Medical data desensitization identification and processing method, system and device based on artificial intelligence

Abstract

The invention relates to the technical field of medical data processing, and particularly provides a medical data desensitization identification and processing method, system and device based on artificial intelligence. According to the method, an original medical text is obtained through a data access module, a pre-training language model and a sequence labeling model are fused to recognize sensitive fields, a desensitization strategy is dynamically matched on the basis of field sensitivity levels, user roles and data purposes, and character shielding, forgery replacement or encryption processing is executed. And meanwhile, the system generates a desensitization mapping table and a hash chain log, so that the whole process can be audited and the field can be controllably restored. And when the identification confidence is low or the strategy execution is abnormal, an exception reexamination mechanism is automatically triggered to ensure that the processing quality is in compliance with the data. The system has a strategy self-learning capability, supports multi-scene deployment, and improves the intelligence, compliance and availability of medical data privacy protection.

Description

Technical Field

[0001] This invention relates to the field of data security and privacy protection technology, and in particular to a method, system and device for de-identifying and processing medical data based on artificial intelligence. Background Technology

[0002] With the deepening of medical informatization, medical institutions, research platforms, and health management systems have accumulated massive amounts of electronic medical data containing personal privacy information, including but not limited to electronic medical records, outpatient records, medical images, laboratory reports, and medical insurance settlement statements. This data has extremely high reuse value in scenarios such as artificial intelligence modeling, public health research, and medical research. However, this type of data generally contains a large amount of personally identifiable information (PII) and health privacy information (PHI). During sharing, transmission, or cross-border use, it is essential to legally de-identify the data to meet the requirements of regulations such as the "Personal Information Protection Law of the People's Republic of China," the "Cybersecurity Law of the People's Republic of China," and the "Electronic Medical Record Management Standards" regarding the compliant use of personal data.

[0003] Existing medical data anonymization technologies mainly include the following two types of solutions: 1. Rule-based static desensitization method This method relies on pre-defined rules for sensitive fields, such as keyword matching and regular expressions, to identify and replace sensitive fields. Such methods are simple to implement, consume few computational resources, and are suitable for structured data or template-based text scenarios. However, due to their limited ability to understand contextual semantics, they often fail to accurately identify semantically complex and highly concealed sensitive words, easily leading to misidentification or omissions in the desensitization process.

[0004] 2. Field-oriented structured masking system These systems perform data anonymization on a field-by-field basis, uniformly masking fields such as patient name, ID number, and contact information. While effective for structured data, this approach is almost entirely incapable of recognizing unstructured free text widely present in medical documents (such as chief complaint, medical history, and examination descriptions), and even less capable of handling variant expressions, indirect privacy exposures, or context-dependent information within the text.

[0005] The aforementioned prior art has the following main technical defects: Insufficient accuracy in identifying sensitive information: It lacks context modeling capabilities, has poor recognition capabilities in complex sentence structures or polysemous scenarios, and is especially unable to identify privacy fields after semantic transformation; The processing strategies are simplistic and rigid: Most solutions lack the ability to dynamically respond to user permissions, usage scenarios, or field sensitivity levels. The de-identification strategies are too fixed, often resulting in over-identification (affecting usability) or under-identification (posing a risk of leakage). Lack of anomaly handling and review mechanisms: Existing solutions generally lack automatic processing and manual review processes when model recognition fails, policy execution is abnormal, or recognition conflicts occur, resulting in poor system fault tolerance. Unauditable and untraceable: Most systems lack a complete log chain and mapping structure, making it difficult to meet the compliance requirements for post-event auditing, accountability, and behavior verification. The system is not adaptive and lacks evolutionary capabilities: it lacks model update mechanisms and feedback loop capabilities, and cannot adapt to changes in data distribution or the emergence of new types of privacy modes during long-term operation, resulting in a gradual degradation of processing capabilities.

[0006] Although recent studies have attempted to introduce natural language processing techniques, especially named entity recognition (NER) methods, to improve the ability to identify privacy fields in medical texts, these solutions have mostly remained at the model experiment level and have not yet built a complete integrated de-identification system framework that includes "identification → policy judgment → execution processing → log auditing → anomaly review → permission restoration".

[0007] Therefore, there is an urgent need at present for an integrated desensitization system that is oriented towards the actual application scenarios of medical data, integrates artificial intelligence and rule strategies, and has the functions of intelligent identification, strategy control, fault tolerance and compliance auditing. This system can effectively solve the systemic deficiencies of traditional solutions in terms of identification accuracy, strategy flexibility, compliance and evolution capabilities, and meet the dual technical requirements of "high-value sharing" and "high-strength protection" of medical data. Summary of the Invention

[0008] To achieve the aforementioned objectives, this invention proposes an artificial intelligence-based method for de-identifying and processing medical data, and outlines several feasible technical solutions based on its core processes and system structure. The following describes possible implementations of this invention in conjunction with the technical features defined in the claims, to better understand the composition and functional positioning of each technical solution. The terminology and structures in these implementations are substantially consistent with the statements in the claims, with only appropriate adjustments made to the form of expression.

[0009] In one embodiment of the present invention, an artificial intelligence-based method for de-identifying and processing medical data is provided, comprising the following steps: S1. The control data access module obtains raw medical data and performs format parsing to convert unstructured or semi-structured data into a structured field sequence. S2. Integrate the pre-trained language model and the sequence labeling model to perform privacy field identification on the field sequence and obtain a labeling mapping that includes field type, location information and identification confidence. S3. Match a de-identification strategy based on the sensitivity level of the field, the user access role, and the purpose of data use. The de-identification strategy includes at least one of character masking, forgery replacement, and encryption processing. S4. Execute the matching de-identification strategy to generate de-identified data; S5. Generate processing logs and establish a mapping structure between original fields and de-identified fields; S6. When the recognition confidence level is lower than the preset threshold, the field hits an abnormal word set, or the strategy execution fails, the exception handling mechanism is triggered, the field is added to the list to be reviewed, and an alarm message is sent.

[0010] Furthermore, the privacy field identification model includes: a pre-trained language model for generating word vector representations; a bidirectional long short-term memory neural network (BiLSTM) for extracting contextual semantic features; a conditional random field (CRF) for predicting the optimal label path; and a rule engine for supplementing the identification of sensitive fields with obvious structured features, thereby improving the identification accuracy in complex semantic environments.

[0011] Furthermore, the selection of the desensitization strategy is based on the field sensitivity level L, the user access role R, and the data purpose U, satisfying the following strategy matching function: ;

[0012] Preferably, when L is high-level and R is a researcher, a character masking strategy is preferred; when L is medium-level and U is a display scenario, a forgery and replacement strategy is selected; when L is high-level and U is for reporting purposes, an encryption strategy is selected to meet the data usage needs under different privacy protection scenarios.

[0013] Furthermore, the anomaly handling mechanism includes any one or more of the following: determining whether the identification confidence level is lower than a preset threshold; determining whether the field hits the blacklist of words; determining whether the strategy execution has failed; and when any of the conditions are met, marking the corresponding field as an anomaly, adding it to the review queue and generating an alarm message.

[0014] Furthermore, before performing privacy field identification, the method may optionally include: performing field integrity verification to determine whether the input data has missing fields, incorrect format, or illegal symbols; if an anomaly is found, the desensitization process is stopped and the user is prompted to correct the data to prevent erroneous data from interfering with the identification results.

[0015] Furthermore, the method also includes a strategy evolution module, which records historical desensitization strategies and their execution effects. Based on user feedback and failure samples, the module adjusts the strategy selection rules through a machine learning model, supports strategy self-learning and optimization, and improves the accuracy and adaptability of strategy matching.

[0016] Furthermore, in a preferred embodiment of the present invention, the encryption processing method in the desensitization strategy preferably employs the AES-256 symmetric encryption algorithm and performs the encryption operation in CBC mode (Cipher Block Chaining). The system invokes this algorithm when processing highly sensitive fields and generates a 256-bit key and a dynamic initialization vector (IV) through the key management module to enhance the data processing's replay resistance and encryption strength.

[0017] Furthermore, when constructing the field mapping relationship structure, the mapping table preferably adopts a JSON data structure model, including but not limited to fields such as field unique identifier (field_id), original value, de-identified value, field type, strategy type, processing confidence level, processing timestamp, session identifier (session_id), encryption key number and corresponding log chain hash reference, so as to support functions such as controllable field restoration, compliance auditing and log tracking.

[0018] In one embodiment of the present invention, a medical data anonymization identification and processing system for implementing the above method is provided, comprising: a data access and parsing module for receiving raw medical data and completing format standardization; a privacy field identification module for performing named entity recognition based on an artificial intelligence model and outputting field type, location information, and confidence level; a policy matching module for determining anonymization strategies based on field sensitivity level, user role, and purpose; an anonymization execution module for implementing anonymization processing according to the matching results; a log recording module for generating anonymization logs and field mapping structures; and an anomaly handling module for handling low-confidence or rule-abnormal fields and triggering a review process.

[0019] Preferably, the log recording module constructs a de-identified hash chain and generates the log chain in the following manner: ; ;

[0020] This ensures that the de-identification process is tamper-proof and traceable, thereby enhancing the security and auditing capabilities of data processing.

[0021] Furthermore, the system also includes a permission recovery mechanism for restoring the original value of a field when the following conditions are met: Verify that the user's identity level, access scenario tag, and time window satisfy the following rule function: ;

[0022] Where U is the user ID, T is the access timestamp, and P is the usage scenario tag. This mechanism ensures reversible control over data within specific permission ranges, complying with medical data security regulatory requirements.

[0023] In one embodiment of the present invention, a medical data de-identification and processing device deployed on a server or terminal device is provided, comprising: a storage unit for storing model parameters, strategy configuration and de-identification mapping information; and a network interface module for receiving raw medical data input. The processor is configured to execute the above method steps and generate de-identification results and audit logs.

[0024] Based on the above technical solutions, the AI-based medical data desensitization identification and processing method, system, and device of the present invention address the shortcomings of existing desensitization technologies, such as poor identification accuracy, rigid strategy control, weak anomaly handling capabilities, and lack of audit trails. They construct a fully intelligent desensitization processing solution that integrates a natural language processing model and a multi-dimensional parameter control mechanism, achieving the following significant technical effects and practical value: 1. Significantly improved recognition accuracy By introducing a fusion-based artificial intelligence model architecture, pre-trained language models (such as BERT) and sequence labeling structures (BiLSTM-CRF) are used in collaborative computation, enabling the system to understand semantic context and accurately identify sensitive fields such as names, ID numbers, contact information, addresses, and diagnostic information in medical texts. It is particularly adaptable to unstructured electronic medical record texts, effectively solving the problems of low accuracy and high misidentification rate of traditional rule matching methods in complex semantic scenarios.

[0025] 2. Supports dynamic policy decision-making based on multiple parameters. The system has a built-in multi-dimensional control parameter set (including field sensitivity level, user role permissions, data usage scenarios, etc.) and uses a policy matching function to dynamically allocate desensitization strategies. It can select appropriate processing methods such as character masking, forgery replacement, or reversible encryption according to different business environments, avoiding the "one-size-fits-all" approach of traditional solutions that result in excessive or insufficient desensitization, and improving the usability and compliance of the desensitized data.

[0026] 3. Implement automatic detection and fault tolerance for abnormal scenarios. By setting a confidence threshold and a blacklist word set matching mechanism, the system can automatically identify abnormal fields that are uncertain in model recognition, fail in strategy execution, or hit sensitive patterns, triggering the desensitization pause and review process, thereby improving the system's robustness and security control capabilities in complex or unpredictable scenarios.

[0027] 4. Construct a fully auditable processing structure. The system generates a chained hash log after each field processing and constructs a de-identification mapping table based on the mapping relationship between the original field and the de-identified field, achieving full traceability and immutability, and meeting the mandatory requirements of the "Data Security Law of the People's Republic of China" and the "Personal Information Protection Law" for data processing compliance, auditability, and accountability.

[0028] 5. Supports reversible control over permission restoration and desensitization. By setting the permission verification function f recover (U, T, P) allows specific users to controllably restore specific fields under specific identity levels, usage time limits, and scenario requirements, thus meeting the core privacy compliance requirements of "using while protecting and protecting while using" in medical business scenarios.

[0029] 6. Possesses self-evolution and continuous optimization capabilities. The system supports the access of a policy evolution module, which can automatically update policy matching rules based on historical policy execution results and user feedback. It continuously optimizes the policy library and desensitization effect with the help of decision trees or reinforcement learning models, thereby enhancing the intelligent adaptability of the system in the long-term operation process.

[0030] 7. Excellent system deployment adaptability and engineering feasibility The method and system of this invention can be embedded and deployed in hospital HIS systems, scientific research data platforms, government data export channels or privacy computing frameworks. It is compatible with various medical data formats such as JSON, XML, text and image reports, and has horizontal scalability and vertical integration capabilities. It has extremely high engineering implementation value and industrial promotion prospects.

[0031] In summary, this invention constructs an intelligent de-identification platform suitable for real-world medical scenarios through a combination of model-based intelligent recognition, fine-grained policy control, dynamic anomaly handling, log compliance auditing, reversible permission control, and policy self-learning evolution. This comprehensively improves the intelligence level, privacy protection strength, and regulatory compliance capabilities of medical data processing, demonstrating significant technological advancement and practical application value. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and do not constitute a limitation on the scope of protection of the present invention.

[0033] Figure 1 This is a flowchart illustrating the medical data desensitization identification and processing method in an embodiment of the present invention, showing the overall method structure of the present invention; Figure 2This is a functional module structure diagram of the medical data desensitization identification and processing system in an embodiment of the present invention, illustrating the various modules of the system and their interrelationships; Figure 3 This is a schematic diagram of the artificial intelligence privacy field recognition model structure in an embodiment of the present invention, illustrating the structural hierarchy of BERT, BiLSTM, and CRF working together; Figure 4 This is a decision logic diagram of the desensitization strategy selection function in an embodiment of the present invention, which describes the logical path of strategy matching based on sensitivity level, role permissions, and data usage. Figure 5 This is a schematic diagram of the de-identification processing log chain construction structure in an embodiment of the present invention, illustrating an immutable log chain structure built based on a hash function; Figure 6 This is a flowchart illustrating the exception handling and review triggering process in an embodiment of the present invention, showing the fault-tolerant control process when the identification confidence is insufficient or the strategy fails. Detailed Implementation

[0034] To more clearly understand the technical solution and beneficial effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and several specific embodiments. Those skilled in the art should understand that the described embodiments are only used to illustrate the technical concept and implementation path of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the essential content of the present invention, any functional modifications, parameter adjustments, or module replacements made to the structure of each module, algorithm details, or processing flow should be considered equivalent technical solutions of the present invention and are also included within the scope of protection of the present invention.

[0035] Implementation Method 1: Medical Data Desensitization Method and Flowchart (Combined with...) Figure 1 ) In one embodiment of the present invention, such as Figure 1 As shown, an artificial intelligence-based method for desensitizing and processing medical data is provided. This method is applicable to electronic medical texts containing patient personal identification information (PII) and health privacy information (PHI), such as electronic medical records, test reports, discharge summaries, and medical imaging reports.

[0036] This implementation method mainly includes the following three technical logic layers: I. System Structure Layer: Functional Modules and Interaction Relationships This method relies on the collaborative operation of multiple functional modules to form a complete desensitization processing system, specifically including: 1. Data Access and Parsing Module: Used to receive raw medical text data (structured, semi-structured, or unstructured) and complete format recognition, field extraction, and standardization preprocessing.

[0037] 2. Privacy Field Recognition Module: This module integrates deep learning-based semantic modeling and rule reinforcement mechanisms to accurately identify potentially private entities (such as names, ID numbers, contact information, diagnostic information, etc.) from text.

[0038] 3. Strategy Matching and Decision Module: Based on the sensitivity level of the identified fields, user access permissions, and data usage, automatically determine the most suitable de-identification strategy.

[0039] 4. De-identification Execution Module: Used to perform specific character masking, forgery replacement, or reversible encryption operations on the target field.

[0040] 5. Log and Mapping Generation Module: Used to establish the mapping structure between the original text and the de-identified fields, and to generate operation logs to support subsequent auditing and reversible field recovery.

[0041] 6. Exception Handling and Review Module: Used for manual review and fault tolerance of fields that fail to be identified, have abnormal policy execution, or hit risk rules.

[0042] The above modules are in accordance with Figure 1 The processes shown work together sequentially to form a closed-loop processing system from the input of original medical text to the output of compliant de-identified data.

[0043] II. Method and Flow Layer: Step Logic and Control Mechanism Combination Figure 1 The specific processing flow is as follows: Step S1: Data Access and Format Parsing The system receives various types of raw medical data through an interface and determines their data type. For unstructured text, the system calls tools such as word segmenters and paragraph cutters to construct standardized field sequences.

[0044] For example, the input text T is represented as: Each sentence It is a word sequence.

[0045] Step S2: Privacy Field Identification This step integrates context vector generation and sequence labeling mechanisms for named entity recognition (NER): Generate word embedding vectors using pre-trained language models (such as BERT). ; Extracting contextual latent semantics using a bidirectional LSTM embedded in the input. ; The label path is decoded through the CRF layer to output the entity category and its position in the original text; And record the confidence level of each identification tag. .

[0046] The field recognition results are shown below: ;

[0047] Step S3: Strategy Judgment and Matching The system calculates the strategy function based on the field's sensitivity level (L), visitor role level (R), and purpose tag (U): ;

[0048] For example: if L = high and U = report, the matching strategy is encryption; if L = medium and R = doctor, the matching strategy is masking; if L = low and U = display, the matching strategy is forgery and replacement.

[0049] Step S4: Perform desensitization processing Based on the matching results, select a specific de-identification function to process the sensitive fields: Masking strategy:

[0050] Replacement strategy:

[0051] Encryption strategy:

[0052] The system generates a new de-identified text T'.

[0053] To further enhance the security and technical exclusivity of this invention in scenarios involving highly sensitive fields, preferably, the AES-256 (Advanced Encryption Standard) symmetric encryption algorithm is used when implementing the encryption strategy. The encryption process details are as follows: Algorithm mode: Adopts AES-256 encryption standard, with high security, supports ECB / CBC mode configuration, and CBC mode is preferred to improve the ability to resist replay attacks; Key management: The system has a preset key generation module with a key length of 256 bits, and a periodic rotation mechanism ensures long-term security. Initialization Vector (IV): In CBC mode, an IV is dynamically generated using a pseudo-random number generator to prevent the same plaintext from generating the same ciphertext in different sessions; Encrypted Field Scope: This encryption strategy is mainly applied to highly sensitive fields (such as ID card number, name, diagnosis name, etc.) to avoid directly exposing user privacy information; Reversible control logic: In conjunction with the permission recovery mechanism, under the verification functions that satisfy user roles, access time and usage scenarios, controlled restoration operations are allowed on encrypted fields.

[0054] This AES encryption strategy has a mature engineering implementation scheme, combining high security and performance efficiency, making it suitable for high-risk scenarios involving compliant de-identification of medical data.

[0055] Step S5: Log and Map Generation Construct mapping table M: json { Original text field: "Zhang San", "Desensitized field": "***", Type: "Name" "Strategy": "mask", Confidence level: 0.94 } ``` Simultaneously generate de-identified chained logs:

[0056]

[0057] To enhance the structural consistency of field restoration control, permission recovery, and log auditing mechanisms in the system, this invention further standardizes the data organization structure of the field mapping table.

[0058] In this invention's system, the correspondence between original fields and de-identified fields is stored and managed in the form of a structured mapping table. This mapping table is implemented using a JSON-formatted structured data model, facilitating machine reading and field restoration control. A schematic structure is shown below: json { "field_id": "F123456", "original_value": "Zhang San", "masked_value": "***", "field_type": "Name", "strategy": "mask", "confidence": 0.92, "timestamp": "2025-12-14T09:21:00Z", "session_id": "S202512140001", "encryption_key_id": "K-AES-2048-001", "hash_reference": "e3a9c7d..." } ``` Field description: field name meaning field_id Unique field identifier, used for field-level indexing original_value Raw field values ​​(visible only under controlled permissions) masked_value De-identified field values, such as `***` or encrypted ciphertext. field_type Field type labels (such as name, phone number, diagnosis, etc.) strategy The desensitization strategies implemented (such as mask, replace, encrypt) confidence Field recognition confidence comes from AI recognition model timestamp Processing timestamps session_id The current processing session number is used for log chain tracing. encryption_key_id If the encryption strategy is used, the corresponding key number hash_reference The hash digest value of the corresponding node in the log chain supports auditability and immutability. This mapping table is not only used to record the results of field processing, but also serves as the basic data structure for the permission recovery mechanism and the audit log mechanism, and has good scalability, security and audit adaptability.

[0059] Step S6: Exception Handling and Review Triggering When the confidence level of field recognition is lower than the threshold If the blacklist field set B is hit, or if the policy execution fails, the exception handling module will be automatically triggered, and the field will be added to the review queue. And issue a manual review prompt.

[0060] III. Algorithm and Physical Meaning Layer: Parameter Interpretation and Actual Effects Field Sensitivity Level L: Represents the degree of risk of privacy leakage for a field (e.g., "Name" is highly sensitive, "Gender" is low sensitive), and is directly proportional to the degree of desensitization. User role R: Represents the user's permission level, common values ​​are such as administrator > doctor > researcher; Purpose label U: Indicates the purpose of the data (display / training / reporting), affecting whether reversible restoration is allowed; Identification confidence score P(y|x): Reflects the model's confidence in the current prediction result; below a set threshold. Manual intervention is triggered at certain times; Policy function Physical meaning: This function achieves a dynamic balance between "desensitization intensity and scenario usability", that is, while ensuring data security, it avoids excessive desensitization from affecting data mining; Hash Log Chain Recording the processing order and content of fields provides an unalterable traceability capability for the desensitization process, serving as technical support for compliance auditing.

[0061] By implementing the above methods and processes, this invention can maximize the protection of patient privacy information while maintaining the availability of medical data, and has the ability to control anomalies, support permissions, and audit logs throughout the entire process, thus meeting the needs of the medical industry for compliant use and intelligent governance of data in real-world environments.

[0062] Implementation Method Two: System Structure and Module Description (combined with...) Figure 2 ) In one embodiment of the present invention, such as Figure 2As shown, a medical data desensitization identification and processing system is provided. This system is used to automatically desensitize and identify electronic medical record-type medical text data containing sensitive information, determine strategies, execute processing, control anomalies, and audit logs, thereby improving the level of intelligence in data sharing and compliant use.

[0063] The system can be deployed as an independent platform in hospital HIS systems, scientific research data platforms, privacy computing frameworks, or data export gateways, supporting unified desensitization processing of structured and unstructured medical data.

[0064] I. System Structure Layer: Module Composition and Functional Division like Figure 2 As shown, the system includes the following core modules: 1. Data Access and Parsing Module: Used to receive, standardize, and serialize fields of raw medical data. It supports multiple input formats (JSON, TXT, XML, DICOM, etc.) and outputs a unified field structure that can be recognized by the model.

[0065] 2. Privacy Field Recognition Module: This module integrates the BERT-BiLSTM-CRF model with a rule matching engine to accurately locate and label privacy fields that may be contained in the input text, and outputs the field location, type, and confidence information.

[0066] 3. Policy Judgment and Matching Module: Based on the sensitivity level L of the identified result field, the access user's permission role R, and the current data purpose U, the policy function f is used to determine the appropriate match. s (L, R, U) is used to select the optimal desensitization strategy, ensuring the dynamism and controllability of strategy matching.

[0067] 4. De-identification Execution Module: Receives the policy matching results and performs corresponding de-identification operations, including character masking, forgery replacement, or symmetric encryption. Supports field-level granular differential processing.

[0068] 5. Anomaly Handling Module: This module is used to monitor and identify risky situations such as low confidence levels, policy failures, or fields hitting abnormal rules. It automatically submits abnormal fields to the review process and supports configuring alarm triggering and permission prompt mechanisms.

[0069] 6. Log recording and mapping module: Used to generate a mapping table between the original fields and the de-identified fields, and to perform chained logging of the processing behavior (based on SHA256 hash function) to ensure that the de-identification process is traceable, verifiable and tamper-proof, and to support data compliance auditing and permission recovery requirements.

[0070] II. Module Functional Layer: Sub-functional Units and Internal Logic 1. The data access and parsing module further includes: Format recognition unit: Determines the structure of the input data and processes it accordingly; Word segmentation and paragraph segmentation units: Transforming free text into analysis units; Field normalization unit: Construct a standard field sequence T = {w1, w2, ..., w}.

[0071] 2. The privacy field identification module further includes: BERT embedding encoder: transforms each word into a context vector; BiLSTM layer: Obtains the forward and backward context of a word; CRF decoder: Outputs entity labels and boundaries; Rule engine: Used to detect fields with specific formats (such as ID card number, mobile phone number) and supplement the identification.

[0072] 3. The strategy judgment and matching module includes: Policy configuration unit: Preset de-identification rule library; Decision function unit: Dynamically selects the optimal strategy based on (L, R, U); Parameter validation module: verifies the validity of input combinations to avoid conflicts.

[0073] 4. The desensitization execution module includes: Masking processor: performs character-level masking on the target field (e.g., "***"); Replacement processor: Calls the semantic replacement library D and selects a high similarity value from it; Encryption processor: Performs symmetric encryption operations, such as the AES encryption function.

[0074] 5. The exception handling module includes: Threshold detector: Determines whether the recognition confidence level is lower than a set threshold θ; Blacklist rule matcher: detects whether a field contains sensitive blacklisted words; Fault tolerance logic: Stop the desensitization of abnormal fields and push them to the review queue.

[0075] 6. The logging and mapping module includes: Desensitization mapping generator: Constructs a field mapping structure for controllable restoration; Hash log generator: Generates an immutable log chain in the following way:

[0076]

[0077] III. Interaction Mechanism Layer: Explanation of Logical Flow Between Modules Medical data is first input and parsed by the data access module, and then transformed into a standardized field stream; The field stream is fed into the privacy recognition module to complete entity recognition and risk field labeling; The identified field information is passed to the strategy matching module to determine the best processing method; The desensitization module performs specific processing on the fields and generates desensitized text; Meanwhile, processing actions are recorded in the log and mapping module, establishing an audit foundation; If any uncertainty or anomaly occurs during the identification or processing, the anomaly handling module will immediately intervene and initiate a review process to ensure system security and stability.

[0078] Through the above-mentioned module composition and interaction relationships, the system of the present invention can achieve "intelligent identification, dynamic strategy, auditable behavior, and controllable restoration of results" in the process of de-identification of medical data, thereby improving the availability and transparency of data circulation while ensuring privacy compliance.

[0079] Implementation Method 3: Privacy Field Recognition Model Mechanism (Combined with...) Figure 3 ) In one embodiment of the present invention, such as Figure 3 As shown, a privacy field recognition model mechanism integrating deep language model is provided to accurately identify and label privacy information (such as name, ID number, contact information, address, disease name, etc.) in unstructured medical texts such as electronic medical records and medical records.

[0080] This model is based on the BERT-BiLSTM-CRF architecture, which integrates pre-trained semantic modeling and sequence labeling capabilities, and combines a medical-specific rule engine to enhance the results. It has high entity recognition accuracy and robustness, and can adapt to complex language environments in the medical field, such as polysemous words, abbreviations, and non-standard sentence structures.

[0081] I. Model Structure Composition (corresponding to) Figure 3 ) The overall structure of the model can be divided into the following three main sub-modules from bottom to top: 1. Embedding Layer Pre-trained language models (such as BERT) are used to perform context encoding on the input text to obtain the semantic vector representation of each word.

[0082] 2. Sequence Modeling Layer (BiLSTM Layer) Bidirectional Long Short-Term Memory (BiLSTM) networks are used to model the sequence of context vectors and extract the dependency information between the current word and the words before and after it.

[0083] 3. Label Prediction Layer (CRF Layer) The Conditional Random Field (CRF) is used to predict the label for each word and output the optimal named entity path, which is used to label the category of the field (such as B-NAME, I-NAME, B-ID, etc.).

[0084] II. Processing Flow and Calculation Logic Assume the input text is a sequence of sentences: T = {w1, w2, ..., w} The entire recognition process is as follows: (1) BERT embedding generation Obtain the contextual embedding representation of each word using the BERT model: ;

[0085] Where (d) represents the word vector dimension, which is usually 768.

[0086] (2) BiLSTM sequence modeling Inputting word vectors into a BiLSTM network yields the concatenation of the forward and backward hidden states for each word: ;

[0087] Where h is the hidden layer dimension of the unidirectional LSTM.

[0088] (3) Linear mapping and label score The BiLSTM output is mapped to the label space using a linear layer:

[0089] in, K is the number of label categories, and b is the bias term.

[0090] (4) CRF prediction of optimal label sequence All words are jointly decoded using a conditional random field layer, and the tag path is output. The optimization objective is to maximize the path score. ; in, Let be the label transition matrix, representing the transition from label Move to label The score.

[0091] Optimal path search using the Viterbi algorithm: ; Finally, the label type of each word (e.g., B-NAME, I-ID, O, etc.) is output, and the field recognition result is generated by combining the word position.

[0092] III. Analysis of Model Parameters and Physical Meaning

[0093] Furthermore, this invention introduces a medical-specific rule engine to enhance the model's ability to accurately identify special fields (such as 18-digit ID card numbers, 11-digit mobile phone numbers, etc.). The rule recognition results will be fused with the model results; if the two are inconsistent, a "low-confidence review mechanism" will be triggered.

[0094] IV. Explanation of Technical Effects and Innovations This invention achieves the following technical effects through this recognition model: 1. Improved recognition accuracy: By integrating deep semantics and sequence dependency modeling, it significantly outperforms traditional rule matching methods; 2. Enhanced adaptability to complex language: Maintains high robustness in medical terminology, abbreviations, and polysemous expressions; 3. Supports field confidence level output: facilitates integration with the strategy module to achieve confidence-based anomaly detection and review; 4. Transferable and controllable training: Supports fine-tuning mechanisms, allowing for continued training on specific hospital or research data to improve local adaptability.

[0095] In a preferred embodiment of the present invention, the privacy field identification model (BERT-BiLSTM-CRF) is trained in a supervised manner based on an annotated medical privacy dataset. The training data mainly comes from the following typical datasets: 1. Real-world desensitized and annotated medical datasets: including desensitized and authorized electronic medical records (EMR), inpatient summaries, laboratory reports, medical image description text, etc., with field annotations covering sensitive information (PII / PHI) such as name, ID number, contact information, address, disease name, and drug name; 2. Publicly available medical text corpora: such as anonymized international medical datasets like i2b2 and MIMIC-III, used for model pre-training and semantic transfer; 3. Rule-enhanced corpus: Field variant samples (such as different formats of ID cards and phone numbers) generated by regular expressions and template engines are used to enhance the model's ability to recognize non-standard expressions.

[0096] To ensure the generalization ability of the model, this invention adopts the following strategy: A domain-adaptive fine-tuning mechanism is introduced to support retraining of the model in target deployment scenarios (such as a specific format of medical records in a hospital), thereby improving scenario adaptability. Multi-regional and multi-institutional data cross-validation was used on the model training samples to improve the robustness of recognition of different language styles, word usage habits and medical formats; The model output includes a confidence score. When predicting fields with low confidence, the system automatically guides the model into an anomaly review process, enhancing the model's controllability and robustness. It supports the fusion of rule engine recognition results and neural network prediction results, and improves the final recognition accuracy through result fusion strategies (such as weighted average / confidence priority).

[0097] In summary, the privacy field recognition model adopted in this invention is built on real medical text corpora and has good semantic understanding ability, cross-corpus generalization ability and application scenario transfer ability. It can be widely applied to electronic text desensitization tasks in different medical institutions, data formats and business scenarios.

[0098] Implementation Method Four: Desensitization Strategy Control Mechanism (Combined with...) Figure 4 ) In one embodiment of the present invention, such as Figure 4 As shown, a medical data desensitization strategy control mechanism is provided. After the privacy fields are identified, the mechanism dynamically matches the most suitable desensitization strategy based on multiple control parameters (field sensitivity level, access role, data purpose, etc.) and ensures that the processing results achieve the optimal balance between compliance and usability.

[0099] The core of this control mechanism lies in constructing a strategy decision function that, combined with field context attributes and visitor permission environment, automatically completes the entire process of "selecting the de-identification method, generating the processing path, and outputting parameter configuration".

[0100] I. Definition and Source of Control Parameters This invention defines the following three key control parameters: 1. Field sensitivity level (L) This is used to indicate the sensitivity of a field to patient privacy, and can be set to three categories: High (L=3): such as name, ID number, medical record number; (L=2): such as age, address, and phone number; Low (L=1): such as consultation time, name of examination equipment.

[0101] 2. Access User Roles (R) Reflecting the access level of data users, typical roles include: Administrator (R=3): Has full access and auditing privileges; Doctor (R=2): Has access to diagnosis and treatment information; Researchers (R=1): Used only for non-identifiable research analyses.

[0102] 3. Data Usage (U) Mark the current data usage scenario: Display (U=1): Used for interface display or reporting; Training (U=2): Used for model training or testing; Reporting (U=3): Used to transmit data to regulatory authorities or third parties.

[0103] The above parameters are automatically injected by the upstream and downstream environment of the system, or selected / input by the user in the platform configuration.

[0104] II. Strategy Matching Function Design The system has a set of preset desensitization strategies: ;

[0105] Based on the parameters ((L, R, U)), the system calls the policy matching function: ;

[0106] This function is implemented using a rule-based decision table and a priority weighting mechanism, such as Figure 4 As shown.

[0107] III. Strategic Decision-Making Logic Flow (combined with...) Figure 4 ) Figure 4 The following diagram illustrates the decision path of the policy function, showing the key judgment rules: Combination of conditions Matching results Explanation of processing meaning L = height, R ≤ 2 (non-administrator) mask (to cover up) Preventing the leakage of core identity information L = in, U = display replace Ensure that the displayed content has reasonable semantics. L = High, U = Report encrypt Meet compliance data reporting requirements L = Low, R = Doctor, U = Training Optional: Non-desensitizing The data risk is low and it can be used directly. Other combinations mask is the default policy. If a clear strategy is not found, a conservative approach will be adopted. IV. Physical Meaning of Parameters and Strategies The core value of this strategy function lies in achieving dynamic strategy adaptation among "field-user-scenario", with the following technical effects: 1. Field sensitivity level L ↔ Desensitization intensity The more sensitive the field, the more stringent the desensitization. L determines the boundaries of the strategy selection.

[0108] 2. User access role R↔ permission control The higher the R value, the more likely a weak desensitization or restoration path will be chosen, reflecting the concept of hierarchical access control.

[0109] 3. Usage label U↔ Availability trade-off U controls whether fuzzy replacement is allowed or encrypted transmission is mandatory, ensuring adaptability to various business scenarios.

[0110] By combining and mapping the three parameters, this invention avoids the problem of "single strategy and unadjustable" in traditional systems, and achieves dynamic control capabilities driven by diverse strategies and usage scenarios.

[0111] V. Extended Functionality and Configurability In a preferred embodiment, the strategy determination module further includes: 1. Policy rule configuration interface Administrators can configure the platform through the backend to customize and add new rule combinations and policy mappings, and extend processing strategies (such as desensitization levels, mask styles, etc.).

[0112] 2. Strategy evolution mechanism support (in conjunction with implementation method seven) Combining model feedback, policy failure logs, and user preferences, the policy matching function f is optimized. s (L, R, U) is used for self-learning optimization to improve the accuracy of strategy adaptation.

[0113] By implementing the above-mentioned desensitization strategy control mechanism, this invention effectively achieves differentiated processing of different types of privacy fields, avoiding the risk of "over-desensitization" affecting data analysis or "insufficient desensitization" causing privacy leakage. It has high adaptability and engineering practical value in actual medical data processing systems.

[0114] Implementation Method 5: Log Chain Structure and Audit Mechanism (Combined) Figure 5 ) In one embodiment of the present invention, such as Figure 5 As shown, a hash chain-based medical data anonymization log structure and its auditing mechanism are proposed to record all key operational events generated during the anonymization process of medical data, and to provide reliable technical support for subsequent compliance audits, behavior tracing, responsibility definition and access restoration.

[0115] This mechanism addresses the common problem of lack of operational verifiability and traceability in existing desensitization systems, and is particularly suitable for the technical environment required by laws and regulations such as the Personal Information Protection Law, the Data Security Law of the People's Republic of China, and the Guidelines for Cross-border Transfer of Medical and Health Data, which require data operations to have a traceable, verifiable, and non-repudiable technical environment.

[0116] I. Log Chain Structure Design (Combined with...) Figure 5 ) like Figure 5 As shown, this invention designs a chain-structured operation log chain (Hash Chain), whose basic components include: Original session identifier (session_id) Operation timestamp Process field content (field) The desensitization strategy used. Previous log hash (H) n-1 ) Log node digests are generated using standard one-way hash functions (such as SHA-256) to form a chain structure.

[0117] The hash value of each node in the log chain is defined as follows: Initial node (H0):

[0118] The hash node corresponding to the processing behavior of the (i)th field is:

[0119] Ultimately, this forms a log chain structure that is sequential, complete, and immutable.

[0120] II. Log Record Content and Field Structure Each log node records the following: field name meaning session_id Unique identifier for the current data processing session timestamp Operation time field_i The original identifier of the field currently being processed (can be encrypted). strategy_i The desensitization strategies employed (such as mask, replace, encrypt) hash_i Hash value of this node prev_hash The hash value of the previous node (chain link) The system supports storing logs in a local secure area in JSON format, or uploading them to the blockchain (such as a blockchain platform) for non-repudiable evidence preservation.

[0121] III. Log Generation Process and Timing Control Based on the aforementioned methods and steps, the log chain generation process is as follows: 1. At the start of each data anonymization processing session, the system initializes the log chain: Python H0= SHA256(session_id + timestamp0) ``` 2. For each field processed (e.g., upon detecting the "Name" field and executing a masking strategy), the system automatically generates a corresponding log node: Python H1 = SHA256(H0 + "Zhang San" + "mask") ``` 3. And so on, all processing actions are chained together to generate a complete log chain.

[0122] 4. After the task is completed, the system will synchronize and store the log chain with the de-identified mapping structure.

[0123] IV. Audit Verification Mechanism and Usage Scenarios The log chain structure of this invention has the following technical effects and application value: 1. Immutability Since each log node contains a summary of the previous node, any tampering with any field will invalidate all subsequent hashes, making it easier for the system to automatically verify data integrity.

[0124] 2. Traceability The log chain can quickly locate processing records based on session_id and restore the processing time, strategy, and operation order of each field, supporting compliance auditing.

[0125] 3. Permission verification and recovery assistance Combined with the permission recovery mechanism (see Implementation Method 7), the log chain records can be used to determine whether the restoration conditions are met, ensuring the compliance of data recovery operations.

[0126] 4. Compliance-based evidence preservation and integration with external regulatory bodies Logs can be encrypted and stored in a local security module, or connected to an external blockchain platform to form trusted and anonymized evidence of behavior, meeting regulatory and auditing requirements.

[0127] V. Advantages of Technological Implementation This mechanism is built on the standard SHA-256 algorithm and has the following engineering feasibility and advantages: It has low computational cost and is suitable for high-concurrency processing systems; It does not depend on a specific platform or database and has independent storage and external migration capabilities; It has a simple structure and clear interfaces, making it easy to integrate with the system. Optional digital signature mechanism can be used to enhance audit non-repudiation.

[0128] Through the log chain structure and auditing mechanism described in this embodiment, the present invention significantly improves the transparency, security, and regulatory oversight of the medical data anonymization process, and technically meets the core requirements of "full-process traceability and verifiable behavior" in the current data governance environment.

[0129] Implementation Method Six: Anomaly Identification and Review Control Process (combined with...) Figure 6 ) In one embodiment of the present invention, such as Figure 6 As shown, an anomaly identification and review control mechanism is provided in the medical data anonymization process. This mechanism is used to identify situations such as insufficient confidence in field identification, strategy execution failure, or violation of risk rules, and automatically triggers the review process, suspends processing, or issues an alarm, thereby ensuring the reliability of the anonymization results and the fault tolerance of the system.

[0130] This mechanism is particularly suitable for addressing issues such as uncertainty, high-risk field identification, and declining model prediction confidence encountered in the operation of actual desensitization systems. It is an important supporting means for this invention to improve system security and desensitization quality.

[0131] I. Exception Handling Process Structure (combined with...) Figure 6 ) like Figure 6 As shown, the anomaly identification and handling process of the present invention includes the following core steps: 1. Receive model recognition results This includes field type, location information, recognition confidence P(y|x), matching strategy, etc. 2. Perform multi-condition anomaly detection. The system performs the following checks before processing each field: [A] Is the confidence level low (less than the threshold θ)? [B] Whether it hits the blacklist or sensitive word list; [C] Whether a failure occurred during the execution of the strategy (such as encryption error, mapping exception); 3. Exception Type Classification and Branching If any condition is met, the field is marked as an exception; And take the following actions based on the level of abnormality: Processing is suspended; Add to the review queue; Trigger system alarm; Write to the exception log chain node.

[0132] 4. Generate review tasks and provide manual prompts. Package the abnormal field information into a task to be reviewed (including the original field location, identification label, confidence value, and abnormality type); The system will then provide feedback to authorized users (such as data auditors) for manual confirmation.

[0133] 5. Processing of Review Results Reviewers can choose to confirm the processing, modify the label, or manually set the policy; After receiving the review result, the system continues to process the field, updates the log, and rewrites the policy path.

[0134] II. Algorithm Conditions and Mathematical Judgment Mechanism This invention determines abnormal states during field processing using the following formulas and rules: 1. Confidence threshold determination:

[0135] in The default reliability threshold is 0.85. This value can be dynamically adjusted based on model performance.

[0136] (2) Blacklist word set hit judgment: ; in The system maintains a blacklist of terms, including high-risk terms, implicitly sensitive terms, and disease codes.

[0137] (3) Strategy execution failure judgment: ; If the desensitization function returns a null value or an exception flag when processing a field, it is considered to have failed.

[0138] III. Field Review Queue Structure and Control Logic The abnormal field will be added to the system's pending review queue: ; Each review record includes field content, label prediction, confidence value, and reason for the anomaly, and also has a processing status field (unreviewed / confirmed / corrected).

[0139] The system supports centralized processing or role-based review of review fields through the management interface and retains operation logs.

[0140] IV. Improved Technical Effects and System Security The anomaly identification and review mechanism provided in this embodiment has the following technical advantages: 1. Automatic protection against identification failures. To avoid the accidental desensitization of fields with extremely low confidence or high risk, and to reduce the risk of sensitive information leakage; 2. Ensure system stability The failure of strategy execution will not lead to system crash or loss of results, demonstrating good error tolerance. 3. Supports closed-loop manual review. The system guides human intervention to handle difficult-to-determine fields, improves the accuracy of desensitization, and builds a "human-machine collaborative" processing model; 4. Strong compliance support capabilities Review logs and field statuses can be retained long-term, supporting data supervision, accountability audits, and error correction and backtracking. 5. The algorithm is adjustable and the mechanism is scalable. It supports configuring blacklists, adjusting thresholds, and customizing exception types, and has high engineering adaptability.

[0141] By implementing this method, the present invention not only improves the robustness and confidence guarantee of the de-identification results, but also effectively compensates for the problems of missed detection and misprocessing that are prone to occur in real and complex scenarios by pure model de-identification mechanisms. This demonstrates the technical innovation and practical value of the present invention in terms of compliance controllability and data quality assurance.

[0142] Implementation Method Seven: Optional Implementation Method – Permission Restoration and Policy Evolution Mechanism In a preferred embodiment of the present invention, to further meet the need for controllable restoration of medical data in specific scenarios and to improve the strategy adaptability and intelligence level of the system during long-term operation, the following two optional extension mechanisms are proposed: 1. Access Control Restoration Mechanism: Used to reversibly restore de-identified fields when access authorization conditions are met; 2. Strategy Evolution Mechanism: Used to dynamically optimize the strategy decision function based on user feedback and historical execution results, thereby improving the accuracy of strategy matching.

[0143] I. Permission Restoration Mechanism This invention provides a multi-dimensional configurable permission restoration control mechanism to ensure that authorization for sensitive fields is restored within a legal and compliant scope, such as when doctors view data within the hospital, when regulatory audits query data, or when data is debugged within a secure window.

[0144] (1) Design of recovery function Define the recovery control function: ;

[0145] The parameters have the following meanings: U: User identification (such as UserID, role level, etc.); T: Access request timestamp; P: Business usage scenario tag (e.g., in-hospital access, clinical emergency care, regulatory inspection, etc.); The system will perform a permission verification process before the field is restored, only in f recover Only when the value is 1 can the de-identification mapping table be called to restore the fields.

[0146] (2) Multi-level permission matrix The recovery control mechanism supports configuring multi-level access control rules. Common configurations are as follows: Role Scene limitations Time window control Restore permissions Attending physician in the hospital Patient data display Mission day ± 1 day allow Researchers Batch desensitization data analysis none prohibit Auditor Risk control and audit scenarios All time period allow The system supports storing this permission matrix as a JSON / YAML format rule file and provides an interface for administrators to update it.

[0147] (3) The recovery process is linked with the audit. The field restoration operation will be automatically written to the de-identified log chain, including the restored user ID, operation time, field ID, and restoration reason, to ensure that the restoration process is auditable and non-repudiable.

[0148] II. Strategy Evolution Mechanism To avoid the failure, rigidity, or mismatch problems of static policy matching mechanisms in long-term operation, this invention introduces a policy evolution mechanism, which enables the policy decision function to be continuously optimized over time, with data and feedback.

[0149] (1) Policy feedback data caching The system establishes a policy execution feedback cache area to record the following data: Field recognition and policy decision records (field type, selection policy, sensitivity level, etc.); Feedback on actual processing effects (such as whether desensitization is incorrect, manual adjustment by the user, etc.); Model or user review suggestions (such as recommended replacement with encryption, masking ineffectiveness, etc.); This cache provides the training data basis for policy evolution.

[0150] (2) Evolution model design Optional evolution models include: Incremental update mechanism based on decision tree According to the feedback data, dynamically modify the policy matching path. For example, change the original "L = medium, U = display" from replace to mask; Policy update mechanism based on reinforcement learning Regarding the desensitization policy as the action space, set the policy reward function (such as "recognition accuracy × data availability × user satisfaction"), and update the policy function through the policy optimization algorithm.

[0151] (3) Policy evolution process The execution process of the evolution mechanism is as follows: 1. Collect policy failure records and user feedback; 2. Train / update the evolution model; 3. Output suggestions for updating the policy function; 4. The administrator confirms and executes the update; 5. Automatically synchronize the update to the policy matching module.

[0152] The system also supports simulating and testing the evolution effect to avoid policy drift caused by incorrect learning.

[0153] III. Technical effect description By introducing the permission restoration and policy evolution mechanisms, the present invention has the following advanced functional advantages: Functional modules Technical effect Access Control To achieve secure and reversible data handling for sensitive data and meet the needs of special scenarios. Multidimensional verification logic Reduce risks such as unauthorized access and misuse of data, and improve compliance. Self-learning evolution mechanism Enhance the adaptability of strategy decisions and support collaborative optimization of data and rules. Log-linked auditing Ensure that recovery behaviors and strategy evolution are traceable and verifiable. In summary, the permission restoration and policy evolution mechanisms are effective supplements to the intelligence, controllability, and life cycle maintenance capabilities of the system of the present invention, further enhancing the practicality, security, and sustainable evolution capabilities of the medical data desensitization platform.

[0154] In summary, this invention, by integrating natural language processing models, multi-source policy matching logic, and audit log chain mechanisms, constructs a medical data desensitization and processing system that combines accuracy, compliance, and controllability. It can effectively identify privacy fields in electronic medical records and implement differentiated desensitization strategies based on access roles and data usage. Furthermore, by incorporating anomaly handling, permission restoration, and policy evolution mechanisms, it achieves adaptive support for complex application scenarios, demonstrating good engineering practicality and promotional value.

[0155] It should be understood that the embodiments described above are merely preferred specific examples of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, improvements, or equivalent solutions made in accordance with the concept of the present invention without departing from the original design idea of ​​the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for de-identifying and processing medical data based on artificial intelligence, characterized in that, Includes the following steps: S1. Control the data access module to obtain raw medical data and complete the format parsing, converting unstructured or semi-structured data into a structured field sequence; S2. Integrate the pre-trained language model and the sequence labeling model to perform privacy field identification on the field sequence and obtain a labeling mapping that includes field type, location information and identification confidence. S3. Match a de-identification strategy based on the sensitivity level of the field, the user access role, and the purpose of data use. The de-identification strategy includes at least one of character masking, forgery replacement, and encryption processing. S4. Execute the matching de-identification strategy to generate de-identified data; S5. Generate processing logs and establish a mapping structure between original fields and de-identified fields; S6. When the recognition confidence level is lower than the preset threshold, the field hits an abnormal word set, or the strategy execution fails, the exception handling mechanism is triggered, the field is added to the list to be reviewed, and an alarm message is sent.

2. The method according to claim 1, characterized in that, The privacy field identification model includes: Pre-trained language models are used to generate word vector representations; Bidirectional Long Short-Term Memory (BiLSTM) neural network is used to extract contextual semantic features; Conditional Random Fields (CRFs) are used to predict the optimal label path. The rules engine is used to supplement the identification of sensitive fields with obvious structured characteristics.

3. The method according to claim 1, characterized in that, The selection of the desensitization strategy is based on the field sensitivity level L, user access role R, and data purpose U, satisfying the following strategy matching function: Specifically, when L is high-level and R is a researcher, the character masking strategy is preferred; when L is medium-level and U is for display, the forgery and replacement strategy is selected; when L is high-level and U is for reporting purposes, the encryption strategy is selected.

4. The method according to claim 1, characterized in that, The exception handling mechanism includes any one or more of the following: Determine whether the confidence level of the identification is lower than a preset threshold; Determine if a field matches the blacklist of words; Determine whether the strategy execution failed; When any condition is met, the corresponding field is marked as abnormal, added to the review queue, and an alarm message is generated.

5. The method according to claim 1, characterized in that, The process before performing privacy field identification further includes: Perform field integrity checks to determine if the input data contains missing fields, incorrect formats, or illegal symbols; If any anomalies are found, the desensitization process will be stopped and the user will be prompted to correct the data.

6. The method according to claim 1, characterized in that, It also includes a strategy evolution module, used for: Record historical anonymization strategies and their effectiveness; Based on user feedback and failure samples, the strategy selection rules are adjusted through machine learning models, supporting strategy self-learning and optimization.

7. A medical data desensitization identification and processing system for implementing the method of claim 1, characterized in that, include: The data access and parsing module is used to receive raw medical data and standardize its format. The privacy field recognition module is used to perform named entity recognition based on an artificial intelligence model, and outputs the field type, location information and confidence level. The strategy matching module is used to determine the de-identification strategy based on the field sensitivity level, user role, and purpose. The desensitization execution module is used to perform desensitization processing according to the matching results; The logging module is used to generate de-identified logs and field mapping structures; The exception handling module is used to handle low-confidence or rule-based exception fields and trigger the review process.

8. The system according to claim 7, characterized in that, The log recording module constructs a de-identified hash chain, generating the log chain in the following manner: This ensures that the desensitization process is unalterable and traceable.

9. The system according to claim 7, characterized in that, It also includes a permission recovery mechanism to restore the original value of a field when the following conditions are met: Verify that the user's identity level, access scenario tag, and time window satisfy the following rule function: Where U is the user ID, T is the timestamp, and P is the usage scenario tag.

10. A medical data anonymization identification and processing device deployed on a server or terminal device, characterized in that, include: Storage unit, used to store model parameters, strategy configuration, and de-identification mapping information; The network interface module is used to receive raw medical data input; The processor is configured to perform the steps of the method described in any one of claims 1 to 6 and generate de-identification results and audit records.

Images