Differential privacy-based medical data desensitization traceability auditing method

Abstract

The application discloses a medical data desensitization and traceability auditing method based on differential privacy, and relates to the technical field of medical data security, and specifically comprises the following steps: collecting two types of medical data and standardizing the format, extracting the clinical decision weight of numerical data and the privacy risk label of text data, and then implementing differential desensitization based on the weight and the label, adding Laplace noise with correlation weight to the numerical data, replacing or masking the key information of the text data in stages, embedding a unique traceability identifier and verifying privacy compliance and availability, finally establishing a closed-loop auditing mechanism to support identifier extraction, compliance verification and leakage tracking, and continuously optimizing relevant parameters and strategies based on feedback, effectively guaranteeing data security and traceability.

Description

Technical Field

[0001] This invention belongs to the field of medical data security technology, specifically involving a method for de-identifying and tracing the source of medical data based on differential privacy. Background Technology

[0002] With the rapid development of medical informatization, the value of medical data (including structured numerical data and text-based diagnostic data) in areas such as clinical diagnosis and treatment optimization and medical research innovation is becoming increasingly prominent, and the demand for data sharing and cross-scenario applications continues to grow. However, medical data contains a large amount of sensitive and private information, and its leakage may lead to serious consequences such as discrimination and reputational damage to patients, making privacy protection compliance requirements increasingly stringent.

[0003] Existing medical data anonymization technologies have significant shortcomings: First, they lack differentiated processing strategies and fail to accurately match the anonymization intensity with the data's impact on clinical decision-making and privacy risk level, easily leading to over-anonymization causing decreased data usability or under-anonymization causing privacy leaks. Second, their source tracing and auditing capabilities are weak; anonymized data lacks effective identification and flow records, making it difficult to trace the source and flow path after a data leak. Third, the balance mechanism between privacy protection and usability is imperfect, lacking quantitative verification based on strict privacy definitions (such as differential privacy) and lacking compliance verification methods specific to the characteristics of medical data. Fourth, feature extraction and risk labeling are not precise enough, failing to fully integrate professional information such as clinical guidelines and treatment pathways, making it difficult to match the specific needs of medical scenarios.

[0004] Therefore, there is an urgent need for a differential privacy-based method for de-identifying and tracing medical data to address the aforementioned issues. Summary of the Invention

[0005] The purpose of this invention is to provide a method for de-identifying and tracing medical data based on differential privacy, which solves the technical problem in the prior art that the de-identification intensity is not adapted to the degree of influence of medical data on clinical decision-making and the level of privacy risk, resulting in a rigid de-identification strategy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] Differential privacy-based methods for anonymizing and tracing the source of medical data include:

[0008] Step 1: Collect two types of medical data and unify them into a standardized format. For numerical data, calculate the clinical decision weight by adjusting the guideline evidence level, treatment pathway, and application frequency. For text data, extract semantic vectors and match them with a list of high-risk diseases to label privacy risks.

[0009] Step 2: Implement differentiated desensitization based on data weight and privacy risk labels. For numerical data, add Laplace noise with associated weights. For textual data, replace or mask key information in a hierarchical manner. Embed a unique traceability identifier and establish an association library. At the same time, verify privacy protection compliance and clinical usability.

[0010] Step 3: Establish a closed-loop audit mechanism, extract traceability identifiers and verify the compliance of desensitization, track data flow nodes when leakage occurs, and combine audit results with clinical feedback to regularly optimize weight calculation coefficients, desensitization parameters and audit rules.

[0011] Furthermore, two types of medical data were collected and standardized into a unified format. The specific method is as follows:

[0012] Two types of medical data are collected: one is structured numerical data, including blood glucose levels, tumor marker concentrations, and routine blood tests; the other is text-based diagnostic data, including disease diagnosis conclusions, symptom descriptions, and treatment recommendations.

[0013] Both types of target data are unified into a standardized JSON structure. The structured numerical data retains the original measurement accuracy and simultaneously labels the measurement units and reference ranges. The text-based diagnostic data is based on industry-related literature, and medical terminology is standardized and segmented to remove meaningless auxiliary words, retaining only the core semantic fields.

[0014] Furthermore, numerical data are used to calculate clinical decision weights through guideline evidence grading, treatment pathway adaptation, and application frequency adjustments. The specific method is as follows:

[0015] The guideline quantifies evidence by collecting relevant documents, extracting recommendation clauses related to the target data fields, classifying recommendation levels according to the GRADE evidence grading standard, assigning values ​​to different levels, and determining the recommendation level score for each data field. ;

[0016] The number K of key nodes in the statistical data field in the standard treatment pathway. Calculate the path fit coefficient based on the maximum number of key nodes for the same type of data field. ;

[0017] Based on historical clinical case data, the frequency (f) of each data field used by doctors in decision-making was statistically analyzed. f is equal to the ratio of the number of decision records for a particular data field to the total number of decision records. This is achieved using the formula... Clinical decision weights are calculated, where a1 and a2 represent the weight coefficients of the recommendation grade score and the path fit coefficient, respectively, and a1 + a2 = 1. Multiple linear regression analysis is performed based on historical clinical diagnosis and treatment decision data, with the decision accuracy improvement rate as the dependent variable and the recommendation grade score and path fit coefficient as independent variables. The weight coefficients of both are calculated using stepwise regression. g represents the number of recommendation grades, and A represents the basic weight percentage of f. This represents the percentage of the correction factor for f, which is set based on historical data and actual conditions.

[0018] Furthermore, semantic vectors are extracted from text-based data and matched with a list of high-risk diseases to label privacy risks. The specific method is as follows:

[0019] Based on historical real-world cases of medical data breaches, a list of disease types that would cause serious harm such as discrimination, significant reputational damage, or health loss after a breach is compiled. Using a BERT medical pre-trained model, semantic vectors are extracted from standardized text-based diagnostic data to capture core semantic information. The extracted semantic vectors are then semantically matched with disease terms in the high-risk disease list. Pre-set matching criteria are used; if a match is successful, the corresponding high-risk label is applied, and if no match is found, the corresponding common disease label is applied.

[0020] Furthermore, for numerical models, Laplace noise is added to the correlation weights. The specific method is as follows:

[0021] Using the Laplace noise mechanism, the noise intensity is negatively correlated with the clinical decision weight ω, and the desensitization formula is: Where Δf is the data sensitivity, ε is the preset privacy budget, x is the original value, and x' is the anonymized value, ensuring that the differential privacy definition is met. This indicates Laplace distribution sampling, with the constraint that after desensitization, x' must fall within [reference range minimum value - c × Δf, reference range maximum value + c × Δf]. If it exceeds this range, the boundary value is taken. c is a constant value set according to historical data and actual needs.

[0022] Δf = Maximum value of the reference range for this field - Minimum value of the reference range;

[0023] The principle behind setting the privacy budget ε is based on the dynamic adaptation of the clinical decision weight ω of medical data to the privacy risk level. High-weight data takes a higher ε value, and low-weight data takes a lower ε value. Differential privacy is defined as a constraint to balance data privacy security and clinical / research availability.

[0024] Furthermore, for text-based hierarchical replacement or masking of key information, the specific methods are as follows:

[0025] For high-risk labels, a key term replacement and semantic preservation strategy is adopted to build a term replacement library. Privacy information such as disease names and symptom characteristics are replaced with synonymous generalized terms. The semantic similarity of the text before and after replacement is calculated using a medical BERT model. If the similarity is ≥d, it is considered qualified. If it is not qualified, the replacement term is reselected. d is a constant value set according to historical data and actual needs.

[0026] For common disease labels, a lightweight masking strategy is adopted, which randomly masks only non-core diagnosis and treatment information, while core semantic fields remain unchanged. The lightweight masking rule is: using the [MASK] identifier of the BERT model, only non-core fields in the text are masked, and core diagnosis and treatment fields are not masked.

[0027] Furthermore, a unique traceability identifier is embedded and an association database is established. The specific method is as follows:

[0028] An invisible traceability identifier is embedded in the anonymized data. Based on the standardized characteristics of each data entry, and combined with the anonymization batch and processing node information, a unique hash identifier is generated. ;

[0029] Structured numerical data will Extended fields embedded in JSON structure, text data will After Base64 encoding, the low-frequency dimension of the embedded semantic vector is used, and the embedding rule is to replace the original vector value with the ASCII code of the encoded character minus 32.

[0030] The mapping relationship with key information for data anonymization, including the original data summary, anonymization algorithm type, noise intensity and replacement rules, executors, and processing time, forms an immutable anonymization log, completing the establishment of the associated database.

[0031] Furthermore, a closed-loop audit mechanism will be established to verify the compliance of anonymization after extracting traceability identifiers. The specific method is as follows:

[0032] When auditing is required, extract the anonymized data using parsing tools. For structured numerical data, the `data_verify` field of the JSON can be read directly using a Python script to obtain the data. For text-based data, extract dimensions 128-160 of the semantic vector, convert the vector values ​​+32 to ASCII code, concatenate them, and then decode using Base64. If decoding fails, it is determined that the data has been tampered with. If the semantic vector of text data is tampered with, the low-frequency dimension identification information is restored through the medical BERT model.

[0033] Then conduct a compliance verification and pass. By linking and tracing the relevant database, we can verify whether the de-identification algorithm matches the data type, whether the noise intensity and replacement rules meet the ω value and privacy risk label level, whether the personnel performing the operation have the necessary permissions, and whether the processing time is within the compliance window, and generate a de-identification compliance report.

[0034] Furthermore, in the event of a leak, the data flow nodes are traced, specifically through the following methods:

[0035] When an anonymized data leak is discovered, through By tracing the data flow path and using the data_flow_path field of the source-related database, combined with the access logs of the hospital information system, we can locate all nodes of data download, use, and secondary sharing. We can also locate the leakage points by combining the de-identified logs. By comparing the audit logs and operation logs, we can investigate abnormal operations and output a source tracing report.

[0036] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0037] 1. This invention implements differentiated desensitization by using clinical decision weights and privacy risk labels. Numerical data uses a Laplace noise mechanism to dynamically adjust the noise intensity, while textual data is classified by label and uses term replacement or lightweight masking strategies. Furthermore, the privacy protection effect is quantitatively verified through differential privacy budget calculation and adjacent dataset attack tests. At the same time, semantic similarity and relative error rate are used to ensure data availability, avoiding the problems of data invalidation due to over-desensitization and privacy leakage due to insufficient desensitization in traditional desensitization techniques.

[0038] 2. This invention embeds unique traceability identifiers into de-identified data, embeds JSON extended fields into structured data, and encodes text data to embed low-frequency dimensions of semantic vectors. Combined with a traceability database storing key de-identified information, this forms an immutable flow record. During auditing, specialized tools can quickly extract the identifiers, verify the compatibility of the de-identification algorithm and data type, the matching degree of processing parameters, weights, and tags, and the compliance of the operating entity's permissions and processing time, accurately locating any violations.

[0039] 3. This invention establishes a continuous optimization mechanism based on audit results and clinical feedback, updating the clinical decision weight calculation coefficient, guideline recommendation score, and high-risk disease list every quarter. It adjusts desensitization parameters such as noise intensity and terminology replacement library based on error rate and semantic similarity feedback in data availability verification. At the same time, it improves the anti-tampering capability of traceability labels and the dimensions of audit log recording by combining leakage tracking cases. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 The diagram illustrates the steps of the medical data anonymization and traceability auditing method based on differential privacy according to the present invention.

[0042] Figure 2 The diagram illustrates the steps of the medical data differential desensitization method of the present invention.

[0043] Figure 3 The diagram illustrates the steps of the closed-loop auditing and optimization method of the present invention. Detailed Implementation

[0044] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1, such as Figure 1 The differential privacy-based medical data anonymization and source tracing auditing method shown includes the following steps:

[0046] Step 1: Collect two types of medical data and standardize them into a unified format. For numerical data, calculate the clinical decision weight by adjusting the guideline evidence level, treatment pathway, and application frequency. For text data, extract semantic vectors and match them with a list of high-risk diseases to label privacy risks.

[0047] Medical data was collected and categorized into two types based on data format: structured numerical data (such as blood glucose levels, tumor marker concentrations, and routine blood tests) and text-based diagnostic data (such as disease diagnoses, symptom descriptions, and treatment recommendations). Both types of data were formatted using a standardized JSON structure. Structured numerical data retained its original measurement accuracy, with measurement units (such as mmol / L and ng / mL) and reference ranges annotated. Text-based diagnostic data (such as disease diagnoses, symptom descriptions, and treatment recommendations) underwent standardized medical terminology segmentation (based on the Medical Subject Headings, MeSH), removing meaningless auxiliary words and retaining core semantic fields.

[0048] Focusing on structured numerical data (such as blood glucose and tumor markers) and text-based diagnostic data in electronic medical records, this study standardizes the JSON format and extracts domain-specific features. For numerical data, clinical decision weights ω are assigned, where ω is defined as the degree of influence of the data on clinical diagnosis and treatment decisions. An objective value determination method based on guideline evidence quantification and clinical application frequency correction is adopted.

[0049] The quantification of evidence in the guidelines involves collecting relevant documents issued by the National Health Commission, such as the "Clinical Practice Guidelines" and the "Clinical Laboratory Results Mutual Recognition Criteria," extracting recommendation clauses related to the target data fields, and classifying them into recommendation levels according to the GRADE evidence grading standard (e.g., strong recommendation = 4 points, moderate recommendation = 3 points, weak recommendation = 2 points, no recommendation = 1 point). Values ​​are assigned to different levels to determine the recommendation level score for each data field. ;

[0050] Treatment pathway adaptation: The number K of key nodes in the standard treatment pathway (key nodes include diagnosis confirmation, treatment plan selection, efficacy evaluation, and prognosis) of statistical data fields. The maximum number of key nodes for a data field of the same type (such as test data). ), calculate path adaptation coefficient ;

[0051] Clinical application frequency correction: Based on historical clinical case data, the frequency percentage f of each data field used by doctors for decision-making is calculated (i.e., the number of decision records containing one data field / the total number of decision records). The formula is then used to correct this. Clinical decision weights are calculated, where a1 and a2 represent the weight coefficients of the recommendation grade score and the path fit coefficient, respectively, and a1 + a2 = 1. Multiple linear regression analysis is performed based on historical clinical diagnosis and treatment decision data, with the decision accuracy improvement rate as the dependent variable and the recommendation grade score and path fit coefficient as independent variables. The weight coefficients of both are calculated using stepwise regression. g represents the number of recommendation grades, and A represents the basic weight percentage of f. This represents the percentage of the correction factor for f, set based on historical data and actual conditions;

[0052] Semantic vectors are extracted using the BERT medical pre-trained model, and privacy risk labels associated with diseases are annotated. These privacy risk labels are defined as the type of privacy leakage risk associated with the data-related diseases (e.g., high-risk labels such as infectious diseases, malignant tumors, etc.). The method for determining the values ​​is to match a list of high-risk diseases in a medical data breach case library. Specifically, by analyzing historical real-world medical data breach cases, a standardized list of high-risk diseases is formed by statistically analyzing disease types that caused serious harm (e.g., discrimination, significant reputation / health loss) after a breach. Semantic vectors of textual diagnostic data are extracted using the medical BERT model and semantically matched with disease terms in the high-risk disease list (e.g., a similarity score ≥ 0.85 is considered a successful match). If a match is successful, the corresponding high-risk label is applied (e.g., high-risk infectious disease, high-risk malignant tumor). If no match is found, that label is not applied. In the example, HIV diagnostic data is tagged with "high-risk infectious disease," lung cancer diagnostic data with "high-risk malignant tumor," and common cold diagnostic data with "common disease."

[0053] Step 2: Implement differentiated desensitization based on data weight and privacy risk labels. For numerical data, add Laplace noise with associated weights. For textual data, replace or mask key information in a hierarchical manner. Embed a unique traceability identifier and establish an association library, while verifying privacy protection compliance and clinical usability.

[0054] like Figure 2 As shown, based on the clinical decision weights ω (numerical data) obtained in step one and the privacy risk labels (textual data), a differentiated anonymization strategy is adopted to ensure a balance between privacy protection and data availability:

[0055] Structured numerical data: Laplace noise mechanism is used, where noise intensity is negatively correlated with clinical decision weight ω (the higher the ω, the stronger the data availability requirement, and the lower the noise intensity). The desensitization formula is as follows: Where Δf is the data sensitivity, ε is the preset privacy budget, x is the original value, and x' is the anonymized value, ensuring that the differential privacy definition is met. This indicates Laplace distribution sampling, with the constraint that after desensitization, x' must fall within [reference range minimum value - 5% × Δf, reference range maximum value + 5% × Δf], otherwise the boundary value is taken.

[0056] Δf = maximum reference range value of this field - minimum reference range value (e.g., blood glucose reference range 3.9-6.1 mmol / L, Δf = 2.2);

[0057] The principle behind setting the privacy budget ε is based on the dynamic adaptation of the clinical decision weight ω of medical data to the privacy risk level. Data with high weight (high availability requirements) takes a higher ε value, and data with low weight (high privacy protection requirements) takes a lower ε value. Differential privacy is defined as a constraint to balance data privacy security and clinical / research availability. In this embodiment, the default value of the privacy budget is set to ε=1.0, which is adjusted to ε=1.5 when ω≥0.8 and to ε=0.5 when ω≤0.3.

[0058] Text-based diagnostic data: processed according to privacy risk labels;

[0059] High-risk labels (infectious diseases, malignant tumors, etc.): A key term replacement and semantic preservation strategy is adopted to build a term replacement library (e.g., based on the "Medical Terminology Classification and Code", containing 892 sets of synonymous generalized term pairs, with the replacement method being "HIV" → "specific infectious diseases", "lung cancer" → "malignant lung tumors", "liver cancer" → "malignant liver tumors"). Privacy information such as disease names and symptom characteristics are replaced with synonymous generalized terms. The semantic similarity of the text before and after replacement is calculated using a medical BERT model. A similarity of ≥0.8 is considered acceptable. If it is unacceptable, the replacement term is reselected (maximum of 3 retries; if it is still unacceptable, key information is deleted and annotations are added, such as deleting "HIV" and adding "[related to infectious diseases]").

[0060] Common disease labels: A lightweight masking strategy is adopted, which randomly masks only non-core diagnosis and treatment information (such as indirect patient descriptions), while core semantic fields (such as key nodes of the treatment plan) are left as is. The lightweight masking rule is to use the [MASK] label of the BERT model to mask only non-core fields such as "patient self-report" and "family description" in the text, with a masking ratio of ≤10%. Core diagnosis and treatment fields (such as "acute bronchitis" and "oral amoxicillin") are not masked.

[0061] To ensure the traceability of anonymized data, an invisible traceability identifier is embedded in the anonymized data, without affecting the clinical use and format standardization of the data.

[0062] Generate a unique traceability identifier: Based on the standardized features of each data point (such as the ω value for numerical data and the semantic vector hash value for textual data), and combined with the de-identification batch and processing node information, generate a unique hash identifier. (Length is 32 bits);

[0063] Identifier embedding method: Structured numerical data will Extended fields embedded in the JSON structure (field names set to non-sensitive identifiers, such as data_verify), text data will... After Base64 encoding, the low-frequency dimension of the semantic vector is embedded. The embedding rule is to replace the original vector value with the ASCII code of the encoded character minus 32 (to ensure that the overall distribution of the vector is not changed).

[0064] Establish a traceability database: storage The mapping relationship with key information for data anonymization, including the original data summary, anonymization algorithm type, noise intensity and replacement rules, executors, processing time, etc., forms an immutable anonymization log.

[0065] Verify the privacy compliance and clinical usability of anonymized data to avoid over- or under-anonymization:

[0066] Privacy protection verification: using formula Calculate the actual differential privacy budget for each data point, where `scale` is the Laplace noise scaling parameter, ensuring that all data meet the requirements. Less than or equal to a preset threshold (e.g., ≤1.5);

[0067] The adjacent dataset attack test verifies that attackers cannot infer the privacy information of a single patient in the original data from the anonymized data. Two adjacent datasets, D1 and D2, differing only from a single patient data point are constructed. D1' and D2' are then anonymized. A machine learning model (logistic regression) is used to attempt to infer the difference data. The attack success rate is calculated as the number of correct inferences divided by the total number of tests. The attack success rate is required to meet a preset requirement, such as ≤5%.

[0068] Clinical usability validation: Numerical data using formulas Calculate the relative error rate; for text-based data, calculate the semantic similarity before and after de-identification using a medical BERT model.

[0069] Step 3: Establish a closed-loop audit mechanism, extract traceability identifiers and verify the compliance of desensitization, and track data flow nodes in the event of leakage; combine audit results with clinical feedback, and regularly optimize weight calculation coefficients, desensitization parameters and audit rules.

[0070] like Figure 3 As shown, the source identification is extracted. When auditing is required, a dedicated parsing tool is used to extract the source identification from the anonymized data. For example, for structured numerical data, a Python script can be used to read the data_verify field of the JSON and directly obtain the data. For text-based data, extract dimensions 128-160 of the semantic vector, convert the vector values ​​+32 to ASCII code, concatenate them, and then decode using Base64. (If decoding fails, it is determined that the data has been tampered with). If the semantic vector of text data is tampered with, the low-frequency dimension identification information is restored through the medical BERT model.

[0071] Then conduct a compliance verification and pass. By linking and tracing the relevant database, we can verify whether the de-identification algorithm matches the data type, whether the noise intensity and replacement rules meet the ω value and privacy risk label level, whether the personnel performing the operation have the necessary permissions, and whether the processing time is within the compliance window period, and generate a de-identification compliance report.

[0072] When an anonymized data leak is discovered, through By tracing the data flow path and using the "data_flow_path" field in the source-related database, combined with the access logs of the Hospital Information System (HIS), all nodes of data download, use, and secondary sharing are located. By combining the de-identified logs, the leakage links are located. By comparing the audit logs and operation logs, abnormal operations (such as unauthorized downloads, tampering with labels, and unauthorized forwarding) are investigated, and a source tracing report is output.

[0073] Based on audit results and clinical feedback, continuously optimize model parameters and desensitization strategies:

[0074] Weight and label optimization: The calculation coefficients (a1, a2, α) of the clinical decision weight ω are updated quarterly and adjusted in accordance with the latest clinical practice guidelines. Scoring, supplementing the medical data breach case database to update the list of high-risk diseases;

[0075] Desensitization parameter optimization: Based on the error rate and semantic similarity feedback in usability validation, adjust the Laplace noise intensity coefficient and term substitution library to ensure a dynamic balance between privacy protection and clinical use;

[0076] Audit rule optimization: Based on leakage tracing cases, improve the anti-tampering capability of the source identification (such as adding digital signatures) and expand the recording dimensions of audit logs (such as data access duration and modification operation traces).

[0077] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0078] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0079] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for de-identifying and tracing the source of medical data based on differential privacy, characterized in that, include: Step 1: Collect two types of medical data and unify them into a standardized format. For numerical data, calculate the clinical decision weight by adjusting the guideline evidence level, treatment pathway, and application frequency. For text data, extract semantic vectors and match them with a list of high-risk diseases to label privacy risks. The guideline quantifies evidence by collecting relevant documents, extracting recommendation clauses related to the target data fields, classifying recommendation levels according to the GRADE evidence grading standard, assigning values ​​to different levels, and determining the recommendation level score for each data field. ; The number K of key nodes in the statistical data field in the standard treatment pathway. Calculate the path fit coefficient based on the maximum number of key nodes for the same type of data field. ; Based on historical clinical case data, the frequency (f) of each data field used by doctors in decision-making was statistically analyzed. f is equal to the ratio of the number of decision records for a particular data field to the total number of decision records. This was achieved using the formula... Clinical decision weights are calculated, where a1 and a2 represent the weight coefficients of the recommendation grade score and the path fit coefficient, respectively, and a1 + a2 = 1. Multiple linear regression analysis is performed based on historical clinical diagnosis and treatment decision data, with the decision accuracy improvement rate as the dependent variable and the recommendation grade score and path fit coefficient as independent variables. The weight coefficients of both are calculated using stepwise regression. g represents the number of recommendation grades, and A represents the basic weight percentage of f. This represents the percentage of the correction factor for f, set based on historical data and actual conditions; Step 2: Implement differentiated desensitization based on data weight and privacy risk labels. For numerical data, add Laplace noise with associated weights. The Laplace noise mechanism is used, and the noise intensity is negatively correlated with the clinical decision weight ω. For textual data, replace or mask key information in a hierarchical manner. For high-risk labels, adopt key term replacement and semantic preservation strategies, build a term replacement library, and replace disease names and symptom feature privacy information with synonymous generalized terms. For common disease labels, adopt a lightweight masking strategy, embed a unique traceability identifier and establish an association library, and verify privacy protection compliance and clinical usability. Step 3: Establish a closed-loop audit mechanism, extract the source identification mark and verify the compliance of desensitization, track the data flow nodes when leakage occurs, and combine the audit results with clinical feedback to regularly optimize the calculation coefficients a1, a2, α, desensitization parameters and audit rules of the clinical decision weight ω. The desensitization parameter optimization method is as follows: adjust the Laplace noise intensity coefficient and term substitution library according to the error rate and semantic similarity feedback in usability verification.

2. The method for de-identifying and tracing medical data based on differential privacy according to claim 1, characterized in that, Two types of medical data were collected and standardized into a uniform format. The specific method is as follows: Two types of medical data are collected: one is structured numerical data, including blood glucose levels, tumor marker concentrations, and routine blood tests; the other is text-based diagnostic data, including disease diagnosis conclusions, symptom descriptions, and treatment recommendations. Both types of target data are unified into a standardized JSON structure. The structured numerical data retains the original measurement accuracy and simultaneously labels the measurement units and reference ranges. The text-based diagnostic data is based on industry-related literature, and medical terminology is standardized and segmented to remove meaningless auxiliary words, retaining only the core semantic fields.

3. The method for de-identifying and tracing medical data based on differential privacy according to claim 1, characterized in that, Extract semantic vectors from text-based data, match them with a list of high-risk diseases, and label them with privacy risk tags. The specific method is as follows: Based on historical real-world medical data breaches, a list of disease types that would cause serious harm after a breach is compiled. Using the BERT medical pre-trained model, semantic vectors are extracted from the standardized text-based diagnostic data to capture core semantic information. The extracted semantic vectors are then semantically matched with disease terms in the high-risk disease list. Pre-set matching criteria are used; if a match is successful, the corresponding high-risk label is applied, and if no match is found, the corresponding common disease label is applied.

4. The method for de-identifying and tracing medical data based on differential privacy according to claim 1, characterized in that, The specific method for adding Laplace noise with correlation weights to numerical data is as follows: The desensitization formula is: Where Δf is the data sensitivity, ε is the preset privacy budget, x is the original value, and x' is the anonymized value, ensuring that the differential privacy definition is met. This indicates Laplace distribution sampling, with the constraint that after desensitization, x' must fall within [reference range minimum value - c × Δf, reference range maximum value + c × Δf]. If it exceeds this range, the boundary value is taken. c is a constant value set according to historical data and actual needs. Δf = Maximum value of reference range - Minimum value of reference range; The principle behind setting the privacy budget ε is based on the dynamic adaptation of the clinical decision weight ω of medical data to the privacy risk level. High-weight data takes a higher ε value, and low-weight data takes a lower ε value. Differential privacy is defined as a constraint to balance data privacy security and clinical / research availability.

5. The method for de-identifying and tracing medical data based on differential privacy according to claim 1, characterized in that, The specific method for text-based hierarchical replacement or masking of key information is as follows: For high-risk tags, the semantic similarity of the text before and after replacement is calculated using the medical BERT model. If the similarity is ≥d, it is considered acceptable. If it is not acceptable, the replacement term is reselected. d is a constant value set according to historical data and actual needs. For common disease labels, only non-core diagnostic and treatment information is randomly masked, while core semantic fields are left as is. The lightweight masking rule is: use the [MASK] identifier of the BERT model to mask only non-core fields in the text, and prohibit masking core diagnostic and treatment fields.

6. The method for de-identifying and tracing medical data based on differential privacy according to claim 1, characterized in that, Embed a unique traceability identifier and establish an associated database, the specific method is as follows: An invisible traceability identifier is embedded in the anonymized data. Based on the standardized characteristics of each data entry, and combined with the anonymization batch and processing node information, a unique hash identifier is generated. ; Structured numerical data will Extended fields embedded in JSON structure, text data will After Base64 encoding, the low-frequency dimension of the embedded semantic vector is used, and the embedding rule is to replace the original vector value with the ASCII code of the encoded character minus 32. The mapping relationship with key information for data anonymization, including the original data summary, anonymization algorithm type, noise intensity and replacement rules, executors, and processing time, forms an immutable anonymization log, completing the establishment of the associated database.

7. The method for de-identifying and tracing medical data based on differential privacy according to claim 6, characterized in that, Establish a closed-loop audit mechanism, extract traceability identifiers, and verify the compliance of anonymization. The specific method is as follows: When auditing is required, extract the anonymized data using parsing tools. For structured numerical data, the `data_verify` field of the JSON can be read directly using a Python script to obtain the data. For text-based data, extract dimensions 128-160 of the semantic vector, convert the vector values ​​+32 to ASCII code, concatenate them, and then decode using Base64. If decoding fails, it is determined that the data has been tampered with. If the semantic vector of text data is tampered with, the low-frequency dimension identification information is restored through the medical BERT model. Then conduct a compliance verification and pass. By linking and tracing the relevant database, we can verify whether the de-identification algorithm matches the data type, whether the noise intensity and replacement rules meet the ω value and privacy risk label level, whether the personnel performing the operation have the necessary permissions, and whether the processing time is within the compliance window, and generate a de-identification compliance report.

8. The method for de-identifying and tracing medical data based on differential privacy according to claim 6, characterized in that, When a data leak occurs, the data flow nodes are traced. The specific method is as follows: When an anonymized data leak is discovered, through By tracing the data flow path and using the data_flow_path field of the source-related database, combined with the access logs of the hospital information system, we can locate all nodes of data download, use, and secondary sharing. We can also locate the leakage points by combining the de-identified logs. By comparing the audit logs and operation logs, we can investigate abnormal operations and output a source tracing report.

Images