Pet diagnosis and treatment data sharing method and system

Abstract

The application discloses a pet diagnosis and treatment data sharing method and system, relates to the technical field of pet medical information processing and data sharing, and comprises the following steps: acquiring pet diagnosis and treatment original data stored locally by each participating institution and performing standardization processing, extracting pet identity information, treatment event information, symptom information, examination information, diagnosis information, medication information and outcome information to generate a clinical semantic anchor set, further generating anchor fingerprints and local anchor association graphs, completing cross-institution semantic alignment to obtain pet diagnosis and treatment semantic units, and performing local training, privacy disturbance, parameter encryption and platform aggregation distribution based on the pet diagnosis and treatment semantic units; so that the safe sharing, semantic consistent utilization and reliable output of cross-institution diagnosis and treatment data are realized without sharing the content of pet diagnosis and treatment original data.

Description

Technical Field

[0001] This invention relates to the field of pet medical information processing and data sharing technology, and more specifically, to a method and system for sharing pet medical data. Background Technology

[0002] Currently, pet medical data sharing mainly occurs between pet clinics, referral hospitals, testing institutions, imaging institutions, and regional collaboration platforms. Although the industry has gradually adopted information technology methods such as electronic medical records, digital imaging, laboratory information management, and remote collaboration, there are still significant differences among different institutions in pet identification methods, disease naming habits, record standards for different stages of treatment, presentation of examination results, and organization of medication information. Data generated for the same pet at different institutions and at different stages of treatment often exhibits unstable object identification, inconsistent clinical semantic granularity, and a lack of contextual relationships. Furthermore, even if cross-institutional data transfer or parameter-level sharing is achieved through technologies such as interface integration, data anonymization, access control, federated learning, and differential privacy, it is still difficult to ensure that the data processed and utilized by each participant accurately corresponds to the same patient, the same stage of the disease, and the same treatment facts in clinical meaning. This can easily lead to shared data that is transmissible but difficult to interpret, aggregated but difficult to reuse, and modelable but difficult to ensure semantic consistency during training. Ultimately, this results in distorted, mismatched, or unreliable shared results in scenarios such as remote diagnosis, joint analysis, disease early warning, and model training.

[0003] Therefore, the problem is that cross-institutional pet medical data lacks a stable, unified, and computable clinical semantic alignment mechanism before sharing, which means that even if privacy protection and sharing are completed, the data cannot be used securely, accurately, and effectively. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides a method and system for sharing pet medical data, addressing the problems of prominent data silos, inconsistent data standards, difficulty in accurately corresponding clinical semantics, difficulty in balancing privacy protection and data utilization, and difficulty in reusing shared results stably during the cross-institutional sharing of existing pet medical data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The method for sharing pet medical data includes the following steps: S1. Obtain the original pet medical data stored locally by each participating institution. Based on the unified data dictionary, perform field extraction, field name matching, data type conversion, missing item marking, time format unification, and treatment time sorting on the original pet medical data, and output the standardized medical dataset corresponding to each participating institution. S2. Input the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process, extract pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information from the standardized diagnosis and treatment datasets, and generate a set of clinical semantic anchors according to the information categories; S3. Perform local encoding, anchor relationship binding and irreversible transformation processing on each clinical semantic anchor in the clinical semantic anchor set of each participating institution to generate anchor fingerprints that correspond one-to-one with each clinical semantic anchor, and construct a local anchor association graph based on the order of visits, examination and diagnosis and diagnosis and medication relationship between each clinical semantic anchor. S4. Input the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. Perform cross-institutional matching calculations according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical events, and the contextual correspondence of examination information and medication information. Output a unified cross-institutional pet medical semantic unit. S5. Input the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process. Perform model training, privacy perturbation processing and parameter encryption processing on the local basis of the pet diagnosis and treatment semantic units in each participating institution, and output encrypted shared parameters that do not contain the original pet diagnosis and treatment data. S6. Upload the encrypted shared parameters output by each participating institution to the sharing platform. The sharing platform performs aggregation calculation, access control verification, security audit recording, and sharing result distribution processing on the encrypted shared parameters, and outputs shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training.

[0006] A pet medical data sharing system for performing the method, including: The data standardization module acquires the original pet medical data stored locally by each participating institution, and performs field extraction, field name matching, data type conversion, missing item marking, time format standardization, and treatment time sorting on the original pet medical data according to the unified data dictionary, and outputs the standardized medical dataset corresponding to each participating institution. The local semantic parsing module inputs the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process, extracts pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information from the standardized diagnosis and treatment datasets, and generates a set of clinical semantic anchors according to the information categories; The anchor coding and association module performs local encoding, anchor relationship binding and irreversible transformation processing on each clinical semantic anchor in the clinical semantic anchor set of each participating institution, generates anchor fingerprints that correspond one-to-one with each clinical semantic anchor, and constructs a local anchor association graph based on the order of visits, examination and diagnosis and diagnosis and medication relationship between each clinical semantic anchor. The cross-institutional semantic alignment module inputs the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. It performs cross-institutional matching calculations according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical events, and the contextual correspondence between examination information and medication information, and outputs a unified cross-institutional pet medical semantic unit. The local shared training module inputs the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process. Each participating institution performs model training, privacy perturbation processing, and parameter encryption processing based on the pet diagnosis and treatment semantic units, and outputs encrypted shared parameters that do not contain the original pet diagnosis and treatment data. The shared platform processing module uploads the encrypted shared parameters output by each participating institution to the shared platform. The shared platform then performs aggregation calculations, access control verification, security audit logging, and shared result distribution on the encrypted shared parameters, outputting shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training.

[0007] The technical effects and advantages of this invention are as follows: 1. First, perform standardized processing driven by a unified data dictionary on the original pet medical data of each participating institution. Then, extract clinical semantic anchors and construct anchor fingerprints and local anchor association graphs. Subsequently, without directly uploading the original pet medical data, complete local shared training, parameter encrypted upload, and platform aggregation and distribution. This can address the problems of prominent data silos, heterogeneous data formats, inconsistent record quality, insufficient privacy protection, and low sharing efficiency in the existing pet medical data sharing process. It establishes a coherent processing link between institutions from data processing and semantic organization to shared application, thereby improving the consistency, security, and availability of pet medical data circulation across institutions, while taking into account data utilization needs and compliant sharing requirements. 2. First, standardized medical data is converted into clinical semantic anchors. Then, cross-institutional semantic alignment is performed through anchor fingerprints and local anchor association graphs. Based on this, a unified cross-institutional pet medical semantic unit is generated. This addresses the core technical problem in existing technologies where, even after data anonymization, parameter sharing, or platform interconnection, it is still difficult to ensure that different institutions form a unified semantic expression for the same pet, the same medical event, and the same medical facts. It reduces the risk of mismatch in cross-institutional medical records in terms of identity recognition, time sequence, event correspondence, and examination, diagnosis, and medication context. This improves the semantic consistency, result reliability, and reuse accuracy of shared data in subsequent shared training, remote diagnosis, disease early warning, and collaborative analysis. Attached Figure Description

[0008] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0009] 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.

[0010] Refer to the instruction manual appendix Figure 1 The method for sharing pet medical data includes the following steps: S1. Obtain the original pet medical data stored locally by each participating institution. Based on the unified data dictionary, perform field extraction, field name matching, data type conversion, missing item marking, time format unification, and treatment time sorting on the original pet medical data, and output the standardized medical dataset corresponding to each participating institution. In the process of sharing pet medical data, the raw pet medical data stored locally by participating institutions usually comes from different business systems. Not only are field names different, but the same medical information may also differ in field data type, field value range, and field temporal semantics. This makes it difficult to directly form a unified, computable, standardized medical dataset. If field extraction, field attribution, and standardization correction are not completed before inputting into the local semantic parsing process, the subsequent generation of clinical semantic anchor sets is easily based on unstable data, thus affecting the accuracy of cross-institutional semantic alignment and shared training. Therefore, this implementation method performs field extraction, field attribution, standardization correction, and optimal merging processing on the raw pet medical data using a unified data dictionary to obtain standardized medical datasets for each participating institution. This implementation process includes the following steps: In S1.1, when reading the pet medical treatment raw data stored locally by each participating institution and matching it with the standard field names, field data types, field value ranges, and field time semantics in the unified data dictionary, the process first reads the raw fields in the pet medical treatment raw data according to the participating institution, business system, data table, record number, and field position, and simultaneously reads the field names, field meanings, field data types, field value ranges, and field time semantics corresponding to each standard field in the unified data dictionary. Then, for each raw field, the similarity of field names, consistency of field data types, matching degree of field value ranges, and co-occurrence relationships between the raw field and each standard field are calculated. Specifically, field name similarity is obtained by calculating the character overlap ratio and word mapping results after character normalization of the raw field names and standard field names; field data type consistency is obtained by judging whether the current value form of the raw field is consistent with the data type specified by the standard field; and field value range matching... The degree of field co-occurrence is obtained by statistically analyzing the proportion of values ​​of the original field in the current participating institution's historical records that fall within the corresponding field's value range. The field co-occurrence relationship is obtained by statistically analyzing the frequency of co-occurrence of the original field with other known original fields in the same record and comparing it with the field combination relationship in the unified data dictionary. After completing the above four calculations, candidate results for each original field are generated for multiple standard fields. The field assignment result is determined according to the weighted result of field name similarity, field data type consistency, field value range matching degree, and field co-occurrence relationship. When the highest weighted result is higher than the assignment threshold, the corresponding standard field is determined as the field assignment result of the original field. When multiple candidate results are close, they are retained as multiple candidate states. Finally, candidate standardized data fragments with field source identifiers and field assignment results are output. The field source identifiers include at least the participating institution identifier, business system identifier, data table identifier, and original record identifier, so that they can be called in the subsequent standardization correction process. In S1.2, after the candidate standardized data fragments are input into the standardization correction process, data type conversion, missing item marking, time format unification, and appointment time sorting are performed on each candidate standardized data fragment according to the field name correspondence results. Specifically, data type conversion converts character, numeric, date, or enumeration values ​​to a unified form based on the field data type corresponding to the field assignment results; missing item marking writes missing status for fields that should appear in the unified data dictionary but were not successfully assigned to the current candidate standardized data fragment; time format unification converts different time representations to a unified time format based on the field time semantics; and appointment time sorting arranges the candidate standardized data fragments sequentially according to the appointment time of the same pet within the same participating institution. After completing the above processing, for cases where multiple... Data fragments for each candidate field are then selected and merged based on field source stability, temporal correlation integrity, and cross-record consistency. Field source stability is determined by statistically analyzing the continuous usage ratio of the same field source identifier in the historical records of the current participating institutions. Temporal correlation integrity is determined by judging whether the time corresponding to the candidate field satisfies a reasonable chronological relationship with other time fields in the same record. Cross-record consistency is determined by comparing the consistency of the same field value in multiple records for the same pet, the same medical event, or the same record group. When multiple candidate fields exist simultaneously, the candidate field with higher field source stability, better temporal correlation integrity, and higher cross-record consistency is prioritized and retained as the final standard field value. The remaining candidate fields are retained as source references, and the final output is the standardized diagnosis and treatment dataset corresponding to each participating institution. Through the above implementation method, the raw pet medical data stored locally by each participating institution can first undergo field extraction, field attribution, standardization correction, and optimal merging under the constraints of a unified data dictionary. This results in a standardized medical dataset with unified field names, unified field data types, verifiable field value ranges, clear field time semantics, and consistent consultation time order. This provides a stable data input foundation for the subsequent generation of clinical semantic anchor sets and reduces subsequent processing deviations caused by field heterogeneity, record missingness, and time confusion. In practical applications: the outpatient management system of a participating institution records consultation time as outpatient time, and the laboratory system records relevant time as submission time. The system first calculates the similarity of field names, consistency of field data types, matching degree of field value range, and co-occurrence relationship between these two raw fields and the standard fields consultation time and sampling time, respectively, based on the unified data dictionary. This yields corresponding candidate standardized data fragments. Then, the candidate standardized data fragments undergo data type conversion, missing item marking, time format unification, consultation time sorting, and optimal merging processes to finally form the standardized medical dataset corresponding to the participating institution, which can be directly called by the subsequent local semantic parsing process. S2. Input the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process, extract pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information from the standardized diagnosis and treatment datasets, and generate a set of clinical semantic anchors according to the information categories; In the process of sharing pet medical data, even though participating institutions have formed corresponding standardized medical datasets, there are still problems such as the same medical facts being scattered across different records, different fields carrying different clinical meanings, and the same clinical meaning appearing repeatedly in different records. If it is not possible to stably extract pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information, and outcome information from the standardized medical dataset, and organize this information into a set of clinical semantic anchors that can be directly entered into the subsequent local encoding and cross-institutional semantic alignment process, then subsequent processing is prone to semantic fragmentation, interruption of context connection, and inconsistent category attribution. Therefore, this implementation method first completes information extraction and category attribution calculation through a local semantic parsing process, and then completes fragment merging, conflict resolution, and category binding through a semantic anchor generation process to generate a set of information categorized clinical semantic anchors corresponding to pet identity, medical event, symptoms, examination, diagnosis, medication, and outcome, respectively. This implementation process includes the following steps: In S2.1, when inputting the standardized medical datasets corresponding to each participating institution into the local semantic parsing process, the standardized medical datasets are first read sequentially according to record number, field position, generation time, and the corresponding medical record. The pet basic information field, medical record field, symptom description field, examination result field, diagnosis record field, medication record field, and outcome record field in each standardized record are marked as fields to be parsed. Subsequently, based on the value of each field in the field group, information extraction and category assignment calculation are performed according to field position relationship, temporal adjacency relationship, and record co-occurrence relationship. Among these, the field position relationship is used to determine the structural position of a field in the current record and its sequential arrangement with adjacent fields; the temporal adjacency relationship is used to determine the temporal proximity between the corresponding content of a field and adjacent time fields. Co-occurrence relationships are used to determine the frequency of co-occurrence and combination stability of different fields in the same record, sub-record, or record group. After calculating the above three types of relationships, a category classification score is generated for each field value, corresponding to pet identity information, medical visit information, symptom information, examination information, diagnosis information, medication information, and outcome information. The category with the highest score that is higher than the classification threshold is taken as the category classification result of the field value. If the scores of multiple categories are close, the field value is retained as a multi-category candidate state. Finally, a set of candidate semantic fragments corresponding to pet identity information, medical visit information, symptom information, examination information, diagnosis information, medication information, and outcome information is output. Each candidate semantic fragment includes at least the fragment content, source record number, source field number, category classification result, generation time, and the medical visit record identifier to which it belongs. In S2.2, after inputting the candidate semantic fragment set into the semantic anchor generation process, the set is first merged according to the information association within the same medical event, and then conflict resolution and category binding are performed according to the temporal sequence between different medical events. The information association within the same medical event refers to the correspondence between candidate semantic fragments formed within the same medical record identifier, the same consultation time range, or the same treatment process chain, which can be determined by comparing the medical record identifier, time difference, and record co-occurrence results of the candidate semantic fragments. The temporal sequence between different medical events refers to the temporal order and clinical continuity between candidate semantic fragments formed during multiple medical visits for the same pet, which can be determined by the pet's identity information correspondence, the order of previous and subsequent medical visits, and the referral, re-visit, re-examination, or follow-up markers between the previous and next medical events. During the fragment merging process, candidate semantic fragments with identical or complementary content are merged. During the conflict resolution process... During the process, candidate semantic segments with inconsistent content within the same category are prioritized and compared based on temporal relevance, source integrity, and contextual closure. Temporal relevance is determined by comparing the generation time of the corresponding record; source integrity is determined by comparing whether the candidate semantic segment simultaneously possesses the source record number, source field number, and the identifier of the medical record to which it belongs; and contextual closure is determined by comparing whether the candidate semantic segment can form a complete diagnostic and treatment context with other candidate semantic segments within the same medical event. In the category binding process, the candidate semantic segments after segment merging and conflict resolution are bound to information categorized clinical semantic anchors for pet identity, medical event, symptoms, examination, diagnosis, medication, and outcome, respectively. Segments corresponding to the same diagnostic and treatment fact within the same category are merged into the same clinical semantic anchor. Finally, information categorized clinical semantic anchor sets corresponding to pet identity, medical event, symptoms, examination, diagnosis, medication, and outcome are generated, and these sets are used as input for subsequent local encoding processes. Through the above implementation method, pet identity information, visit event information, symptom information, examination information, diagnosis information, medication information, and outcome information can be stably extracted from standardized medical datasets. Then, through fragment merging, conflict resolution, and category binding, a set of categorized clinical semantic anchors is formed. This organizes medical information scattered across different fields, records, and stages of treatment into unified expression units with clear semantic boundaries, well-defined contextual relationships, and direct participation in subsequent local encoding and cross-institutional semantic alignment. This reduces subsequent processing biases caused by information fragmentation, unstable category attribution, and unclear connection between records. In practical applications: the standard of a participating institution... The clinical diagnosis dataset contains one outpatient visit record and one follow-up visit record for the same pet. The system first extracts the fields related to chief complaint, test, diagnosis, medication, and outcome into corresponding candidate semantic fragments in the local semantic parsing process. Then, based on the information association relationship within the same medical event, the relevant candidate semantic fragments of the outpatient stage are merged. Based on the time sequence relationship between different medical events, the outcome information of the follow-up visit stage is connected with the previous medical process. Finally, the corresponding clinical semantic anchors for pet identity, medical event, symptoms, examination, diagnosis, medication, and outcome are generated and can be used in subsequent processes. S3. Perform local encoding, anchor relationship binding and irreversible transformation processing on each clinical semantic anchor in the clinical semantic anchor set of each participating institution to generate anchor fingerprints that correspond one-to-one with each clinical semantic anchor, and construct a local anchor association graph based on the order of visits, examination and diagnosis and diagnosis and medication relationship between each clinical semantic anchor. In the process of sharing pet medical data, although the clinical semantic anchor set has organized the scattered medical information into categorized expressions such as pet identity, medical visit events, symptoms, examinations, diagnoses, medications, and outcomes, if these clinical semantic anchors remain at the level of raw text or ordinary structured fragments, it is difficult for different participating institutions to achieve stable comparison and continuous connection without exposing the original medical content. Therefore, this implementation method first performs local encoding, relation binding, and irreversible transformation processing on each clinical semantic anchor to obtain anchor fingerprints that correspond one-to-one with each clinical semantic anchor. Then, node registration, relation determination, relation filtering, and graph structure construction are performed around the anchor fingerprints to generate a local anchor association graph containing medical visit order relations, examination and diagnosis relations, and diagnosis and medication relations. This provides a local relation expression that is comparable without directly exposing the original content for subsequent cross-institutional semantic alignment. The implementation process includes the following steps: In S3.1, when inputting each clinical semantic anchor from the clinical semantic anchor set of each participating institution into the local encoding process, the information category identifier, field content, time position, and medical event affiliation corresponding to each clinical semantic anchor are read one by one. The information category identifier identifies which category the current clinical semantic anchor belongs to: pet identity, medical event, symptoms, examination, diagnosis, medication, or outcome. The field content records the specific standardized content carried by the clinical semantic anchor. The time position records the chronological position of the clinical semantic anchor within its respective medical process. The medical event affiliation records the medical event number corresponding to the clinical semantic anchor. Subsequently... Fixed-length coding and category bit concatenation are performed according to information category, field order, time sequence, and medical event attribution. Specifically, fixed category coding bits are first assigned to different information categories. Then, positional coding is performed on field content according to a unified field order, sequential coding is performed on time position, and event number coding is performed on medical event attribution. The coding length of each part is fixed in advance and kept consistent within the same participating institution. After completing the fixed-length coding of each part, the category coding bits, field content coding bits, time position coding bits, and medical event attribution coding bits are concatenated in sequence to form an initial coding fragment that corresponds one-to-one with the current clinical semantic anchor. In S3.2, after the initial encoded fragment is input into the anchor relationship binding process, the co-occurrence relationship of the same visit, the sequential relationship of visits before and after visits, and the contextual relationship between examination information, diagnosis information, and medication information are calculated based on the attribution and time position of the visit events of each clinical semantic anchor within the same participating institution. Specifically, the co-occurrence relationship of the same visit is obtained by determining whether the visit event attribution of two clinical semantic anchors is the same, whether the time position is within the same visit window, and whether the field content appears together within the same record group. The sequential relationship of visits before and after visits is obtained by comparing the visits to which the two clinical semantic anchors belong. The sequence of events, the pet's identity, and whether there are follow-up, re-examination, or follow-up markers between the previous and subsequent visits are used to determine the contextual relationship between examination information, diagnosis information, and medication information. This is achieved by comparing the temporal position order, record co-occurrence, and field content association of the three types of clinical semantic anchors within the same visit. After the determination is completed, a corresponding relationship number is assigned to each relationship, and the relationship number is written into the relationship marker bit of the corresponding initial coding segment. At the same time, the association bits reflecting the relationship direction and associated objects are bound to the corresponding positions to form an associated coding segment with relationship binding markers. In S3.3, after the associated coded fragment is input into the irreversible transformation process, the associated coded fragment is first irreversibly mapped according to the preset one-way transformation rules, and then a fixed-length truncation process is performed to generate anchor fingerprints that correspond one-to-one with each clinical semantic anchor. The preset one-way transformation rules can be implemented using one-way hash mapping rules. Specifically, the associated coded fragment is written as an input string into a one-way transformation function, and the irreversible mapping result is output. To ensure consistent length within the same participating institution and facilitate subsequent comparison, a fixed-length truncation process is performed on the mapping result, and a result fragment of a preset length is truncated as the anchor fingerprint. After completing the irreversible mapping and fixed-length truncation, each associated coded fragment generates a unique anchor fingerprint, and a one-to-one correspondence is established between the anchor fingerprint and the original clinical semantic anchor. This correspondence is only stored locally and is not uploaded with the sharing process. In S3.4, when registering anchor fingerprints generated by participating institutions according to the corresponding clinical semantic anchors and their corresponding medical events and time sequences, the anchor fingerprints are first used as node identifiers. All anchor fingerprints within the same participating institution are grouped according to their respective medical event numbers, and within each group, they are arranged from earliest to latest according to their time position, forming a node registration sequence. After node registration is completed, consecutive anchor fingerprints in adjacent medical events are subjected to a succession judgment. If the conditions are met, a succession candidate relationship is written. At the same time, the anchor fingerprints corresponding to examination information and the anchor fingerprints corresponding to diagnosis information in the same medical event are subjected to an examination-diagnosis correspondence judgment. If the conditions are met, an examination-diagnosis candidate relationship is written. Then, the anchor fingerprints corresponding to diagnosis information and the anchor fingerprints corresponding to medication information are subjected to a diagnosis-medication correspondence judgment. If the conditions are met, a diagnosis-medication candidate relationship is written. Finally, a set of candidate relationships between anchors is output, where each candidate relationship includes at least the starting anchor fingerprint, the target anchor fingerprint, the relationship type, and the relationship generation time. In S3.5, after inputting the set of candidate anchor relationships into the relationship filtering process, the process first performs relationship retention and conflict resolution based on whether the time difference falls within the scope of the same medical visit, whether the information category combination satisfies the correspondence between examination and diagnosis or diagnosis and medication, and whether the same anchor fingerprint forms a conflict in multiple candidate relationships. Specifically, whether the time difference falls within the scope of the same medical visit is determined by comparing the difference between the time positions corresponding to the anchor fingerprints at both ends of the candidate relationship with the preset medical visit window. When the time difference is less than or equal to the preset medical visit window, the candidate relationship for the same medical visit is retained; otherwise, it is discarded. Whether the information category combination satisfies the correspondence between examination and diagnosis or diagnosis and medication is determined by comparing... The clinical semantic anchor categories corresponding to the anchor fingerprints at both ends of the candidate relationship are obtained. Whether the same anchor fingerprint conflicts in multiple candidate relationships is determined by statistically analyzing whether the same anchor fingerprint points to multiple target anchor fingerprints or is pointed to by multiple starting anchor fingerprints under the same relationship type. When multiple candidate relationships compete, priority comparison is performed according to the principles of smaller time difference, stronger contextual continuity, and more complete source record chain. Only the candidate relationship with the highest priority is retained, and the other candidate relationships are eliminated. After the above screening, a set of valid anchor relationships is output, in which each retained relationship meets the requirements of time range, category combination, and relationship uniqueness. In S3.6, when connecting the set of valid anchor relationships with each anchor fingerprint using nodes and edges, each anchor fingerprint is first used as a graph node, and then each valid relationship in the set of valid anchor relationships is used as a graph edge. Each edge is written into the graph structure according to the starting anchor fingerprint, the target anchor fingerprint, and the relationship type. Specifically, sequential relationships correspond to the order of medical visits, examination and diagnosis relationships correspond to the examination and diagnosis relationship, and diagnosis and medication relationships correspond to the diagnosis and medication relationship. During the writing process, the event number and time position of each node are saved, and the relationship type and relationship generation order are saved for each edge, so that they can be called in subsequent cross-institutional semantic alignment processes. After all nodes and edges are written, a local anchor association graph containing the order of medical visits, examination and diagnosis, and diagnosis and medication relationships is generated. Figure 1 On the one hand, it preserves the temporal and contextual relationships between clinical semantic anchors within the same participating institution; on the other hand, together with the anchor fingerprint, it constitutes the local basic input for subsequent cross-institutional matching calculations. Through the above implementation method, each clinical semantic anchor in the clinical semantic anchor set can be converted into an anchor fingerprint with a uniform structure, fixed length, and inability to directly deduce the original content. Then, a local anchor association graph containing the relationships of visit order, examination and diagnosis, and diagnosis and medication can be constructed around the anchor fingerprint. This allows participating institutions to retain key contextual relationships within the same visit and between previous and subsequent visits without directly exposing the original expressions of field content, time location, and visit event attribution. This provides a stable, continuous, and computable basic input for subsequent cross-institutional semantic alignment and reduces the risk of sharing and matching bias caused by exposure of original content or breakage of relationships. In practical application: a participating institution generated examination-related clinical semantics for a single outpatient record of the same pet. For anchors, diagnostic clinical semantic anchors, and medication clinical semantic anchors, the system first reads the information category identifier, field content, time position, and visit event affiliation corresponding to these clinical semantic anchors, performs fixed-length encoding and category bit concatenation processing to generate initial encoding fragments, and then completes the relationship number writing and association bit binding processing based on the co-occurrence relationship of the same visit and the contextual pointing relationship between examination, diagnosis, and medication. Subsequently, the system generates corresponding anchor fingerprints through one-way transformation and fixed-length truncation. After that, these anchor fingerprints are registered as nodes according to the visit event and time sequence, the examination-diagnosis correspondence relationship and the diagnosis-medication correspondence relationship are determined, and after relationship filtering, node edges are connected to finally generate the local anchor association graph of the participating institution, which can be directly called by the subsequent cross-institutional semantic alignment process. S4. Input the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. Perform cross-institutional matching calculations according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical events, and the contextual correspondence of examination information and medication information. Output a unified cross-institutional pet medical semantic unit. In the process of sharing pet medical data, although participating institutions have formed anchor fingerprints and local anchor association graphs, there may still be issues such as time misalignment, category mismatch, and context breakage in the expression of the same pet, the same medical event, and the same medical facts among different participating institutions. If anchor fingerprints and their relationship structures cannot be uniformly matched, filtered, merged, and registered at the cross-institutional level, even if each participating institution has completed local encoding and local relationship construction, it will still be difficult to obtain a unified pet medical semantic unit across institutions, thus affecting the accuracy and stability of subsequent local sharing training and sharing platform processing. Therefore, this implementation method establishes a cross-institutional semantic alignment process based on anchor fingerprints and local anchor association graphs. First, a candidate matching set between anchors is formed among participating institutions, and then effective matching filtering, semantic merging, semantic unit generation, conflict resolution, and unified registration are gradually completed to output a unified pet medical semantic unit across institutions. This implementation process includes the following steps: In S4.1, when inputting the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process, the clinical semantic anchor category, the corresponding medical event number, the time position, and the node connection relationship in the local anchor association graph are first read according to the participating institution's identifier. Then, taking any participating institution's anchor fingerprint as the current object to be matched, identity matching calculation, time continuity calculation, event association calculation, and context consistency calculation are performed one by one on anchor fingerprints with the same or corresponding clinical semantic anchor categories from other participating institutions. Among them, identity matching calculation is used to determine whether two anchor fingerprints may correspond to the same pet identity information main line; time continuity calculation is used to determine whether the medical events of the two anchor fingerprints have a sequential or synchronous correspondence relationship between different participating institutions; and event association calculation is used to determine whether the two anchor fingerprints have a sequential or synchronous correspondence relationship between different participating institutions. The internal relationship structure of the medical treatment events where the anchor fingerprints are located is similar. Context consistency calculation is used to determine the degree of contextual correspondence between examination information and medication information across different participating institutions. After completing the above four types of calculations, a comprehensive matching score is generated for each pair of candidate anchor fingerprints. The comprehensive matching score can be obtained by multiplying the identity matching calculation result, time continuity calculation result, event association calculation result, and context consistency calculation result by their respective weights and then summing them. The sum of each weight is set to 1. The specific weights can be determined based on the statistical analysis of correct cross-institutional matching results in historical review samples. When the comprehensive matching score is higher than the candidate retention threshold, the pair of anchor fingerprints is written into the anchor candidate matching set between each participating institution, and the starting participating institution identifier, the target participating institution identifier, the matching anchor fingerprint pair, the calculation results of each item, and the comprehensive matching score are recorded. In S4.2, after inputting the set of candidate anchor matches into the matching and filtering process, the first round of filtering is performed according to the order of visit time under the same pet identity. Specifically, only candidate matches that meet the reasonable time order under the correspondence of the same pet identity information are retained. Then, the second round of filtering is performed according to the matching constraints of information categories under the same visit event. Specifically, it checks whether a candidate match has a matching category relationship that matches the current clinical semantic anchor category within the visit event. After that, the third round of filtering is performed according to the consistency constraints of the relationship between cross-institutional anchor relationships and the local anchor association graph. Specifically, it compares whether the predecessor, successor, and relationship types of the anchor fingerprints at both ends of the candidate match are consistent in their respective local anchor association graphs. For candidate matches that have passed the above three rounds of filtering but still have one-to-many, many-to-one, or many-to-many competition, conflict match elimination is performed. Specifically, the only retained result is determined according to the principle of higher comprehensive matching score, stronger relationship consistency, and more stable time order, and the remaining conflicting candidate matches are eliminated. Finally, the set of effective cross-institutional matches is output, where each effective match includes at least the participating institution pair, anchor fingerprint pair, clinical semantic anchor category, matching relationship type, and retention priority. In S4.3, after inputting the cross-institutional valid match set into the semantic merging process, all valid matches are first classified and grouped according to the clinical semantic anchor category corresponding to the anchor fingerprint. This allows pet identity, medical visit events, symptoms, examinations, diagnoses, medications, and outcomes to enter their respective merging channels. Then, within each merging channel, anchor group merging is performed based on the cross-institutional valid match relationship. Specifically, using any anchor fingerprint pair in a valid match as the initial seed node, anchor fingerprints of the same category that are directly or indirectly connected to that anchor fingerprint pair through the cross-institutional valid match relationship are gradually incorporated into the same anchor group. During the incorporation process, it is checked whether the newly added anchor fingerprints are consistent with the time sequence, relationship structure, and event thread in the existing anchor groups. When a newly added anchor fingerprint merges with an existing anchor group... If most anchor fingerprints in the data satisfy the correspondence of the same pet identity information, the same medical treatment event, or the same context chain, the newly added anchor fingerprint is determined to be merged into the current anchor group; otherwise, a new anchor group is created for it. For anchor fingerprints with similar content but no direct effective matching relationship in different participating institutions, they can be written into the corresponding anchor group as supplementary members if they simultaneously satisfy the same clinical semantic anchor category, the same identity main line, and adjacent event positions, but a low priority mark should be attached. After merging all effective matches, a unified pet medical semantic anchor group across institutions is output. Each anchor group contains at least the anchor group number, clinical semantic anchor category, member anchor fingerprint set, the set of participating institutions to which it belongs, and the time sequence information within the anchor group. In S4.4, after inputting the unified cross-institutional pet medical semantic anchor group into the semantic unit generation process, a cross-institutional master index is first established according to the pet identity mainline. Specifically, the anchor group of pet identity category is used as the root anchor group, and the anchor groups of medical event categories with valid connection relationships with it are sequentially attached to the root anchor group to form a medical event mainline around the same pet identity. Subsequently, within each medical event mainline, anchor group reorganization is performed according to the contextual relationship between examination information, diagnosis information, and medication information. Specifically, the examination category anchor group, diagnosis category anchor group, and medication category anchor group under the same medical event are arranged in the order of examination first, diagnosis in the middle, and medication last, and based on their valid matching set in the cross-institutional system and the current system. The correspondence in the anchor association diagram determines the attribution boundary; for symptom category anchor groups and outcome category anchor groups, they are respectively attached to the front end and back end of the corresponding medical treatment event mainline; after completing the anchor group reorganization, the unit encapsulation process is performed, specifically, the pet identity category anchor group, medical treatment event category anchor group, symptom category anchor group, examination category anchor group, diagnosis category anchor group, medication category anchor group and outcome category anchor group under the same pet identity mainline and belonging to the same medical treatment event mainline are encapsulated into a cross-institutional candidate pet diagnosis and treatment semantic unit, and the candidate unit number, identity mainline number and event mainline number are written for this cross-institutional candidate pet diagnosis and treatment semantic unit; finally, the cross-institutional candidate pet diagnosis and treatment semantic unit is output for subsequent conflict resolution process calls; In S4.5, after inputting cross-institutional candidate pet diagnosis and treatment semantic units into the conflict resolution process, time relevance comparison, information integrity comparison, and context closure comparison are performed on pet identity information, examination information, diagnosis information, medication information, and outcome information where there are content differences between different participating institutions. Specifically, time relevance comparison is obtained by comparing the generation time of the record corresponding to the candidate content with the time position of the corresponding medical event; information integrity comparison is obtained by comparing whether the candidate content simultaneously possesses field content, time position, medical event attribution, participating institution source, and relationship connection results; and context closure comparison is obtained by comparing whether the candidate content can be compared with the same cross-institutional candidate pet diagnosis and treatment semantic units. The preceding and following anchor groups in the unit form a complete diagnostic context chain; after completing the three comparisons, conflict priorities are formed for various types of information with content differences. The conflict priorities can be obtained by multiplying the results of the time relevance comparison, the information integrity comparison, and the context closure comparison by their corresponding weights and summing them, with the sum of each weight set to 1; when a candidate content has the highest conflict priority, it is determined as the content to be retained after conflict resolution, and the remaining content is retained as version reference but not written into the currently effective unit; after all the conflict differences are resolved, the conflict-resolved pet diagnostic semantic unit is output, in which each piece of information in the unit has been uniquely and consistently processed across participating institutions; In S4.6, when inputting the conflict-resolved pet medical semantic unit into the unified registration process, the cross-institutional number writing process is first performed on the pet medical semantic unit. Specifically, a unique cross-institutional number is generated according to the pet's identity main line number, the medical event main line number, and the current registration order, and this cross-institutional number is written into the main identifier of the pet medical semantic unit. Subsequently, the source institution association writing process is performed. Specifically, the participating institution identifiers, source record identifiers, and matching relationship identifiers corresponding to all anchor group members in the pet medical semantic unit are summarized and written into the source institution association area to retain the source trajectory of the pet medical semantic unit. After that, the version identifier writing process is performed. Specifically, the current conflict resolution result is used as the current version content, and the version number is incremented according to the number of changes. When new participating institution data or supplementary data from the same participating institution enters the pet medical semantic unit, a new version is formed by incrementing the version number on the basis of the original version. After completing the cross-institutional number writing, source institution association writing, and version identifier writing, the cross-institutional unified pet medical semantic unit is output and used as the direct input for the subsequent local shared training process. Through the above implementation method, based on anchor fingerprints and local anchor association graphs, cross-institutional candidate matching can be generated first, and then effective matching screening, semantic merging, semantic unit generation, conflict resolution, and unified registration processing can be gradually completed. This enables different participating institutions to form a unified, continuous, and traceable semantic expression of pet diagnosis and treatment around the same pet, the same medical event, and the same medical facts. This reduces semantic deviations caused by time misalignment, category mismatch, relationship inconsistency, and content conflict in cross-institutional data sharing, and provides a semantically consistent and clearly defined input foundation for subsequent local sharing training and sharing platform processing. In practical applications: a pet first completes an examination at participating institution A and forms preliminary diagnostic information, and then completes a follow-up visit at participating institution B and forms updated diagnostic and medication information. The system first generates information based on the data from the two participating institutions. Anchor fingerprints and local anchor association graphs perform identity matching calculations, time sequence calculations, event association calculations, and context consistency calculations to obtain a set of candidate anchor matches. A matching and filtering process then retains valid cross-institutional matches that satisfy the same pet identity, have a reasonable time sequence, and consistent relationships. Subsequently, the corresponding examination category anchor groups, diagnosis category anchor groups, and medication category anchor groups are merged into a unified anchor group and encapsulated into cross-institutional candidate pet diagnosis and treatment semantic units according to the pet identity main line and the treatment event main line. When two participating institutions have different descriptions of the diagnostic content, the superior diagnostic content is retained through time recentity comparison, information completeness comparison, and context closure comparison. Finally, the cross-institutional number, source institution association, and version identifier are written, forming a unified cross-institutional pet diagnosis and treatment semantic unit that can directly enter the subsequent shared training process. S5. Input the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process. Perform model training, privacy perturbation processing and parameter encryption processing on the local basis of the pet diagnosis and treatment semantic units in each participating institution, and output encrypted shared parameters that do not contain the original pet diagnosis and treatment data. In the process of sharing pet medical data, even if a unified pet medical semantic unit has been formed across institutions, directly training based on the original medical records will not only increase the risk of privacy leakage, but also make it difficult to ensure the consistency of semantic structure and parameter form of the content uploaded by different participating institutions. Furthermore, different local models have varying sensitivities to pet identity information, examination information, diagnostic information, medication information, and outcome information when updating parameters. Using the same perturbation method for all parameters can easily lead to insufficient privacy protection or decreased model usability. Therefore, this implementation uses the pet medical semantic unit as the local shared training input. First, training sample assembly, feature vector generation, and local model iterative calculation are completed locally at each participating institution. Then, based on the parameter sensitivity calculation results and parameter category distribution results, sub-parameter perturbation intensity allocation and privacy perturbation processing are performed. Finally, the privacy perturbation parameter set is encrypted and encapsulated, and the institution identifier is isolated and written, thereby outputting encrypted shared parameters that do not contain the original pet medical data content. This implementation process includes the following steps: In S5.1, when inputting the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process, each pet diagnosis and treatment semantic unit is first read in the order of the fields corresponding to pet identity information, medical event information, examination information, diagnosis information, medication information, and outcome information. Training sample assembly is then performed using pet identity information as the primary sample index, medical event information as the sample event boundary, and examination information, diagnosis information, medication information, and outcome information as the sample content area. Specifically, during assembly, information items belonging to the same medical event information under the same pet identity information are grouped into one training sample, and multiple medical events under the same pet identity information are connected in chronological order to form an event sequence sample. A sample number, the identifier of the participating institution, and the training target identifier are written into each training sample or event sequence sample. Subsequently, feature vector generation is performed according to the training task type. When the training task is diagnostic assistance, examination information, symptom information, past outcome information, and current medical event information can be mapped to input features. Diagnostic information is mapped to target features. When the training task is medication recommendation, examination information, diagnostic information, pet identity information, and past medication information can be mapped to input features, and current medication information or outcome information can be mapped to target features. During the mapping process, discrete encoding is performed on the category field in the pet identity information, normalization is performed on the numerical field in the examination information, and term indexing or category indexing is performed on the textual standard fields in the diagnostic and medication information. These are then concatenated according to a unified feature order to generate feature vectors. After the feature vectors are generated, the local model is called to perform iterative calculations. Specifically, the current batch of feature vectors is input into the local model, the error between the prediction result and the target feature is calculated, and then gradient updates or weight updates are performed on the local model parameters based on the error. This process is continued until the preset number of iterations or error convergence conditions are reached. Finally, the updated parameters of the local model corresponding to each participating institution are output. The updated parameters of the local model include at least the parameter number, parameter category, updated value, and corresponding training round. In S5.2, after inputting the local model update parameters corresponding to each participating institution into the privacy protection process, parameter sensitivity calculation and parameter category distribution statistics are first performed on the local model update parameters to obtain the parameter sensitivity calculation results and parameter category distribution results. The parameter sensitivity calculation is based on the responsiveness of a single local model update parameter to changes in pet identity information, examination information, diagnostic information, medication information, and outcome information in the training samples. Specifically, this can be obtained by comparing the magnitude of change of a parameter in adjacent training batches, the dependence of the gradient corresponding to the parameter on different categories of input features, and the influence of the layer containing the parameter on the output results during local training. When a parameter reacts significantly to changes in pet identity information or rare examination information, it is identified as a high-sensitivity parameter; when a parameter mainly reflects common category distribution or low uniqueness features, it is identified as a low-sensitivity parameter. The parameter category distribution results are the results of classifying and statistically analyzing all local model update parameters according to model level, parameter function, and sensitivity level. Specifically, the number of parameters, the proportion of sensitivity levels, and the distribution of update magnitude can be statistically analyzed for input layer parameters, intermediate layer parameters, and output layer parameters. After obtaining the parameter sensitivity calculation results and parameter category distribution results, sub-parameter perturbation intensity allocation is performed on the local model update parameters. Specifically, the local model update parameters are first divided into high-perturbation parameter groups, medium-perturbation parameter groups, and low-perturbation parameter groups. Then, different perturbation intensities are set for the three groups of parameters. The perturbation intensity of the high-perturbation parameter group is greater than that of the medium-perturbation parameter group, and the perturbation intensity of the medium-perturbation parameter group is greater than that of the low-perturbation parameter group. The specific values ​​can be determined according to the privacy budget upper limit and model accuracy lower limit set by the participating institutions. Subsequently, privacy perturbation processing is performed on different parameter groups. Specifically, random perturbation amounts are added to the corresponding parameter update values ​​according to their respective perturbation intensities to form perturbed parameter update results. After completing the perturbation processing of all parameters, a privacy perturbation parameter set is output. The privacy perturbation parameter set includes at least the parameter number, parameter category, perturbed parameter value, and perturbation intensity identifier. In S5.3, after the privacy perturbation parameter set is input into the parameter encryption process, the privacy perturbation parameter set is first subjected to parameter encryption and encapsulation processing. Specifically, the parameter values, parameter categories, training rounds, and perturbation strength identifiers in the privacy perturbation parameter set are organized into parameter messages according to the parameter number order. Then, the local encryption module is called to perform encryption operations on the parameter messages to generate encrypted parameter blocks whose original parameter values ​​cannot be directly read. After the parameter encryption and encapsulation is completed, the institution identifier isolation and writing processing is performed. Specifically, the real name of the participating institution is not directly written into the encrypted parameter block. Instead, an independent institution identifier code is assigned to the current participating institution, and this institution identifier code is written into the institution identifier area of ​​the encrypted parameter block. At the same time, the institution identifier code is compared with the real name of the participating institution. The mapping relationship between participating institutions is stored only in the local security table and is not output with the encrypted parameter block. If the same participating institution generates multiple sets of privacy perturbation parameters in different training rounds, round and batch markers can be added after the institution identifier area so that the sharing platform can collect and verify the order of parameters later. The final output is an encrypted shared parameter that does not contain the original pet diagnosis and treatment data. This encrypted shared parameter only contains the parameter data after privacy perturbation and encryption, the isolated institution identifier code, and the necessary round and batch markers. It does not contain the original pet diagnosis and treatment data, the original feature vector content, or the original pet diagnosis and treatment semantic unit content, thus meeting the upload requirements of the subsequent sharing platform. Through the above implementation method, the unified pet diagnosis and treatment semantic units across institutions can be first converted into training samples and feature vectors that can participate in local shared training. Then, local model iterative calculations, parameter perturbation intensity allocation, privacy perturbation processing, parameter encryption encapsulation, and institution identifier isolation writing are completed locally at each participating institution. This forms encrypted shared parameters that can be aggregated by the sharing platform without outputting the original pet diagnosis and treatment data. This ensures the ability of local shared training to utilize pet diagnosis and treatment patterns, reduces the risk of leakage of pet identity information, examination information, diagnostic information, and medication information during the sharing process, and improves the consistency of uploaded parameters in terms of format and processing flow among different participating institutions. In practical applications: a participating institution uses examination information, diagnostic information, and medication information from the pet diagnosis and treatment semantic units to train a diagnosis and treatment aid. The model first assembles examination, diagnosis, and outcome information corresponding to multiple medical events under the same pet identity information into training samples, and generates input feature vectors and target features. Then, it iterates through multiple rounds to obtain local model update parameters. Subsequently, parameter sensitivity calculation and parameter category distribution statistics are performed on the local model update parameters. Parameters involving rare examination items and pet identity-related features are classified into high-perturbation parameter group, and parameters reflecting general diagnostic patterns are classified into low-perturbation parameter group. Random perturbations of different intensities are added to each group to form a privacy perturbation parameter set. Finally, the privacy perturbation parameter set is encrypted and encapsulated, and isolated institution identification codes, training round markers, and batch markers are written into it to generate encrypted shared parameters that do not contain the original pet medical data content. These parameters are then uploaded to the sharing platform for subsequent aggregation processing. S6. Upload the encrypted shared parameters output by each participating institution to the sharing platform. The sharing platform performs aggregation calculation, access control verification, security audit recording and sharing result distribution processing on the encrypted shared parameters, and outputs shared results for remote diagnosis, disease early warning, medical resource allocation or veterinary model training. In the process of sharing pet medical data, although each participating institution has completed local shared training, privacy perturbation processing, and parameter encryption encapsulation, and outputs encrypted shared parameters that do not contain the original pet medical data, it is still difficult to form a stable and callable shared result if the sharing platform cannot further uniformly receive, verify, collect, aggregate, and distribute encrypted shared parameters from different participating institutions, different training rounds, and different parameter versions. Especially when multiple participating institutions upload in parallel, different calling entities access concurrently, and different application scenarios have different requirements for the form of shared results, the lack of a coherent mechanism for parameter reception registration, parameter integrity verification, access control verification, security audit recording, and result distribution processing can easily lead to problems such as parameter misuse, aggregation deviation, unauthorized access, and inconsistent result delivery. Therefore, this implementation method establishes a unified platform processing flow for encrypted shared parameters, first forming a set of parameters to be aggregated, and then gradually completing parameter weighted aggregation calculation, access control verification, security audit recording, and result distribution processing to output shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training. This implementation process includes the following steps: In S6.1, after the encrypted shared parameters output by each participating institution are uploaded to the sharing platform, the sharing platform first performs parameter reception registration according to the participating institution identifier, parameter source batch, parameter type marker, and parameter version marker. Specifically, for each encrypted shared parameter, it reads the attached participating institution identifier, parameter source batch, parameter type marker, parameter version marker, training round marker, and parameter message length information, and assigns a platform reception number and reception timestamp to the encrypted shared parameter. Subsequently, it performs parameter integrity verification, specifically: it checks whether the message length of the encrypted shared parameter is complete based on the parameter message length information; it checks whether the participating institution identifier, parameter source batch, parameter type marker, and parameter version marker are missing based on the preset verification fields; and it checks whether the parameter message header, parameter body, and end marker meet the requirements based on the encrypted encapsulation format. According to the platform's receiving format requirements, if any verification item fails to meet the requirements, the corresponding encrypted shared parameter is written to the exception queue and marked as pending retransmission. After passing the parameter integrity verification, parameter grouping and aggregation processing is performed. Specifically, encrypted shared parameters from the same training batch are grouped at the first level according to the parameter source batch, at the second level according to the parameter type label, and at the third level according to the parameter version label. Within each group, duplicate registrations are made according to the participating institution identifier to avoid parameters from the same batch uploaded by the same participating institution from entering multiple aggregations. After completing the above processing, the set of parameters to be aggregated is output. The set of parameters to be aggregated includes at least the group number, the set of participating institutions within the group, the parameter type label, the parameter version label, and the parameter quantity statistics, for subsequent use in the platform's aggregation processing flow. In S6.2, after the set of parameters to be aggregated is input into the platform's aggregation processing flow, a weighted aggregation calculation is first performed based on the effectiveness of the participating institutions' contributions, the consistency of parameter update directions, and the historical aggregation stability. The effectiveness of the participating institutions' contributions is determined by comparing whether the parameters uploaded by each participating institution arrive on time, pass parameter integrity checks, meet the minimum training round requirement, and have continuously participated in uploading within the most recent preset period. When a participating institution meets the above conditions in the current batch, its contribution effectiveness is determined to be valid; otherwise, it is determined to be less effective or invalid. The consistency of parameter update directions is determined by comparing whether the update directions of the parameters uploaded by each participating institution under the same parameter type are consistent in the current batch. Specifically, the update values ​​of parameters at the same position can be compared. The symbol direction is selected separately, and the direction consistency ratio is calculated. When the direction consistency ratio is higher than the consistency threshold, the parameter type is determined to have high parameter update direction consistency in the current batch. Historical aggregation stability is obtained by reading the change range of aggregation results of the same parameter type in the previous few batches in the shared platform. When the change range of aggregation results in adjacent historical batches is consistently within a stable range, the current parameter type is determined to have high historical aggregation stability. After obtaining the above three results, an aggregation weight is assigned to the parameter currently uploaded by each participating institution. The aggregation weight can be obtained by multiplying the participating institution's contribution effectiveness score, parameter update direction consistency score, and historical aggregation stability score by their respective weights and then summing them. The sum of each weight is set to 1, and then the parameters are treated according to the aggregation weight. The parameters in the aggregation parameter set undergo weighted aggregation calculation, outputting intermediate parameters of the aggregated shared result. Simultaneously, the aggregation request is processed for access identity verification, access permission matching, and call scope verification. Specifically: first, access identity verification is performed based on the access credentials submitted by the requester to confirm whether the requester is a registered calling entity; then, access permission matching is performed based on the requester's entity type, business role, and preset permission table to confirm whether the requester has permission to call the current shared result type; finally, call scope verification is performed based on the application purpose and calling object scope declared by the requester to confirm whether the requester's calling scope falls within its authorized remote diagnosis, disease early warning, medical resource allocation, or veterinary model. Within the scope of training, if any verification fails, the current result output is stopped and an access denial record is written; if all verifications pass, a security audit record is executed on the parameter reception time, aggregation process, access request content, and processing result. Specifically, the parameter reception time, group number, set of participating organization identifiers, aggregation weight allocation result, access subject identifier, access time, call type, processing conclusion, and output result number are written to the audit log; after completing the parameter weighted aggregation calculation, access control verification, and security audit record, a shared result data packet is output, which includes at least the result number, result type identifier, intermediate parameters or derived results of the aggregated shared result, call permission label, and audit association number. In S6.3, after the shared result data package is input into the result distribution process, the result encapsulation is performed according to the corresponding call type: remote diagnosis, epidemic early warning, medical resource allocation, or veterinary model training. Specifically: when the call type is remote diagnosis, the shared result data package is encapsulated into a diagnostic auxiliary result package and the result fields applicable to the diagnostic call are written; when the call type is epidemic early warning, the shared result data package is encapsulated into an early warning result package and the time, risk, and region fields applicable to the early warning call are written; when the call type is medical resource allocation, the shared result data package is encapsulated into a resource scheduling result package and the institution load and resource requirement fields applicable to the resource allocation call are written; when the call type is veterinary model training, the shared result data package is encapsulated into a model update result package and the model version and training round fields applicable to the model training call are written; After the results are encapsulated, target organization matching is performed. Specifically, based on the result type identifier, access permission tag, and the identifier of the calling entity's organization in the shared result data packet, a set of target organizations that meet the receiving conditions is selected from the platform's registered organization table, and organizations that are not within the current calling scope or do not have receiving permissions are removed. Subsequently, distribution channel writing is performed. Specifically, based on the target organization's access method, terminal type, and business system interface, the encapsulated shared result data packet is written to the corresponding distribution channel. The distribution channel can be one of an interface call channel, a message push channel, or a batch download channel. After writing is completed, the result distribution time, target organization identifier, distribution channel type, and result receipt status are recorded, and shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training are output for the corresponding target organizations to directly call or receive. Through the above implementation method, the encrypted shared parameters uploaded by each participating institution can first complete parameter reception registration, parameter integrity verification, and parameter grouping and aggregation on the sharing platform. Then, under unified rules, parameter weighted aggregation calculation, access control verification, and security audit recording are completed. Finally, according to different call types, result encapsulation, target institution matching, and distribution channel writing are completed. Thus, without exposing the original pet medical data content, a shared result that can be uniformly managed, accessed according to permissions, audited and traced, and applied to different business scenarios is formed. This also reduces the application risks caused by parameter mixing, unauthorized access, or result distribution mismatch during parallel sharing by multiple participating institutions. In practical applications: a sharing platform simultaneously receives data from three participating institutions in the same training batch. The platform first registers, verifies, and groups the transmitted encrypted shared parameters based on the participating organization's identifier, parameter source batch, parameter type, and parameter version, forming a set of parameters to be aggregated. Then, the platform assigns aggregation weights based on the effectiveness of the contributions of the three participating organizations, the consistency of parameter update directions, and historical aggregation stability, completing the weighted aggregation calculation. It also performs access identity verification, access permission matching, and call scope verification on a remote diagnostic call request, while simultaneously writing corresponding security audit records. Finally, the platform encapsulates the aggregated shared result data packet into a diagnostic auxiliary result packet, matches it with target organizations that have receiving permissions, and distributes the shared results through the interface call channel for the target organizations to directly use in their remote diagnostic business.

[0011] Refer to the instruction manual appendix Figure 2 Furthermore, the present invention also includes a pet medical data sharing system for performing the above method, comprising a data standardization module, a local semantic parsing module, an anchor coding association module, a cross-institutional semantic alignment module, a local shared training module, and a shared platform processing module, which respectively execute S1-S6.

[0012] Working Principle: This solution can be understood as a processing chain from messy data to securely shareable results. First, the system integrates the raw pet medical data stored locally by participating institutions. Based on a unified data dictionary, it performs field extraction, field name mapping, data type conversion, missing item marking, time format standardization, and appointment time sorting, thus organizing the originally inconsistent and scattered data into a standardized medical dataset. On this basis, it extracts pet identity information, appointment event information, symptom information, examination information, diagnosis information, medication information, and outcome information from the standardized medical dataset and organizes this information into a set of clinical semantic anchors. Subsequently, the system performs local encoding, relation binding, and irreversible transformation on the clinical semantic anchors. The process involves generating comparable anchor fingerprints that do not directly expose the original content, and further constructing a local anchor association graph to preserve contextual relationships within the same visit and between previous and subsequent visits. Then, the anchor fingerprints and local anchor association graphs from different participating institutions are semantically aligned across institutions to form a unified cross-institutional pet medical semantic unit. Finally, each participating institution completes model training, privacy perturbation processing, and parameter encryption locally based on this pet medical semantic unit, and then uploads the encrypted shared parameters to the sharing platform. The sharing platform performs aggregation calculations, access control verification, security audit recording, and result distribution, thereby achieving secure, efficient, and traceable cross-institutional data sharing and result application without directly sharing the original pet medical data. For example, a pet first has its initial visit to a veterinary clinic, where its chief complaint, examination results, and preliminary diagnosis are recorded. It then goes to another veterinary hospital for a follow-up visit, where new diagnoses and medication records are added. In practice, this solution first reads the treatment data stored by both institutions, organizing the visit time, examination time, diagnosis, and medication information into a unified format. It then extracts the pet's corresponding symptom information, examination information, diagnosis information, and medication information to form clinical semantic anchors representing the treatment facts. Next, the system converts these clinical semantic anchors into anchor fingerprints and, combined with the local anchor association graph, determines which information belongs to the same visit and which information belongs to subsequent follow-up visits. The process involves two institutions' data entering a cross-institutional semantic alignment process. The system then identifies that the data actually corresponds to the same pet and the same continuous treatment chain, merging the initial and follow-up consultation information into a unified pet treatment semantic unit. Each institution then uses this unified semantic unit locally for model training, uploading only parameters that have undergone privacy perturbation and encryption. The sharing platform then aggregates these parameters and provides the aggregated results for applications such as remote diagnosis, disease early warning, or veterinary model training. This allows participating institutions to utilize cross-institutional data to form a more complete understanding of treatment without directly exposing their own stored original medical records.

[0013] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for sharing pet medical treatment data, characterized in that, Includes the following steps: S1. Obtain the original pet medical data stored locally by each participating institution. Based on the unified data dictionary, perform field extraction, field name matching, data type conversion, missing item marking, time format unification, and treatment time sorting on the original pet medical data, and output the standardized medical dataset corresponding to each participating institution. S2. Input the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process, extract pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information from the standardized diagnosis and treatment datasets, and generate a set of clinical semantic anchors according to the information categories; S3. Perform local encoding, anchor relationship binding and irreversible transformation processing on each clinical semantic anchor in the clinical semantic anchor set of each participating institution to generate anchor fingerprints that correspond one-to-one with each clinical semantic anchor, and construct a local anchor association graph based on the order of visits, examination and diagnosis and diagnosis and medication relationship between each clinical semantic anchor. S4. Input the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. Perform cross-institutional matching calculations according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical events, and the contextual correspondence of examination information and medication information. Output a unified cross-institutional pet medical semantic unit. S5. Input the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process. Perform model training, privacy perturbation processing and parameter encryption processing on the local basis of the pet diagnosis and treatment semantic units in each participating institution, and output encrypted shared parameters that do not contain the original pet diagnosis and treatment data. S6. Upload the encrypted shared parameters output by each participating institution to the sharing platform. The sharing platform performs aggregation calculation, access control verification, security audit recording, and sharing result distribution processing on the encrypted shared parameters, and outputs shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training.

2. The method according to claim 1, characterized in that, S1 includes the following steps: S1.1 Read the original pet diagnosis and treatment data stored locally by each participating institution and read the standard field names, field data types, field value ranges and field time semantics in the unified data dictionary. Perform field extraction and field attribution calculation on the original pet diagnosis and treatment data according to field name similarity, field data type consistency, field value range matching degree and field co-occurrence relationship, and output candidate standardized data fragments with field source identifiers and field attribution results. S1.2 Input the candidate standardized data fragments into the standardization correction process, perform data type conversion, missing item marking, time format unification and consultation time sorting according to the results corresponding to the field names, and merge data fragments with multiple candidate fields according to the stability of field source, time association integrity and cross-record consistency, and output the standardized diagnosis and treatment datasets corresponding to each participating institution.

3. The method according to claim 2, characterized in that, S2 includes the following steps: S2.1 Input the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process. Perform information extraction and category classification calculation according to the field position relationship, temporal adjacency relationship and record co-occurrence relationship between the pet basic information field, medical record field, symptom description field, examination result field, diagnosis record field, medication record field and outcome record field. Output a set of candidate semantic fragments corresponding to pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information respectively. S2.2 Input the candidate semantic fragment set into the semantic anchor generation process. According to the information association relationship within the same medical event and the time sequence relationship between different medical events, perform fragment merging, conflict resolution and category binding processing on the candidate semantic fragment set to generate information categorized clinical semantic anchor sets corresponding to pet identity, medical event, symptoms, examination, diagnosis, medication and outcome.

4. The method according to claim 3, characterized in that, S3 includes the following steps: S3.1 Input each clinical semantic anchor in the clinical semantic anchor set of each participating institution into the local encoding process, read the information category identifier, field content, time position and medical event affiliation corresponding to each clinical semantic anchor, and perform fixed-length encoding and category bit concatenation processing according to information category, field order, time sequence and medical event affiliation, and output the initial encoding fragment that corresponds one-to-one with each clinical semantic anchor. S3.2 Input the initial encoded fragment into the anchor relationship binding process, and perform relationship number writing and association bit binding processing according to the co-occurrence relationship of the same visit, the succession relationship of previous and subsequent visits, and the contextual pointing relationship between examination information, diagnosis information and medication information among each clinical semantic anchor, and output the associated encoded fragment with relationship binding mark; S3.3 Input the associated coding fragment into the irreversible conversion process, perform irreversible mapping and fixed-length truncation on the associated coding fragment according to the preset one-way conversion rules, and generate anchor fingerprints that correspond one-to-one with each clinical semantic anchor.

5. The method according to claim 4, characterized in that, S3 further includes the following steps: S3.4 Register the anchor fingerprints generated by each participating institution according to the order of the medical treatment events and time sequence to which the corresponding clinical semantic anchors belong. Perform the succession judgment on consecutive anchor fingerprints in adjacent medical treatment events. Perform the examination and diagnosis correspondence judgment on the anchor fingerprints corresponding to examination information and the anchor fingerprints corresponding to diagnosis information in the same medical treatment event. Perform the diagnosis and medication correspondence judgment on the anchor fingerprints corresponding to diagnosis information and the anchor fingerprints corresponding to medication information. Output the set of candidate relationships between anchors. S3.5 Input the set of candidate anchor relationships into the relationship filtering process, and perform relationship retention and conflict resolution processing according to whether the time difference is within the same medical visit range, whether the information category combination meets the requirements of examination diagnosis or diagnosis and medication, and whether the same anchor fingerprint forms a conflict in multiple candidate relationships, and output the set of valid anchor relationships. S3.

6. Connect the effective set of anchor relationships with each anchor fingerprint using node edges to generate a local anchor association graph that includes the order of visits, examination and diagnosis, and diagnosis and medication relationships.

6. The method according to claim 5, characterized in that, S4 includes the following steps S4.1 Input the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. Perform identity matching calculation, time continuity calculation, event association calculation and context consistency calculation according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical treatment events and the contextual correspondence of examination information and medication information, and output the candidate matching set between anchors among the participating institutions. S4.2 Input the candidate matching set between anchors into the matching and filtering process. Perform candidate matching retention and conflict matching removal according to the constraints of the order of medical treatment time under the same pet identity, the matching constraints of information category under the same medical treatment event, and the consistency constraints of the relationship between cross-organization anchors and the local anchor association graph. Output the effective matching set across organizations. S4.3 Input the cross-institutional effective matching set into the semantic merging process, perform anchor group merging processing according to the clinical semantic anchor category corresponding to the anchor fingerprint and the cross-institutional effective matching relationship, and output a unified cross-institutional pet diagnosis and treatment semantic anchor group.

7. The method according to claim 6, characterized in that, S4 further includes the following steps: S4.4 Input the cross-institutional unified pet diagnosis and treatment semantic anchor group into the semantic unit generation process, and perform anchor group reorganization and unit encapsulation processing according to the pet identity main line, the medical treatment event main line, and the contextual attribution relationship between examination information, diagnosis information and medication information, and output cross-institutional candidate pet diagnosis and treatment semantic units. S4.5 Input the cross-institutional candidate pet diagnosis and treatment semantic units into the conflict resolution process, and perform time-recentness comparison, information integrity comparison, and context closure comparison on the pet identity information, examination information, diagnosis information, medication information, and outcome information that have content differences between different participating institutions, and output the pet diagnosis and treatment semantic units after conflict resolution. S4.6 Input the conflict-resolved pet diagnosis and treatment semantic units into the unified registration process, perform cross-institutional number writing, source institution association writing, and version identifier writing processing on the pet diagnosis and treatment semantic units, and output cross-institutional unified pet diagnosis and treatment semantic units.

8. The method according to claim 7, characterized in that, S5 includes the following steps: S5.1 Input the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process, and perform training sample assembly, feature vector generation and local model iterative calculation according to the pet identity information, medical event information, examination information, diagnosis information, medication information and outcome information in the pet diagnosis and treatment semantic units, and output the local model update parameters corresponding to each participating institution. S5.2 Input the local model update parameters corresponding to each participating institution into the privacy protection process, perform sub-parameter perturbation intensity allocation and privacy perturbation processing on the local model update parameters according to the parameter sensitivity calculation results and parameter category distribution results, and output the privacy perturbation parameter set; S5.

3. Input the privacy perturbation parameter set into the parameter encryption process, perform parameter encryption encapsulation and institution identifier isolation writing processing on the privacy perturbation parameter set, and output encrypted shared parameters that do not contain the original pet diagnosis and treatment data.

9. The method according to claim 8, characterized in that, S6 includes the following steps: S6.1 Upload the encrypted shared parameters output by each participating institution to the sharing platform. The sharing platform will perform parameter reception registration, parameter integrity verification and parameter grouping and aggregation processing according to the participating institution identifier, parameter source batch, parameter type tag and parameter version tag, and output the set of parameters to be aggregated. S6.2 Input the set of parameters to be aggregated into the platform's aggregation processing flow, perform parameter weighted aggregation calculation according to the validity of the contributions of participating institutions, the consistency of parameter update direction, and the stability of historical aggregation, and perform access identity verification, access permission matching, and call scope verification on the aggregation request. At the same time, perform security audit recording on parameter reception time, aggregation processing process, access request content, and processing results, and output a shared result data packet. S6.3 Input the shared result data packet into the result distribution process, perform result encapsulation, target institution matching and distribution channel writing processing according to the call type corresponding to remote diagnosis, disease early warning, medical resource allocation or veterinary model training, and output the shared results for remote diagnosis, disease early warning, medical resource allocation or veterinary model training.

10. A pet medical data sharing system, characterized in that, For performing the method according to any one of claims 1-9, comprising: The data standardization module acquires the original pet medical data stored locally by each participating institution, and performs field extraction, field name matching, data type conversion, missing item marking, time format standardization, and treatment time sorting on the original pet medical data according to the unified data dictionary, and outputs the standardized medical dataset corresponding to each participating institution. The local semantic parsing module inputs the standardized diagnosis and treatment datasets corresponding to each participating institution into the local semantic parsing process, extracts pet identity information, medical event information, symptom information, examination information, diagnosis information, medication information and outcome information from the standardized diagnosis and treatment datasets, and generates a set of clinical semantic anchors according to the information categories; The anchor coding and association module performs local encoding, anchor relationship binding and irreversible transformation processing on each clinical semantic anchor in the clinical semantic anchor set of each participating institution, generates anchor fingerprints that correspond one-to-one with each clinical semantic anchor, and constructs a local anchor association graph based on the order of visits, examination and diagnosis and diagnosis and medication relationship between each clinical semantic anchor. The cross-institutional semantic alignment module inputs the anchor fingerprints and local anchor association graphs generated by each participating institution into the cross-institutional semantic alignment process. It performs cross-institutional matching calculations according to the correspondence of pet identity information, the continuity of medical treatment time, the association of medical events, and the contextual correspondence between examination information and medication information, and outputs a unified cross-institutional pet medical semantic unit. The local shared training module inputs the pet diagnosis and treatment semantic units output by each participating institution into the local shared training process. Each participating institution performs model training, privacy perturbation processing, and parameter encryption processing based on the pet diagnosis and treatment semantic units, and outputs encrypted shared parameters that do not contain the original pet diagnosis and treatment data. The shared platform processing module uploads the encrypted shared parameters output by each participating institution to the shared platform. The shared platform then performs aggregation calculations, access control verification, security audit logging, and shared result distribution on the encrypted shared parameters, outputting shared results for remote diagnosis, disease early warning, medical resource allocation, or veterinary model training.

Images