threshold for risk classification; and These represent the lower and upper limits of the disturbance intensity, respectively, and is a dimensionless scaling factor. The global disturbance intensity coefficient is determined accordingly. This approach avoids excessive disturbance to low-risk data, which could lead to a loss of usability, while ensuring that high-risk data receives adequate protection.

[0113] Step 3: Adaptive attribute generalization transformation based on global perturbation intensity coefficient;

[0114] Obtain the global disturbance intensity coefficient output in step two. and the initial attribute space vector output in step one Next, this step performs targeted generalization on the attribute space. Generalization does not apply the same level of fuzziness to all attributes, but rather selects attributes based on their properties. The size of the data is dynamically adjusted to change the abstraction level of each attribute, so that the transformed data retains as much granular information as possible that is valuable to downstream analysis tasks while meeting privacy requirements. Specifically, this includes:

[0115] Step 3.1: Using the initial attribute space vector output in Step 1 and the global disturbance intensity coefficient output in step two For common input, firstly... For each attribute field, its corresponding hierarchical generalization tree is loaded. This generalization tree predefines the multi-level generalization path from the original value to the coarsest-grained abstract value. For example, the age field can generalize from a specific age to a five-year-old age group, then to a ten-year-old age group, and finally to categories such as elderly / middle-aged / youth. For each attribute, its default starting level in the generalization tree is determined based on its data type and the number of times it participates in high-frequency association combinations in step two.

[0116] At the same time, the global disturbance intensity coefficient As an adjustment factor, the default starting level is offset and corrected: The larger the value, the more levels the starting level moves upwards; The smaller the value, the unchanged initial level or only slightly shifted upwards. After this operation, each attribute obtains a target generalization level that adapts to the current risk level. Where j is the attribute in The sequence number in the table. The target generalization hierarchy of all attributes constitutes a hierarchical configuration vector. d is The number of dimensions.

[0117] Step 3.2: Based on the hierarchical configuration vector L generated in Step 3.1, process the semantically standardized intermediate attribute set from Step 1. (Note: The object of this operation is) Instead ,because This is just a list of attribute names; the actual data is still stored in... (In Chinese) Generalization mapping is performed record by record and attribute by attribute. Specifically, for the j-th attribute, it is mapped according to its target generalization level. The corresponding generalization rules will All original standard values ​​of this attribute are replaced with the corresponding level of generalized tags.

[0118] For example, if a patient's age is The value in the middle is "37", while the target generalization level of its age attribute is... For the five-year age group, this value is replaced with "35-39". This mapping process strictly follows the parent-child inheritance relationship of the generalization tree, ensuring that the generalized value is a semantically valid superordinate concept of the original value. After full mapping, the intermediate attribute set... Transformed into a generalized set of attributes This set All attribute values ​​have been reduced to a level of abstraction commensurate with their risk levels, but no random perturbations have been introduced. This generalized set of attributes This will be used as input for the next step to verify whether the generalization effect achieves the expected safety level.

[0119] Step 3.3: Using the generalized attribute set generated in Step 3.2 For each object, repeat the same equivalence class size calculation process as steps 2.1 and 2.2 to obtain the new comprehensive equivalence class size for each record after generalization. Based on this, a generalized multidimensional correlation privacy risk index is calculated. The calculation formula is completely identical in form to the formula in step 2.3, except that... Replace with Here This is used to verify whether the current level configuration vector L has reduced the risk to an acceptable level.

[0120] like Still higher than the preset safety threshold If this indicates that the current generalization is insufficient, it is necessary to return to step 3.1 to adjust the hierarchical configuration vector L upwards; if If the generalization effect is satisfactory, no further adjustments are needed. Regardless of whether a callback is required, the final output of this step is a confirmed and effective set of generalized properties. (If no callback is received) If a callback occurs (This is the result after re-generalization). This confirms the valid set of generalized attributes. This will be used as input for the next step to generate the final generalization transformation description metadata.

[0121] Step 3.4: Confirm the valid generalized attribute set based on the output of Step 3.3 Extract the hierarchical configuration vectors used in the actual application during this generalization process. (i.e., the final effective version), and compare it with the global attribute dimension index table in step one. and the multidimensional related privacy risk index in step two Package and generate a generalized transformation description metadata. This metadata This metadata does not contain any individual patient data; it only records which attributes were generalized to what level, the change in risk before and after generalization, and the version of the generalization tree used. The purpose of this metadata is to provide a traceable transformation context for subsequent steps and to provide the data recipient with a basis for understanding the data granularity. This generalization transformation description metadata Together with the confirmed valid generalized attribute set Together, they constitute the complete output of step three.

[0122] Step 4: Adaptive noise injection and structure preservation balance processing based on generalization transformation to describe metadata;

[0123] After obtaining the confirmed valid generalized attribute set output in step three. and generalization transformation description metadata While simple generalization reduces the risk of direct identification, attackers can still exploit residual statistical distribution characteristics or outlier patterns in multidimensional correlation scenarios to make inferences. Therefore, this step requires introducing controlled random perturbations on top of generalization, while strictly constraining the magnitude and direction of the perturbations to ensure that the data after noise injection maintains consistent overall statistical characteristics with the data before generalization, avoiding unacceptable biases to cross-departmental collaborative analysis tasks. The core of this step lies in utilizing… The system dynamically adjusts noise parameters based on recorded risk variations to achieve an adaptive balance between privacy protection and data availability. Specifically, this includes:

[0124] Step 4.1: Describe the metadata using the generalization transformation output from Step 3. Using this as the input source, the difference in the multidimensional correlation privacy risk index before and after generalization is extracted. and the final effective hierarchical configuration vector .based on Calculate a noise scaling factor based on the size of the noise. .

[0125] Specifically, if A larger value indicates that the generalization operation has significantly reduced the risk, and at this point, a smaller noise injection can be allowed to avoid excessive distortion; if A small value indicates limited generalization effect, requiring stronger noise to compensate for the protection gap. Noise scaling factor The calculation uses an inverse proportional saturation form, that is... ,in This is a preset sensitivity constant. The result obtained from this is... It is a dimensionless coefficient between 0 and 1. The larger the value, the smaller the allowable noise amplitude, and vice versa.

[0126] Step 4.2: Based on the noise scaling factor calculated in Step 4.1 Combined with the valid generalized attribute set confirmed in step three The data type and value range of each attribute are defined, and a suitable noise standard deviation is generated for each numeric attribute. For the j-th numerical attribute, its noise standard deviation is... By this attribute interquartile range With noise scaling factor The calculation formula is as follows, jointly determined:

[0127]

[0128] In the formula, This represents the set of attributes after generalization for the j-th attribute. The interquartile range in the data has the same dimensions as the original dimension of the attribute (e.g., age is "years" and blood glucose concentration is "mmol / L"). The preset base noise ratio is a dimensionless constant used to set the baseline level of noise relative to the degree of data dispersion. The noise scaling factor obtained in step 4.1 is dimensionless. Therefore, Dimensions and Complete consistency ensures that the noise amplitude matches the scale of the data itself. For categorical attributes, numerical noise is not directly injected, but rather... Determine the probability of a category replacement This is used for potential subsequent label perturbation. All numerical attributes... With categorical attributes Together they constitute a noise parameter set .

[0129] Step 4.3: Using the noise parameter set generated in Step 4.2 The instruction is to apply the generalized set of attributes that were confirmed to be valid in step three. Noise injection is performed on each record. For numerical attributes, the noise injection operation starts from the value of zero and the corresponding standard deviation. A noise value is independently sampled from a Gaussian distribution and superimposed on the corresponding attribute value of the record; for categorical attributes, a probability is used... The current generalized label is randomly replaced with other valid labels at the same generalization level, following a uniform distribution during replacement.

[0130] It is worth noting that noise injection is not applied indiscriminately, but is subject to structure preservation constraints: if the generalization value of a record on a certain attribute is near the boundary of the attribute's value range, the injected noise will be truncated or reflected to prevent the generation of outliers that exceed semantically reasonable ranges (such as negative age or test values ​​exceeding physiological limits). After the above controlled perturbation, Transformed into a set of noisy attributes This set It introduces uncertainty at the individual level, but retains consistency with the global level. Similar means, variances, and correlation structures. This set of noisy attributes. This will be used as input for the next step to verify whether noise injection has compromised critical statistical consistency.

[0131] Step 4.4: Based on the noisy attribute set generated in Step 4.3 Calculate its generalized attribute set as confirmed in step three. Bias on key statistical moments. Specifically, this includes first-order moment (mean) bias, second-order moment (variance) bias, and Pearson correlation coefficient bias between principal attribute pairs. If any bias exceeds a preset tolerance threshold, the current noise parameter set is determined. The distortion is too large, so we need to go back to step 4.1 and lower the noise scaling factor. The lower limit or adjustment of the basic noise ratio factor Then, steps 4.2 to 4.3 are repeated; if all deviations are within tolerable limits, the noise injection is considered to have achieved a balance between privacy and usability. Regardless of whether a callback occurs, the final output of this step is a statistically validated set of noisy attributes. and the final noise parameter set actually used. These two, along with the generalization transformation in step three, describe the metadata. Together, these constitute the complete output of step four. They will be seamlessly transferred to step five as the basic input for local alignment and consistency verification before cross-departmental data fusion.

[0132] Step 5: Local alignment and consistency verification of cross-departmental data based on noisy attribute sets;

[0133] After obtaining the statistically validated set of noisy attributes output in step four. and final noise parameter set Subsequently, since the data from each participating hospital were processed independently and in parallel, noisy data from different sources may have subtle differences in attribute granularity, noise distribution parameters, and generalization levels. Directly splicing these data could lead to systematic bias or spurious associations during joint analysis. Therefore, this step aims to locally align the noisy data submitted by each hospital to ensure consistent statistical behavior and semantic interpretation within a unified reference framework. This alignment process does not alter the internal structure of individual hospital data; it eliminates inter-hospital heterogeneity only through lightweight calibration operations, specifically including:

[0134] Step 5.1: Using the statistically validated set of noisy attributes output from Step 4. To determine the target data set, first collect similar batches of data from all participating hospitals. For each batch, extract the sample mean and sample variance for each numerical attribute, and compare them with the final noise parameter set from step four. The implicit target statistical moments are compared. This is because each institute uses [the method described in] step four. Slight differences may exist due to variations in local data distribution, causing the overall distribution after merging to deviate from expectations. Therefore, an inter-institutional calibration offset is calculated for each numerical attribute. Where h represents the hospital number and j represents the attribute sequence number. This offset... This equals the difference between the actual sample mean of this attribute in this institute and the global target mean. Inter-institute calibration offset. The dimensions of the calibration offset are consistent with those of the corresponding attribute, and are used to compensate for mean drift between hospitals. The calibration offsets of all hospitals constitute a calibration matrix. The calibration matrix This will serve as the basis for the next step of correction, used to adjust the center position of the noisy data from each institute.

[0135] Step 5.2: Based on the calibration matrix generated in Step 5.1 For the noisy attribute sets submitted by each hospital that have passed statistical verification Perform mean calibration on each record.

[0136] Specifically, for each record from the h-th hospital, subtract the corresponding calibration offset from its j-th numerical attribute value. This aligns the data center of the institute with the global objective. For categorical attributes, no numerical calibration is performed; instead, the label sets of each institute at the same generalization level are checked for complete consistency. If labels are missing or inconsistently named, the global attribute dimension index table established in step one is used. Perform completion or renaming to ensure semantic space consistency of categorical variables. After the above calibration and label unification operations, each institute's... Transformed into a noisy set of locally aligned attributes This set It is highly consistent with the global objective in terms of statistical moments, and achieves cross-departmental mutual recognition in terms of classification labels.

[0137] Step 5.3: The set of noisy attributes after local alignment of each college generated in Step 5.2. For input, perform cross-department consistency checks. The checks include two aspects: first, checking whether the actual variance of noise in each department on the same attribute matches the final noise parameter set in step four. The declared theoretical variance is consistent. If the deviation exceeds the preset tolerance, the attribute of that institute is marked as noise mismatch. Secondly, the similarity of the joint distribution profiles of each institute under the same attribute combination is checked. Differences are quantified by calculating the KL divergence or Wasserstein distance of the joint distribution between institutes. If the distance exceeds a threshold, it is marked as a mismatch in the associated structure. The verification results are summarized into a consistency verification report. This records all mismatches and their severity. If a severe mismatch exists, it is necessary to go back to step four and re-perform noise injection and statistical validation on the relevant hospital data; if there is only a minor mismatch or all data passes, proceed to the next step.

[0138] Step 5.4: Consistency verification report generated in Step 5.3 If the report shows that all verification items pass or there are only acceptable minor mismatches, then the noisy attribute set after local alignment of each institute will be... The data is vertically concatenated based on the patient's anonymous ID (not a real identity identifier, but a temporary token used only for cross-hospital linking) to form a unified cross-hospital fusion dataset. Simultaneously, the final noise parameter set from step four... Step 3: Generalization Transformation Describes Metadata And the consistency verification report for this step. Packaged into a data quality and privacy protection certificate This voucher It does not contain any individual data, but it fully records the parameter selection, verification results, and compliance evidence of the entire data processing chain, for data users to audit and trace.

[0139] Step Six: Generate restricted query response and secure output based on cross-institute fusion dataset;

[0140] Obtain the cross-departmental integrated dataset output in step five. and data quality and privacy protection certificates Afterwards, the data has the basic conditions for secure sharing, but it cannot yet be directly accessed at the raw record level. This step constructs a restricted query response layer. All external analysis requests must be processed through this layer. The system only returns statistical results that have been securely aggregated and perturbed, preventing any form of individual-level data leakage. Specifically, this includes:

[0141] Step 6.1: Using the cross-departmental fusion dataset output in Step 5 As the underlying data source, load the data quality and privacy protection credentials from step five. Extract the final noise parameter set from it. Metadata described by generalization transformation Based on this information, a query parser is initialized. This parser can recognize externally submitted aggregate queries (such as COUNT, SUM, AVG, etc.) and automatically determine whether the attributes involved in the query are present. Generalization in, and in The parser determines the intensity of noise injected into the query. For queries involving ungeneralized or unnoised attributes, the parser will refuse to execute or enforce default protection; for valid queries, the parser will convert them into valid queries. The internal query plan is executed securely. This internal query plan includes the necessary aggregation grouping keys, filtering conditions, and output fields, and indicates the minimum secure aggregation granularity required for each output item.

[0142] Step 6.2: Based on the internal query plan generated in Step 6.1, in the cross-departmental fusion dataset output in Step 5... Aggregate operations are performed on it. Because... Since the noise injected in step four is already present, the result of direct aggregation is inherently random.

[0143] However, to prevent noise from being averaged out due to small-sample aggregation and thus exposing individual information, this step introduces an additional layer of output noise. Specifically, for each aggregated value in the query results, based on its corresponding group size... (i.e., the number of records contained in the aggregation group), an additional noise value is sampled from the Laplace distribution and superimposed on the aggregation result. The scale parameter b of the Laplace distribution is... Inversely proportional, that is ,in This is the preset output noise baseline. When the noise level is small, the added noise is significant, effectively masking the statistical characteristics of the small group; when... When the noise level is sufficiently high, it approaches zero, ensuring the accuracy of large sample results. After this operation, preliminary query results with dual noise protection are obtained. .

[0144] Step 6.3: Preliminary query results generated in Step 6.2 Perform output validity validation on the object. Validation rules include: checking the number of records in each aggregation group. Is it below the preset minimum release threshold? If the result is lower, set the result of that group to empty or merge it into the parent generalization group; check if the confidence interval width of the aggregated value is too large, if so, mark it as "low reliability" and add a prompt in the output; check if the result is consistent with the data quality and privacy protection certificate in step five. The level of privacy protection stated in the document is consistent with the level of privacy protection stated in the document. For example, if the certificate states... - Differential privacy protection verifies whether the output noise meets the corresponding requirements. Value requirements. Only query results that pass all validations are retained; those that fail are filtered or corrected. The validated query results are denoted as... .

[0145] Step 6.4: Based on the security query results output in Step 6.3 Combined with the data quality and privacy protection credentials in step five Generate a complete shared output package The output package contains three parts: first, the secure query results. The output package includes: 1) the query itself; 2) a metadata summary related to the query, including the generalization level used, noise parameters, aggregation group size distribution, and validation status; and 3) instructions for use, informing the data recipient how to correctly interpret the uncertainties and limitations in the results. This shared output packet contains no original records, intermediate processing data, or reversible transformation information, ensuring that individual information cannot be recovered even if the output packet is intercepted. This is the final product of the entire six-step process, which can be directly delivered to authorized analytical parties for use.

[0146] The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for sharing cross-hospital medical data with privacy protection, characterized in that, include: Standardize and semantically align multi-source heterogeneous medical records to construct an initial attribute space vector; A multidimensional correlation privacy risk index is calculated based on the initial attribute space vector, and the global disturbance intensity coefficient is determined accordingly. Based on the global perturbation intensity coefficient, an adaptive attribute generalization transformation is performed on the initial attribute space vector to generate a confirmed and valid set of generalized attributes and metadata describing the generalization transformation. The generalization transformation is used to adjust the noise parameters of the metadata description, and controlled noise is injected into the confirmed effective generalized attribute set to generate a noisy attribute set that has been statistically verified. Based on the statistically validated set of noisy attributes, cross-departmental local alignment and consistency verification are performed to generate a cross-departmental fusion dataset and data quality and privacy protection credentials. Based on the cross-institute fusion dataset and the data quality and privacy protection certificate, a shared output package is generated through a restricted query response.

2. The method for cross-hospital medical data privacy protection and sharing according to claim 1, characterized in that, Standardize and semantically align multi-source heterogeneous medical records to construct an initial attribute space vector, including: Parse the original electronic medical record dataset to extract all field values ​​and form the original attribute set; The cross-institutional general medical terminology ontology is loaded, and semantic matching is performed on the original attribute set to convert it into a semantically standardized intermediate attribute set. Traverse the semantically standardized intermediate attribute set to count the frequency and value distribution of each field, and establish a global attribute dimension index table; The attribute fields that meet the preset conditions are selected based on the global attribute dimension index table to form an initial attribute space vector.

3. The method for cross-hospital medical data privacy protection and sharing according to claim 2, characterized in that, A multidimensional correlation privacy risk index is calculated based on the initial attribute space vector, and the global perturbation strength coefficient is determined accordingly, including: Select attribute combinations from the initial attribute space vector and calculate the joint value frequency distribution to form a multidimensional correlation frequency matrix; The comprehensive equivalence class size of each record is calculated based on the multidimensional association frequency matrix, and a risk indication sequence is generated. A multidimensional associated privacy risk index is calculated using the aforementioned risk indication sequence and security equivalence class threshold; The global disturbance intensity coefficient is determined based on the multidimensional associated privacy risk index and the preset risk classification mapping rule.

4. The method for cross-hospital medical data privacy protection and sharing according to claim 3, characterized in that, Based on the global perturbation intensity coefficient, an adaptive attribute generalization transformation is performed on the initial attribute space vector to generate a confirmed set of generalized attributes and metadata describing the generalization transformation, including: The target generalization level is determined for each attribute in the initial attribute space vector by combining the global perturbation intensity coefficient, thus forming a hierarchical configuration vector; Based on the hierarchical configuration vector, a generalized mapping is performed on the semantically standardized intermediate attribute set to generate a generalized attribute set; The generalized multidimensional association privacy risk index of the generalized attribute set is recalculated, and the generalized attribute set is confirmed to be valid after being judged to be qualified. Extract the final effective hierarchical configuration vector and package it with the global attribute dimension index table and the multidimensional associated privacy risk index to generate generalized transformation description metadata.

5. The method for cross-hospital medical data privacy protection and sharing according to claim 4, characterized in that, The generalization transformation is used to adjust the noise parameters of the metadata description, and controlled noise is injected into the confirmed effective generalized attribute set to generate a statistically validated noisy attribute set, including: The risk difference is parsed from the generalized transformation description metadata, and the noise scaling factor is calculated. The noise standard deviation of each attribute is determined based on the noise scaling factor and the interquartile range of each attribute, thus forming a noise parameter set; According to the noise parameter set, controlled noise is injected into the confirmed effective generalized attribute set one by one to generate a noisy attribute set; The deviation of the statistical moments between the noisy attribute set and the data before generalization is verified. After the deviation is deemed acceptable, the noisy attribute set and the final noise parameter set that have passed statistical verification are obtained.

6. The method for cross-hospital medical data privacy protection and sharing according to claim 5, characterized in that, Based on the statistically validated set of noisy attributes, cross-departmental local alignment and consistency checks are performed to generate a cross-departmental fused dataset and data quality and privacy protection credentials, including: Extract the sample mean of the noisy attribute set that has passed statistical verification from each institute, and calculate the inter-institute calibration offset to form a calibration matrix; Based on the calibration matrix, mean calibration and classification label unification are performed on the data of each institute to generate a locally aligned set of noisy attributes; Perform a consistency check between the noise variance and the associated structure on the noisy attribute set after local alignment of each institute, and generate a consistency check report; The cross-institute fusion dataset is generated by splicing data from each institute based on the consistency verification report, and the final noise parameter set, generalization transformation description metadata, and consistency verification report are packaged into a data quality and privacy protection certificate.

7. A method for cross-hospital medical data privacy protection and sharing according to claim 6, characterized in that, Based on the aforementioned cross-institute fusion dataset and the aforementioned data quality and privacy protection certificate, a shared output package is generated through a restricted query response, including: Load the data quality and privacy protection credentials to initialize the query parser and convert the external aggregate query into an internal query plan; Based on the internal query plan, an aggregation operation is performed on the cross-institute fusion dataset, and output noise is added according to the size of the aggregation group to generate preliminary query results; The validity of the minimum publication threshold and confidence interval width is validated on the preliminary query results to obtain safe query results; The secure query results, along with the metadata summary and usage instructions of the data quality and privacy protection certificate, are encapsulated into a shared output package.

8. A method for cross-hospital medical data privacy protection and sharing according to claim 4, characterized in that, The generalized multidimensional association privacy risk index of the generalized attribute set is recalculated. After passing the judgment, a valid generalized attribute set is obtained, including: Recalculate the new comprehensive equivalence class size for each record on the generalized attribute set; The generalized multidimensional correlation privacy risk index is calculated using the size of the new comprehensive equivalence class and the threshold of the security equivalence class. When the generalized multidimensional association privacy risk index is higher than the security threshold, the hierarchical configuration vector is adjusted upward and the generalization mapping is re-executed. When the generalized multidimensional association privacy risk index is not higher than the security threshold, the current generalization result is used as the confirmed valid set of generalized attributes.

9. A method for cross-hospital medical data privacy protection and sharing according to claim 6, characterized in that, Perform a consistency check on the noise variance and associated structure of the locally aligned noisy attribute set of each institute, and generate a consistency check report, including: Check the deviation between the actual noise variance of each institute on the same attribute and the theoretical variance declared in the final noise parameter set; Calculate the distance metric for the joint distribution of each college under the same attribute combination; When the deviation or distance metric exceeds the preset tolerance, the corresponding mismatch item is marked and the severity is recorded. Generate a consistency verification report by summarizing all mismatches and their severity.

10. A cross-hospital medical data privacy protection and sharing system for implementing the method as described in any one of claims 1-9, characterized in that, include: Semantic alignment unit is used to standardize semantic alignment of multi-source heterogeneous medical records and construct an initial attribute space vector. The risk quantification unit is used to calculate a multidimensional associated privacy risk index based on the initial attribute space vector, and thereby determine the global disturbance intensity coefficient. The generalization transformation unit is used to perform adaptive attribute generalization transformation on the initial attribute space vector according to the global perturbation intensity coefficient, and generate a confirmed valid set of generalized attributes and metadata describing the generalization transformation. The noise injection unit is used to adjust the noise parameters of the metadata description using the generalization transformation, inject controlled noise into the confirmed effective generalized attribute set, and generate a noisy attribute set that has been statistically verified to be qualified. The alignment verification unit is used to perform cross-departmental local alignment and consistency verification based on the statistically verified set of noisy attributes, and generate cross-departmental fusion dataset and data quality and privacy protection certificate; The secure output unit is used to generate a shared output package based on the cross-institute fusion dataset and the data quality and privacy protection certificate through a restricted query response.

Images