Method and apparatus for processing cross-institutional health record data based on multi-dimensional features

Abstract

The present specification provides a method and device for processing cross-institutional health record data based on multi-dimensional features. Based on a plurality of health record texts, a first type of field of the health record texts can be extracted; and according to the first type of field, identification blocking keys of the health record texts are obtained and utilized by using identification indication information of the health record texts to construct a first node graph; and according to a preset processing rule, the identification indication information of the health record texts is used to determine and utilize the association evidence data between two connected nodes in the first node graph to obtain a second node graph; according to the second node graph, a health record text group of a plurality of users is determined through clustering processing; and a health status of a concerned object is accurately analyzed and determined based on the health record text group of the user by using a preset health status prediction model, while determining the health record text group of the user of the concerned object and the associated object.

Description

Technical Field

[0001] This manual belongs to the field of health big data governance technology, and in particular relates to methods and devices for processing cross-institutional health record data based on multi-dimensional features. Background Technology

[0002] As living standards improve, people are paying more and more attention to their health.

[0003] Currently, many users, for various reasons, visit different institutions (e.g., different hospitals, different medical examination centers) at different times to participate in health services such as physical examinations, resulting in a large number of scattered health data records. Due to significant differences in the standardization of data management, the rules for generating file numbers, and user information entry habits among different institutions, the information recorded in these health data records, especially file identifiers and user identifiers, is often chaotic. Problems such as multiple files for one person, multiple people for one file, and conflicting file numbers across institutions frequently occur. This makes it impossible to effectively unify and integrate the health data records generated by users at different times through different institutions, hindering a comprehensive and accurate analysis of users' health status.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This specification provides a method and apparatus for processing cross-institutional health record data based on multi-dimensional features. It can efficiently and accurately find the health record data corresponding to the same user from massive amounts of chaotic health record data, so as to achieve unified governance and effective integration of multi-source health record data across institutions.

[0006] This specification provides a method for processing cross-institutional health record data based on multidimensional features, including: Obtain multiple health record texts; and extract corresponding first type fields from the multiple health record texts; wherein, the first type fields are fields that contain at least identification indication information; Based on the first type of field, obtain the identification indication information of the health record text; and based on the identification indication information, construct the identification blocking key of the health record text; A first node graph is constructed based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text. Based on the preset rules for processing related evidence, the identification information of the health record text is used to determine the related evidence data between two connected nodes in the first node diagram; Based on the associated evidence data, the weight parameters of the edges between two connected nodes are calculated; and the corresponding weight parameters are marked on the edges to obtain the second node graph. Based on the second node graph, multiple connected clusters are identified through clustering; each connected cluster corresponds to one user. Based on the multiple connected clusters, the health record text groups of multiple users are determined by merging the corresponding health record texts. Receive and respond to health information query requests for targets of interest, identify targets of interest and associated objects associated with targets of interest; and query the health record text groups of the corresponding users based on the consensus record identifiers corresponding to targets of interest and associated objects to obtain multiple health record texts; Based on the multiple health record texts, second-type fields related to the objects of interest and associated objects are extracted respectively, serving as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and based on the target fields and auxiliary fields, corresponding target health feature sequences and auxiliary health feature sequences are constructed. A pre-defined health status prediction model is used to process the target health feature sequence and the auxiliary health feature sequence to obtain the health status prediction result of the object of interest. The pre-defined health status prediction model includes at least a first health prediction branch and a second health prediction branch connected in parallel, as well as a joint regression analysis model. The first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0007] In one embodiment, the identification information includes: attribute information of the health record text and attribute information of the user; The attribute information of the health record text includes at least one of the following: the health record text's file number, the physical examination number associated with the health record text, the physical examination date associated with the health record text, the name of the institution that generated the health record text, and the text generation date of the health record text. The user's attribute information includes at least one of the following: user ID card number, user name, user date of birth, user mobile phone number, user age, name of the organization to which the user belongs, and user gender.

[0008] In one embodiment, constructing the first node graph based on the identifier indication information and identifier blocking key of the health record text includes: Based on the identification information of the health record text, set up a node corresponding to the health record text in the initial node diagram; and mark the identification information of the health record text on the node. Based on the identifier blocking key of the health record text, two potentially related health record texts are identified; wherein, at least one identifier blocking key is shared between the two potentially related health record texts. By connecting the two nodes corresponding to two potentially related health record texts with edges, the corresponding first node graph can be obtained.

[0009] In one embodiment, determining the association evidence data between two connected nodes in the first node graph using the identifier information of the health record text according to preset processing rules includes: The evidence data linking the first and second connected nodes is determined in the following manner: The identification information of the first node and the identification information of the second node are processed using a preset feature matching model, and the information matching degree of the corresponding identification information of the first node and the second node is calculated. Based on the preset rules for processing related evidence, and according to the information matching degree of the corresponding identifier indication information of the first node and the second node, the identifier indication information of the first node and the identifier indication information of the second node that meet the requirements are determined and configured to construct one or more pieces of related evidence. By combining the aforementioned related evidence, we obtain the related evidence data between the first node and the second node.

[0010] In one embodiment, calculating the weight parameter of the edge between two connected nodes based on the associated evidence data includes: The weight parameters of the edge connecting the first and second nodes are calculated as follows: Based on the pre-defined rules for handling related evidence, determine the type of evidence for related evidence; Based on multiple pieces of related evidence and the types of evidence associated with them, a similarity score is determined between the first node and the second node. Based on the preset rules for processing related evidence, the weight parameters of the connection between the first and second connected nodes are calculated using the similarity score.

[0011] In one embodiment, the method further includes: The system checks whether the key indication information in the identification indication information of the first node and the identification indication information of the second node is empty; wherein, the key indication information includes: user ID number, user gender, and user date of birth; When the key indication information in the first node's identification indication information and the second node's identification indication information is not empty, check whether the key identification indication information of the first node and the key identification indication information of the second node are the same; When the key identifier information of the first node is different from that of the second node, the weight parameter of the edge connecting the first node and the second node is set to 0.

[0012] In one embodiment, determining multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters includes: Based on the multiple connected clusters, multiple health profile texts of multiple users are determined; Based on the preset file merging rules, a consensus file identifier is determined for multiple health record documents for each user; Using consensus archive identifiers, multiple health record texts from various users are merged and combined to obtain a health record text group for multiple users.

[0013] In one embodiment, determining the consensus file identifier for multiple health record files of each user according to preset file merging rules includes: Check if the file number identifier field of multiple health record texts of a user is empty; When at least one of the health record texts has a non-empty file identifier in its file number identifier field, the existing file identifier is extracted based on the file number identifier field of the user's health record text. Check if the number of existing file identification types is greater than 1; When the number of existing archival identifiers is greater than 1, determine the number of existing archival identifiers and the generating institution for each type; Based on the quantity and generating organization of existing archival identifiers of each type, existing archival identifiers that meet the requirements are identified as candidate archival identifiers. Detect whether the candidate file identifier overlaps with the consensus file identifier of other users; When a candidate profile identifier overlaps with the consensus profile identifier of other users, a consensus profile identifier for multiple health profiles of that user is generated using the candidate profile identifier according to the preset identifier generation rules.

[0014] In one embodiment, after determining multiple user health record text groups by merging the corresponding health record texts according to the multiple connected clusters, the method further includes: Based on the user's health record text group, the second type of fields of the health record text are extracted; wherein, the second type of fields are fields that contain at least the user's health information; Determine the reliability weight coefficient of the second type of field based on the institution level and record type of the institution that generated the health record text; By performing semantic recognition on the second type of field, the target object indicated by the second type of field is determined; Detect whether there is overlap between the target objects indicated by the second type of field; When it is determined that the target objects indicated by the second type of field overlap, the second type of field in which the indicated target objects overlap is selected as the first candidate field; Based on the text generation date of the health record text to which the first candidate field belongs, filter out the second candidate fields whose text generation dates meet the requirements; Check if the number of second candidate fields is greater than 1; When the number of second candidate fields is greater than 1, multiple second candidate fields are merged according to the preset archive data fusion rules and the reliability weight coefficient to generate a health information consensus field about the health information of the user's target object; wherein, the health information consensus field is used to generate a health data fusion summary report for the user.

[0015] This specification also provides a processing apparatus for cross-institutional health record data based on multidimensional features, including: The first acquisition module is used to acquire multiple health record texts and extract corresponding first type fields from the multiple health record texts; wherein, the first type fields are fields that contain at least identification indication information; The first construction module is used to obtain the identification indication information of the health record text based on the first type of field; and to construct the identification blocking key of the health record text based on the identification indication information. The second construction module is used to construct a first node graph based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text. The first determination module is used to determine the associated evidence data between two connected nodes in the first node diagram based on the preset associated evidence processing rules and the identification information of the health record text. The first processing module is used to calculate the weight parameters of the edge between two connected nodes based on the associated evidence data; and to mark the corresponding weight parameters on the edge to obtain the second node graph. The second processing module is used to determine multiple connected clusters based on the second node graph through clustering processing; wherein, one connected cluster corresponds to one user; The second determining module is used to determine multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters. The second acquisition module is used to receive and respond to health information query requests for objects of interest, identify objects of interest and associated objects associated with objects of interest, and query the health record text groups of the corresponding users according to the consensus record identifiers corresponding to the objects of interest and associated objects to obtain multiple health record texts; The extraction module is used to extract second-type fields about the object of interest and the associated object from the multiple health record texts, respectively, as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and construct corresponding target health feature sequences and auxiliary health feature sequences based on the target fields and auxiliary fields; The third processing module is used to process the target health feature sequence and the auxiliary health feature sequence using a preset health status prediction model to obtain the health status prediction result of the object of interest; wherein, the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0016] Based on the multi-dimensional feature-based cross-institutional health record data processing method and apparatus provided in this specification, after acquiring multiple health record texts of different text types generated by different institutions, the method first extracts a first type of field from the multiple health record texts; then, based on the first type of field, obtains the identification indication information of the health record texts; next, based on the identification indication information, constructs the identification blocking key of the health record texts; and based on the identification indication information and the identification blocking key of the health record texts, constructs a first node graph; wherein the first node graph contains multiple nodes, each node corresponding to a health record text, and the nodes are also marked with the corresponding identification indication information of the health record texts; according to preset processing rules, using the identification indication information of the health record texts, the association evidence data between two connected nodes in the first node graph is determined; based on the association evidence data, the weight parameters of the edges between the two connected nodes are calculated; and the corresponding weight parameters are marked on the edges to obtain a second node graph; then, based on the second node graph, multiple connected clusters are determined through graph structure-based clustering processing; wherein each connected cluster corresponds to a user; and based on... The multiple connected clusters, by merging the corresponding health record texts, determine multiple user health record text groups; receive and respond to health information query requests for objects of interest, determine the objects of interest, and associated objects associated with the objects of interest; and query the corresponding user health record text groups according to the consensus record identifiers corresponding to the objects of interest and associated objects to obtain multiple health record texts; extract second-type fields about the objects of interest and associated objects from the multiple health record texts, respectively, as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and construct corresponding target health feature sequences and auxiliary health feature sequences according to the target fields and auxiliary fields; process the target health feature sequences and auxiliary health feature sequences using a preset health status prediction model to obtain the health status prediction result of the objects of interest; wherein, the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure. This allows it to be well-suited for health big data governance scenarios, efficiently and accurately identifying health record data corresponding to the same user from massive amounts of chaotic health record data. This enables unified governance and effective integration of multi-source health record data from different users across institutions and time periods. Furthermore, through a pre-set health status prediction model, it can effectively utilize the unified and integrated user health record text groups to comprehensively and accurately analyze and determine the health status of the target audience, thereby providing users with health consultation services that have high reference value. Attached Figure Description

[0017] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for processing cross-institutional health record data based on multidimensional features, provided in one embodiment of this specification. Figure 2 This is a schematic diagram of an embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Figure 3 This is a schematic diagram of an embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Figure 4 This is a schematic diagram of an embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Figure 5 This is a schematic diagram of an embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Figure 6 This is a schematic diagram of an embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Figure 7 This is a schematic diagram of the structural composition of a server provided in one embodiment of this specification; Figure 8 This is a schematic diagram of the structural composition of a cross-institutional health record data processing device based on multi-dimensional features, provided in one embodiment of this specification. Figure 9 This is a schematic diagram of one embodiment of the method for processing cross-institutional health record data based on multidimensional features provided in the embodiments of this specification, in a scenario example. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] It should be noted that the information and data related to users involved in the embodiments of this specification are all information and data authorized by the user or fully authorized by the relevant parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, and necessary confidentiality measures have been taken. They do not violate public order and good morals, and corresponding operation entry points are provided for users or relevant parties to choose to authorize or refuse.

[0021] It should also be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0022] See Figure 1 As shown in the embodiments of this specification, a method for processing cross-institutional health record data based on multi-dimensional features is provided. In specific implementation, this method may include the following: S101: Obtain multiple health record texts; and extract corresponding first type fields from the multiple health record texts; wherein, the first type fields are fields that contain at least identification indication information; S102: Based on the first type of field, obtain the identification indication information of the health record text; and based on the identification indication information, construct the identification blocking key of the health record text; S103: Construct a first node graph based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text; S104: Based on the preset rules for processing related evidence, use the identification information of the health record text to determine the related evidence data between two connected nodes in the first node diagram; S105: Based on the associated evidence data, calculate the weight parameters of the edges between two connected nodes; and mark the corresponding weight parameters on the edges to obtain the second node graph; S106: Based on the second node graph, multiple connected clusters are determined through clustering; wherein, one connected cluster corresponds to one user; S107: Based on the multiple connected clusters, the health record text groups of multiple users are determined by merging the corresponding health record texts. S108: Receive and respond to a request for querying health information of a subject of interest, identify the subject of interest and associated objects associated with the subject of interest, and query the health record text group of the corresponding user based on the consensus record identifier corresponding to the subject of interest and associated objects to obtain multiple health record texts; S109: Based on the multiple health record texts, extract the second type of fields related to the object of interest and the associated object, respectively, as target fields and auxiliary fields; the second type of fields are fields that contain at least the user's health information; and construct corresponding target health feature sequences and auxiliary health feature sequences based on the target fields and auxiliary fields; S1010: The target health feature sequence and auxiliary health feature sequence are processed using a preset health status prediction model to obtain the health status prediction result of the object of interest; wherein, the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0023] Specifically, the aforementioned method for processing cross-institutional health record data based on multi-dimensional features can be applied to the server side of a user health data service platform. This service platform can connect to databases of different institutions. The server can specifically include a backend server applied to the user health data service platform, capable of data transmission and data processing. Specifically, the server can be, for example, an electronic device with data processing, storage, and network interaction capabilities. Alternatively, the server can be a software program running on the electronic device, providing support for data processing, storage, and network interaction. In this embodiment, the number of servers is not specifically limited. The server can be a single server, several servers, or a server cluster formed by several servers.

[0024] With user authorization, this service platform can interact with databases of different institutions to obtain massive amounts of health record text containing user health information generated by different institutions at different times. Then, applying the multi-dimensional feature-based cross-institutional health record data processing method provided in the embodiments of this specification, the platform processes the massive and disorganized health record text, identifies multiple health record texts that accurately correspond to each user, and merges them to obtain health record text groups for each user. Furthermore, based on each user's health record text groups, a full health database can be constructed for each user, and the health information from each user's health record text groups can be integrated to generate a relatively rich and comprehensive health data fusion summary report for each user. This effectively achieves unified governance and integration of multi-source health record data across institutions, enabling users to efficiently and conveniently query and understand their own health status.

[0025] The aforementioned multiple health record texts may specifically include health record texts of different types generated by different users at different times and based on different institutions.

[0026] The types of institutions mentioned above may specifically include at least one of the following: hospitals, health check-up centers, medical laboratories, and clinics. The types of health record documents mentioned above may specifically include at least one of the following: physical examination reports, medical records, and laboratory test reports.

[0027] Considering the significant differences among different institutions in the standardization of generating and managing health record data, the rules for generating record numbers, and the habits of entering user information—for example, some institutions use physical examination numbers as record numbers, some generate new records for each physical examination of the same user, and some institutions habitually add remarks suffixes when entering user names, such as entering "Zhang San B-ultrasound" in the name field when entering information related to Zhang San's B-ultrasound report—the acquired health record text data is relatively chaotic, frequently resulting in problems such as multiple records for one person, multiple people for one record, conflicting record numbers across institutions, and inconsistent basic user information. Furthermore, due to the sheer volume of the aforementioned health record data, it is difficult to achieve unification and integration of this massive amount of health record data using existing methods.

[0028] Furthermore, the aforementioned chaos is even more complex due to the involvement of health record data from different periods and of different types. For example, early medical examination reports often contain incomplete user identity information; some institutions' health record texts lack user ID numbers; and due to negligence by the staff who created the records, user names may contain typos, making direct matching based on user names impossible. These issues result in not only information chaos in the massive amount of acquired health record texts, but also a wide variety of types and causes of chaos across different users, institutions, and periods. The chaos varies significantly across different health record data sets and cannot be predetermined, further increasing the difficulty of unifying and integrating this health record data.

[0029] To address the aforementioned issues and efficiently identify health record texts corresponding to each user from a vast amount of disorganized health record texts, the applicant, through creative thinking, considered firstly extracting the first type of fields from each record text. This involves acquiring attribute information for both the user and the health record text itself—two types of attribute information that, while different in nature, offer some indicative value. Furthermore, by combining the identifier indication information with a pre-defined identifier blocking key template, identifier blocking keys are constructed for each health record text. Based on these blocking keys, a preliminary screening can be performed relatively quickly to identify potentially related health record texts, constructing a first node graph containing multiple nodes corresponding to the health record texts. Then, by calculating the weight parameters between connected nodes based on the associated evidence data according to pre-defined association evidence processing rules, the first node graph is transformed into a second node graph. Finally, by performing clustering based on graph theory using the second node graph, a refined screening can be achieved more efficiently, automatically identifying groups of health record texts corresponding to each user. This allows for a more comprehensive consideration of various information chaos situations, making it well-suited for health big data governance scenarios. It enables efficient and accurate unified governance and effective integration of multi-source health record data from different users across institutions with relatively low data processing costs.

[0030] Specifically, the first type of field mentioned above can be understood as a field in the health record text that contains at least identification and indication information.

[0031] The aforementioned identifier blocking key can be understood as a feature-based basis constructed from the identifier indication information contained in each health record text, used to search for potentially related health record texts during the initial screening.

[0032] The first-node graph described above can be understood as a graph structure containing all the acquired health record texts. Specifically, the first-node graph can contain multiple nodes, each corresponding one-to-one with a health record text. See also... Figure 2 As shown, during the initial screening, nodes corresponding to two potentially related health record texts are connected by edges based on the identifier blocking key. Furthermore, each node is labeled with the identifier information of the health record text it corresponds to.

[0033] The aforementioned pre-defined rules for processing related evidence may specifically include rules for constructing related evidence and rules for using related evidence, as well as other related rule data. Specifically, related evidence can be understood as data evidence used to determine whether the health record texts corresponding to two connected nodes belong to the same user, based on a combination of identifier information.

[0034] The aforementioned second node graph can be understood as a graph structure that, based on the first node graph, further annotates the weight parameters on the edges between two connected nodes based on the corresponding related evidence data.

[0035] The aforementioned connected clusters can be understood as clusters of points determined through clustering based on the second node graph. Each connected cluster corresponds to a user and may contain multiple nodes corresponding to the same user.

[0036] In practice, the server can interact with the databases of different institutions according to the corresponding protocol rules to obtain the full health record texts of each institution's database, thus obtaining multiple health record texts.

[0037] In practice, the server can determine multiple health record texts for each user based on the multiple connected clusters; then, for each user, a unified consensus record identifier is generated based on the corresponding multiple health record texts; and then, the consensus record identifier is used to merge the multiple health record texts of the user to obtain the user's health record text group, thereby realizing the unification and integration of multi-source health record data of users across institutions and time periods.

[0038] After unifying and integrating massive amounts of multi-source health record data from various institutions and time periods according to the above method, and obtaining relatively stable and reliable long-term accumulated health record data for each user, the unified and integrated health record data can be further utilized to conduct long-term time series analysis of the user's health characteristics to determine and predict the user's current health status and the trend of health status development in future time periods, so as to provide users with more credible and highly valuable health consultation services.

[0039] In specific implementation, the system can further receive health information query requests regarding a target of interest; respond to these requests by identifying the target of interest and other target objects that match and are associated with it in terms of physical data and pathological characteristics; determine the corresponding consensus archive identifier based on the target of interest's user identifier; and query multiple health record texts in the corresponding user's health record text group based on this consensus archive identifier to extract the second type of field (denoted as the target field) related to the target of interest, as well as the text generation date corresponding to the aforementioned second type of text field. Similarly, based on the consensus archive identifier corresponding to the associated object, the system can query and extract the second type of field (denoted as the auxiliary field) related to the associated object, as well as the text generation date corresponding to the aforementioned second field. The associated object can be one or multiple.

[0040] Next, the corresponding target health features can be extracted from the target field and arranged in chronological order based on the text generation date to obtain a target health feature sequence. At the same time, the corresponding auxiliary health features can be extracted from the auxiliary field and arranged in chronological order based on the text generation date to obtain an auxiliary health feature sequence. The above target health feature sequence and auxiliary health feature sequence are then processed using a preset health status prediction model to obtain the health status prediction result of the object of interest. The above health status prediction result is used to characterize the current health status of the object of interest and the future trend of its health status.

[0041] The aforementioned preset health status prediction model includes at least a first health prediction branch and a second health prediction branch connected in parallel. The first health prediction branch processes the target health feature sequence and outputs a corresponding target health feature vector. The second health prediction branch processes the auxiliary health feature sequence and outputs a corresponding auxiliary health feature vector as an auxiliary reference. The preset status prediction model also includes a joint regression analysis model. This joint regression analysis module is connected to the outputs of the first and second health prediction branches and is used to receive and perform joint regression calculations based on the target health feature vector and the auxiliary health feature vector to determine and output the final health status prediction result of the target object.

[0042] Specifically, the aforementioned first health prediction branch may include a first feature processing module based on an LSTM (Long Short-Term Memory) structure and a first hidden information processing module based on a dual GRU structure. Based on this structure, the first health prediction branch can perform refined temporal processing on the health data accumulated over a long period of time for the object of interest.

[0043] The aforementioned second health prediction branch can include a model network based on an LSTM structure. Compared to the first health prediction branch, this second health prediction branch can more quickly and effectively organize and extract health features from long-term accumulated health data of the associated objects, providing supplementary reference value for predicting and analyzing the health status of the objects of interest.

[0044] Specifically, when processing the target health feature sequence and auxiliary health feature sequence using the preset health status prediction model, the first feature processing module in the first health prediction branch can be used to process the input target feature sequence to obtain the corresponding global health feature vector. The global health feature vector includes multiple health feature groups arranged in chronological order, with each health feature group corresponding to a time point. The dual GRU structure in the first hidden information processing module is used to analyze and process each health feature group in the global health feature vector along both the forward and reverse directions to obtain the forward and reverse hidden state information of each health feature group. These information are then combined to generate the corresponding target health feature vector. Simultaneously, the second health prediction branch performs temporal feature processing on the auxiliary health feature sequence to obtain the corresponding auxiliary health feature vector. Next, the joint regression analysis module can be used to combine the target health feature vector and the auxiliary health feature vector, using the auxiliary health feature vector as an auxiliary reference, to perform regression analysis and prediction on the health status of the object of interest based on the target health feature vector, thereby determining the predicted health status of the object of interest.

[0045] Furthermore, based on the predicted health status of the target individual, the current health status and future trends of their health status can be determined. Then, based on these predictions, matching health advice can be generated for that individual, providing users with highly valuable health consultation services.

[0046] Based on the above embodiments, by constructing and using identifier blocking keys, node graphs, and associated evidence data, it is possible to effectively address the information chaos in various health record texts. It is well-suited for health big data governance scenarios and can efficiently and accurately identify health record data corresponding to the same user from massive amounts of health record data with relatively small data processing volume. This enables unified governance and effective integration of multi-source health record data from different users across institutions. Furthermore, through a preset health status prediction model, it is possible to effectively utilize the unified and integrated user health record text groups to comprehensively and accurately analyze the health status of the target audience, providing users with health consultation services with high reference value.

[0047] In some embodiments, the identification information may specifically include: attribute information of the health record text and user attribute information, etc. The attribute information of the health record text may specifically include at least one of the following: the health record text's file number, the physical examination number associated with the health record text, the physical examination date associated with the health record text, the name of the institution that generated the health record text, and the text generation date of the health record text. The user's attribute information may specifically include at least one of the following: user's ID card number, user's name, user's date of birth, user's mobile phone number, user's age, name of the organization to which the user belongs, user's gender, etc.

[0048] Of course, it should be noted that the attribute information of the health record text and the user listed above is only illustrative. In actual implementation, depending on the specific circumstances and processing requirements, the attribute information of the health record text and the user may include other appropriate information. This instruction manual does not limit this.

[0049] In some embodiments, the extraction of the corresponding first type of fields from the plurality of health record texts described above may include the following: S1: Perform preliminary information extraction on the multiple health record texts to determine the name of the organization that generated the health record text and the text type of the health record text; S2: Based on the name of the generating institution and the text type of the health record text, retrieve the preset text template set and determine the matching target text template; S3: Based on the target text template, determine the corresponding text area in the health record text; and extract the corresponding fields from the text area to obtain the first type of fields of the health record text.

[0050] Specifically, the aforementioned preset text template set can be determined in advance through data statistics and clustering learning of a large number of sample health record texts from different institutions and of different text types.

[0051] The aforementioned preset text template set may contain multiple preset text templates, and each preset text template corresponds to at least one combination of organization name and text type.

[0052] When specifically obtaining identification information, for a user's mobile phone number, a pre-defined regular expression can be used to quickly extract the 11-digit mobile phone number from the complex health record text as a type of identification information.

[0053] In some embodiments, obtaining the identification information of the health record text based on the first type of field includes: processing the first type of field using a preset text field processing model to obtain the identification information of the health record text.

[0054] The preset text field processing model is an algorithm model based on NLP (Natural Language Processing) technology, which is trained in advance using a large amount of sample corpus based on sample health record texts through unsupervised learning.

[0055] In some embodiments, after obtaining the identification information of the health record text, the method may further include the following: S1: Detect the character format of the identification information to obtain a first detection result; and based on the first detection result, convert the identification information with full-width character format to half-width character format to obtain the first processed identification information; S2: Based on the first processed identification information, detect whether there are redundant symbols (e.g., parentheses, periods, etc.) in the identification information to obtain a second detection result; according to the second detection result, for the identification information with redundant symbols, use a preset symbol semantic recognition model to perform semantic recognition; according to the semantic recognition result, determine the invalid symbols for the identification information from the redundant symbols, delete the invalid symbols, and obtain the second processed identification information. S3: Based on the second processed identification indication information, detect whether a user name exists and obtain a third detection result; according to the third detection result, when a user name exists, further detect whether there is a prefix or suffix of the user name; for cases where there is a prefix or suffix, perform semantic recognition on the identification information; according to the semantic recognition result, when it is determined that the prefix or suffix is ​​invalid information, delete the prefix or suffix and obtain the third processed identification indication information. S4: Based on the third processed identification information, detect whether there is a user ID number to obtain the fourth detection result; according to the fourth detection result, when a user ID number is detected, infer the user's gender and date of birth using the user ID number according to the corresponding numbering rules; and cross-validate and update the user's attribute information based on the user's gender and date of birth to obtain the fourth processed identification information. S5: Based on the fourth processed indicator information, detect whether there is a physical examination date and obtain the fifth detection result; according to the fifth detection result, when there is a physical examination date, unify the physical examination date to a preset date format and obtain the fifth processed indicator information; S6: Based on the fifth processed identification information, detect whether there is missing identification information to obtain the sixth detection result; according to the sixth detection result, mark the missing information of the identification information of the health record to obtain the sixth processed indicator identification information, so that the identification blocking key corresponding to the health record text can be quickly constructed and used according to the missing information mark.

[0056] In practical implementation, the system can also detect the text type of identifiers containing written characters (e.g., the name of the organization that generated the health record text, the user's name, the name of the unit to which the user belongs, etc.) and determine whether traditional Chinese characters are present. When traditional Chinese characters are detected, they are modified to the corresponding simplified Chinese characters according to a preset traditional-simplified Chinese character mapping table. The text types include both traditional and simplified Chinese characters.

[0057] In practice, the identification information can also be detected and cleaned of leading invalid characters to identify and remove leading invalid characters such as numbers and special symbols that have no actual semantic meaning before the identification information such as the user's name, so as to avoid interference with subsequent data processing.

[0058] In practice, the identification and indication information of the health record text can also be obtained and updated based on the target template of the health record text.

[0059] Based on the above embodiments, by performing relatively comprehensive and complete preprocessing on the acquired signage indication information, possible error information in the signage indication information can be effectively eliminated, and the signage indication information can be made relatively uniform and standardized, thus obtaining signage indication information suitable for further processing.

[0060] In some embodiments, the above-described method of constructing the identifier blocking key for the health record text based on the identifier indication information may, in specific implementation, include: Based on the preset identifier blocking key template and the missing information markers of the identifier indication information of the health record text, a group key pattern suitable for the health record text is determined. Then, using the identifier indication information of the health record text, an identifier blocking key is constructed based on the group key pattern for the health record text. A single health record text identifier blocking key can contain multiple different identifier blocking keys, each corresponding to a group key pattern.

[0061] The aforementioned preset identifier blocking key template may include multiple optional group key patterns. Specifically, the identifier blocking key may include at least one of the following: the year of birth in the user's name and date of birth, the last four digits of the user's name and mobile phone number, the user's ID card number, the name of the organization to which the user belongs and the user's name, the name of the organization that generated the information and the user's name, etc.

[0062] The specific preset identifier blocking key templates mentioned above can be found in Table 1.

[0063] Table 1

[0064] The multiple component patterns included in the aforementioned preset identifier blocking key template can be determined as follows: A certain number of sample health record texts are obtained; multiple key patterns are determined based on the identifier indication information; the statistical distribution results of various key patterns based on the full sample health record texts are statistically analyzed; simultaneously, the aforementioned sample health record texts are divided into health record text sets for multiple users, and the proportion of each key pattern based on each user's health record text set is statistically analyzed; finally, the statistical distribution results of various key patterns based on the full sample health record texts are combined, and each key pattern is determined based on each user's health record text set. The proportion data of health record text sets is used to determine the first type of key pattern from various key patterns. This first type of key pattern is characterized by the proportion of health record text sets of most users exceeding a preset threshold and the data values ​​differing among different users' health record text sets. Then, based on the statistical distribution results of the full sample of health record texts, the second type of key pattern is selected from the remaining key patterns other than the first type of key pattern. This second type of key pattern has a good statistical distribution and can cover most of the remaining health record text sets of users who do not include the first type of key pattern. Finally, the first and second type of key patterns are combined to obtain the preset identifier blocking key template.

[0065] In some embodiments, see Figure 3 As shown, the first node graph is constructed based on the identifier indication information and identifier blocking key of the health record text. In specific implementation, it may include the following: S3-1: Based on the identification information of the health record text, set up a node corresponding to the health record text in the initial node diagram; and mark the identification information of the health record text on the node; S3-2: Based on the identifier blocking key of the health record text, identify two potentially related health record texts; wherein, at least one identifier blocking key is the same between the two potentially related health record texts; S3-3: Connect the two nodes corresponding to two potentially related health record texts using edges to obtain the corresponding first node graph.

[0066] The above method of identifying two potentially related health record texts based on the identifier blocking key of the health record texts may, in a specific implementation, include: detecting whether there is a common identifier blocking key among the identifier blocking keys of any two health record texts; when it is determined that there is a common identifier blocking key among the identifier blocking keys of the two health record texts, the two health record texts can be determined to be two potentially related health record texts.

[0067] The above-mentioned method of identifying two potentially related health record texts based on their identifier blocking keys can, in a more specific implementation, include: when it is determined that the two health record texts contain the same identifier blocking key, identifying the same identifier blocking key in the two health record texts; wherein the same identifier blocking key includes one or more identifier blocking keys; calculating the information entropy of the same identifier blocking key based on the same identifier blocking key; detecting whether the information entropy of the same identifier blocking key is greater than a preset reference threshold; and determining that the two health record texts are potentially related health record texts when the information entropy is greater than the preset reference threshold.

[0068] Furthermore, the aforementioned identifier blocking key may specifically include a first type of identifier blocking key and a second type of identifier blocking key. The first type of identifier blocking key contains part or all of the user's ID number. The second type of identifier blocking key contains only information other than the user's ID number.

[0069] Accordingly, the above-mentioned method of identifying two potentially related health record texts based on the identifier blocking key of the health record text may, in specific implementation, also include: S1: Detect whether there is a common type 1 identifier blocking key in the identifier blocking keys of any two health record texts; S2: When there is no common first-type identifier blocking key in the identifier blocking keys of the two health record texts, detect whether there is a common second-type identifier blocking key in the identifier blocking keys of the two health record texts; S3: When the two health record texts contain the same second-type identifier blocking key, check whether the number of the same second-type identifier blocking key is greater than a preset number threshold. S4; When the number of identical second-class identifier blocking keys is greater than or equal to a preset number threshold, the two health record texts are determined to be two potentially related health record texts.

[0070] Specifically, if at least one of the first-type identifier blocking keys is the same in the identifier blocking keys of the two health record texts mentioned above, the two health record texts can be directly identified as potentially related health record texts.

[0071] Specifically, when the two health record texts do not have the same first type of identifier blocking key and do not have the same second type of identifier blocking key, or when the number of the same second type of identifier blocking key is less than a preset number threshold, the two health record texts can be determined to be two potentially related health record texts.

[0072] Using the above method, the identifier blocking keys of multiple health record texts can be quickly traversed to identify potentially related health record texts.

[0073] In practice, a large amount of test data can be obtained through testing experiments. Based on this data, big data analysis can be conducted to determine the screening effectiveness of each blocking key in the initial screening process. The priority of the blocking keys can then be determined based on the screening effectiveness. The priority of the blocking keys includes at least two categories: first priority and second priority, with the first priority being higher than the second priority.

[0074] Accordingly, in specific implementation, the first screening can be carried out based on the identification information of the health record text and the first priority identification blocking key to construct the first node graph.

[0075] After obtaining the first node graph, it can be divided into multiple blocks; any node within each block can be connected to any other node within the same block directly or indirectly. The system checks whether the number of nodes in each block exceeds a preset threshold (e.g., 800).

[0076] When the number of nodes in a block exceeds a preset threshold, a second screening can be performed on that block based on the identifier information of the health record text corresponding to the nodes in that block and the second priority identifier blocking key, in order to construct an optimized first node graph.

[0077] Based on the above embodiments, the identification blocking key can be used to quickly perform a preliminary screening of a large number of health record texts with a relatively small amount of data processing, accurately identify potentially related health record texts, and construct the corresponding first node graph.

[0078] In some embodiments, the above-mentioned method uses the identification information of the health record text to determine the association evidence data between two connected nodes in the first node graph according to preset processing rules. For specific implementation, please refer to [reference needed]. Figure 4As shown, the evidence data linking the first and second connected nodes can be determined in the following manner: S4-1: Use a preset feature matching model to process the identification information of the first node and the identification information of the second node, and calculate the information matching degree of the corresponding identification information of the first node and the second node. S4-2: Based on the preset rules for processing related evidence, and according to the information matching degree of the corresponding identification indication information of the first node and the second node, determine and use the identification indication information of the first node and the second node that meet the requirements for configuration, so as to construct one or more pieces of related evidence. S4-3: Combine the aforementioned related evidence to obtain the related evidence data between the first node and the second node.

[0079] In this context, the first node and the second node are any two nodes in the first node graph that are connected by an edge.

[0080] In practice, a preset feature matching model can be used to calculate the semantic distance between the corresponding identifiers and indications of the first and second nodes to obtain the corresponding information matching degree. This information matching degree characterizes the degree of similarity between the identifiers and indications of the first and second nodes; a higher information matching degree indicates a higher degree of similarity.

[0081] In specific implementation, based on preset related evidence processing rules, identifiers with a matching degree greater than a preset matching degree threshold (e.g., 95%) can be selected as compliant identifiers based on the matching degree of the identifiers corresponding to the first and second nodes. Then, based on the compliant identifiers of the first and second nodes, the configurable optional evidence types are determined according to the preset related evidence processing rules. Next, the compliant identifiers of the first and second nodes are configured according to the optional evidence types to construct one or more related pieces of evidence. Finally, based on preset related feature processing rules, the aforementioned related evidence is combined to obtain related evidence data between the first and second nodes. This related evidence data may include one or more pieces of related evidence.

[0082] The types of associated evidence determined based on the preset rules for processing associated evidence may include at least one of the following: a configuration combination of user ID number and user name; a configuration combination of user ID number; a configuration combination of user name and user date of birth; a configuration combination of user name and user mobile phone number; a configuration combination of the name of the user's employer, user name, the last few digits of the user's mobile phone number, and the year of birth in the user's date of birth; a configuration combination of the consistency judgment result of user name, the name of the user's employer, and the text generation date of the user's age and health record text, etc. Specifically, the types of associated evidence determined based on the preset rules for processing associated evidence can be found in Table 2.

[0083] Table 2

[0084] It should be noted that the consistency judgment result between the user's age and the text generation date of the health record text specifically refers to whether the absolute value of the difference between the first difference in user age corresponding to the first node and the second node and the second difference in the date of the health record text corresponding to the first node and the second node is less than or equal to a preset age threshold (e.g., 1.2 years). When the absolute value of the difference between the first and second differences is less than or equal to the preset age threshold, it can be determined that the requirement is met, indicating that it is very likely that the same user participated in different health services at different times.

[0085] Based on the above embodiments, the associated evidence data between two connected nodes can be accurately determined according to the preset associated evidence processing rules and the first node graph.

[0086] In some embodiments, calculating the weight parameter of the edge between two connected nodes based on the associated evidence data may include: calculating the weight parameter of the edge between the first and second connected nodes in the following manner: S1: Determine the evidence type of the related evidence according to the preset rules for processing related evidence; S2: Based on multiple related pieces of evidence and the types of evidence associated with them, determine the similarity score between the first node and the second node; S3: Based on the preset rules for processing related evidence, the weight parameters of the connection between the first and second connected nodes are calculated using the similarity score.

[0087] The aforementioned related evidence, based on its effect of indicating user uniqueness, can be further divided into strong supporting evidence and weak supporting evidence. For example, key evidence that typically includes an ID number is often strong supporting evidence. Generally, related evidence that is strong supporting evidence has a higher similarity score than weak supporting evidence.

[0088] In practice, the corresponding similarity score range can be determined first based on the evidence type of the related evidence and the preset similarity score mapping relationship of the related evidence; then, based on the similarity score range and the specific data value of the related evidence, the similarity score of the related evidence can be determined.

[0089] Before implementation, a pre-defined similarity scoring mapping relationship for related evidence can be constructed as follows: Multiple sample-related pieces of evidence are identified based on the sample health record texts; based on multiple pre-defined correlation analysis rules, using multiple sample-related pieces of evidence from different sample health record texts and corresponding user results (whether the sample health record texts correspond to the same user), various correlation analysis results of each sample-related piece of evidence with respect to the user results are calculated; based on the various correlation analysis results of each sample-related piece of evidence with respect to the user results, through statistical data analysis, similarity scoring intervals corresponding to the evidence types of the multiple sample-related pieces of evidence are determined; combining the above multiple similarity scoring intervals yields the pre-defined similarity scoring mapping relationship for related evidence.

[0090] Specifically, the aforementioned pre-defined correlation analysis rules may include: correlation analysis rules based on Pearson correlation coefficient, correlation analysis rules based on Spearman correlation coefficient, correlation analysis rules based on Kendall correlation coefficient, etc. It should be noted that the pre-defined correlation analysis rules listed above are merely illustrative. In actual implementation, other types of correlation analysis rules may be included depending on the specific circumstances and processing requirements. This specification does not limit this.

[0091] In practice, the total similarity score can be calculated using the similarity score according to the preset rules for processing related evidence; and the weight parameters of the connection between the first and second connected nodes can be determined based on the total similarity score.

[0092] Specifically, a higher overall similarity score indicates a greater likelihood that the health record texts corresponding to the two nodes linked by the evidence belong to the same user. In this case, the weight parameter of the edge connecting the two nodes can be set relatively higher. This weight parameter is greater than or equal to 0 and less than or equal to 1.

[0093] When the weight parameter of the edge connecting the first node and the second node is 0, it can be directly determined that the users corresponding to the health record texts of the first node and the second node are not the same user.

[0094] Conversely, when the weight parameter of the edge connecting the first node and the second node is 1, it can be directly determined that the user corresponding to the health record text of the first node and the second node is the same user.

[0095] In some embodiments, the method may further include the following: S1: Detect whether the key indication information in the identification indication information of the first node and the identification indication information of the second node is empty; wherein, the key indication information includes: user ID number, user gender, and user date of birth; S2: When the key indication information in the first node's identification indication information and the second node's identification indication information is not empty, check whether the key identification indication information of the first node and the key identification indication information of the second node are the same; S3: When the key identifier information of the first node is different from that of the second node, the weight parameter of the edge between the first node and the second node is set to 0.

[0096] Based on the above embodiments, key indication information with relatively high reliability and relatively good indication effect can be introduced and used to simplify the processing process, identify two nodes that cannot correspond to the same user in advance, and avoid further fine-tuning of data processing resources for the two nodes in the future. This can reduce the overall data processing cost and improve the overall processing efficiency to a certain extent.

[0097] In some embodiments, the above-described determination of multiple connected clusters based on the second node graph through clustering processing may specifically include: detecting and determining corresponding connected components based on the edges between nodes in the second node graph and the weight parameters of the edges, according to a preset node graph traversal processing rule (e.g., improving a disjoint-set data structure-based clustering algorithm based on graph theory to obtain an algorithm rule suitable for processing node graphs); and determining multiple connected clusters based on the connected components. Each connected cluster contains multiple nodes, and each connected cluster corresponds to one user. Accordingly, the complete health record text of the user corresponding to a connected cluster can be determined based on that connected cluster.

[0098] In some embodiments, after determining multiple connected clusters through clustering based on the second node graph, the method may further include: determining valid association evidence and signature features for each connected cluster; wherein, the aforementioned valid association evidence can be specifically understood as association evidence that can be satisfied by the vast majority of nodes (e.g., 90% of nodes) in the same connected cluster; based on the valid association evidence of each connected cluster, detecting whether there is the same strong evidence between different connected clusters, and obtaining a strong evidence detection result; wherein, the strong evidence can be specifically understood as association evidence with a reliability greater than a preset reliability threshold; based on the strong evidence detection result and signature features, determining whether different connected clusters satisfy bridging conditions with each other; and recording connected clusters that satisfy bridging conditions with each other as candidate bridging connected clusters; then, according to preset security check rules, performing security checks on the candidate bridging connected clusters that satisfy bridging conditions with each other; wherein, the security check at least includes checking whether there are conflicts in user ID numbers and user birth dates between candidate bridging connected clusters; when it is determined that the security check is passed, merging the aforementioned candidate bridging connected clusters that satisfy bridging conditions with each other into one connected cluster.

[0099] In some embodiments, the above-described method of determining multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters may specifically include the following: S1: Based on the multiple connected clusters, determine multiple health profile texts for multiple users; S2: Based on the preset file merging rules, determine the consensus file identifier for multiple health file documents for each user; S3: Using consensus archive identifiers, merge and combine multiple health archive texts from various users to obtain a health archive text group for multiple users.

[0100] The aforementioned consensus record identifier corresponds one-to-one with a user and can be used to indicate all health record texts of the corresponding user.

[0101] The above-mentioned consensus file identifier is used to merge and combine multiple health record texts of various users to obtain a health record text group of multiple users. In specific implementation, it may include: using the consensus file identifier to associate and index the multiple health record texts of users; then combining the associated and indexed health record texts to obtain the health record text group of users; and storing the health record text group in the platform database.

[0102] Based on the above embodiments, by introducing and using a consensus archive identifier unique to each user to merge and combine multiple health record texts for that user, it is possible to effectively unify and integrate multiple health record texts of different types generated by different institutions at different times for that user. Subsequently, the user only needs to use this single consensus archive identifier to efficiently query and retrieve all health record texts corresponding to that user.

[0103] In some embodiments, see Figure 5 As shown, the above-mentioned consensus file identifier for multiple health record files of each user is determined according to the preset file merging rules. In specific implementation, it may include the following: S5-1: Check if the file number identifier field of multiple health record texts of a user is empty; S5-2: When at least one of the file identifier fields in the file number identifier field of a user's multiple health record texts is not empty, extract the existing file identifier based on the file number identifier field of the user's health record text. S5-3: Check if the number of existing file identifiers is greater than 1; S5-4: When the number of existing archival identifiers is greater than 1, determine the number of existing archival identifiers and the generating organization for each type; S5-5: Based on the number of existing archival identifiers of each type and the generating institution, identify the existing archival identifiers that meet the requirements as candidate archival identifiers. S5-6: Detect whether the candidate file identifier overlaps with the consensus file identifier of other users; S5-7: When the candidate file identifier overlaps with the consensus file identifier of other users, according to the preset identifier generation rules, the consensus file identifier of multiple health file documents for that user is generated using the candidate file identifier.

[0104] In practice, when the file identifier field of multiple health record texts of a user is empty, a consensus file identifier for multiple health record texts of that user can be generated according to the preset identifier generation rules.

[0105] Specifically, it may include: according to the preset identifier generation rules, based on the identifier indication information, obtaining and generating the corresponding pinyin letter combination according to the user's name; at the same time, using a random number generator to generate a random number sequence of a specified number of digits (e.g., 8 digits); and then concatenating the above pinyin letter combination and random number sequence to obtain the corresponding consensus file identifier.

[0106] After obtaining the consensus file identifier, the platform database can be searched based on this identifier to determine whether there is any overlap. If overlap is found, an updated random number sequence is generated using a random number generator according to the preset identifier generation rules; this updated random number sequence is then used to replace the previous random number sequence in the consensus file identifier to obtain an updated consensus file identifier; finally, the updated consensus file identifier is checked for overlap.

[0107] In practice, when the number of existing file identifier types is equal to 1, the unique file identifier can be directly obtained based on the identifier indication information. The platform database is then searched based on this file identifier to determine if there is any overlap in the consensus file identifiers. If there is no overlap, the file identifier can be directly identified as the consensus file identifier for multiple health records of the user. If there is overlap, a random number sequence can be generated using a random number generator according to preset identifier generation rules. The file identifier is then concatenated with the random number sequence to obtain the corresponding consensus file identifier.

[0108] In practice, a weighted score can be determined for each type of existing archival identifier based on the quantity and generating institution of each identifier. The higher the level of the generating institution (e.g., a top-tier hospital), the higher the weighted score for the majority of existing archival identifiers. Based on the weighted scores, the archival identifier with the highest weighted score is selected as a candidate archival identifier.

[0109] In practice, when a candidate file identifier does not overlap with the consensus file identifiers of other users, the candidate file identifier can be directly identified as the consensus file identifier for multiple health file documents of that user. Conversely, when a candidate file identifier overlaps with the consensus file identifiers of other users, a random number sequence can be generated according to a preset identifier generation rule; then, the candidate file identifier is concatenated with the random number sequence to obtain the corresponding consensus file identifier.

[0110] In some embodiments, after determining multiple user health record text groups by merging the corresponding health record texts according to the plurality of connected clusters, refer to... Figure 6 As shown, in specific implementations, the method may also include the following: S6-1: Based on the user's health record text group, extract the second type of fields from the health record text; wherein, the second type of fields are fields that contain at least the user's health information; S6-2: Determine the reliability weight coefficient of the second type of field based on the institution level and file type of the institution that generated the health record text; S6-3: By performing semantic recognition on the second type of field, the target object indicated by the second type of field is determined; S6-4: Detect whether there is overlap between the target objects indicated by the second type of field; S6-5: When it is determined that the target objects indicated by the second type of field overlap, the second type of field in which the indicated target objects overlap is selected as the first candidate field; S6-6: Based on the text generation date of the health record text to which the first candidate field belongs, filter out the second candidate field whose text generation date meets the requirements; S6-7: Check if the number of second candidate fields is greater than 1; S6-8: When the number of second candidate fields is greater than 1, according to the preset archive data fusion rules, multiple second candidate fields are fused using the reliability weight coefficient to generate a health information consensus field about the health information of the user's target object; wherein, the health information consensus field is used to generate a health data fusion summary report for the user.

[0111] The second type of field mentioned above can be specifically understood as fields in the health record text that contain the user's health information, such as the medical examination conclusions in a physical examination report or the examination results in a medical record. The target object mentioned above can specifically be the user's tissues or organs, etc.

[0112] When it is determined that the target objects indicated by the second type of field do not overlap, a health information consensus field about the user's target object's health information can be directly generated based on the second type of field.

[0113] When it is determined that the target objects indicated by the second type of field overlap, it indicates that multiple health record texts containing health information of the target object exist simultaneously in the user's health record text set. Then, based on the text generation date, the latest health record text can be determined; the text generation date of this health record text is then used to determine the end date; next, it is checked whether there is a latest neighboring health record text whose time interval between the text generation date and the end date is less than a preset effective duration threshold. If not, the first candidate field in the latest health record text can be determined as a qualified second candidate field. If it exists, the latest health record text and the first candidate field in the latest neighboring health record text can be determined as qualified second candidate fields.

[0114] When the number of second candidate fields is equal to 1, a consensus field for health information about the user's target object can be directly generated based on the second candidate field.

[0115] When the number of second candidate fields is greater than 1, a consensus health information field can be generated by integrating multiple second candidate fields according to the preset archive data fusion rules and using the reliability weight coefficient. This field combines different health information about the target object provided by different institutions over a period of time and can comprehensively represent the target object's relatively complete latest health status.

[0116] Furthermore, based on the generated health information consensus fields, a user's health data fusion summary report can be automatically generated. This health data fusion summary report can effectively integrate multiple health record texts corresponding to the user, comprehensively reflecting the health information of multiple target objects of the user's body.

[0117] Based on the above embodiments, by extracting and utilizing the second type of fields, the integration of health information from multiple health record texts of a user can be effectively achieved.

[0118] In some embodiments, after determining multiple user health record text groups by merging the corresponding health record texts according to the multiple connected clusters, the method may further include the following: The data processing involves storing a master data table, a merged evidence table, and a conflict audit table for each user. The master data table includes consensus profile identifiers and the correspondence between these identifiers, users, and health record texts. The merged evidence table includes associated evidence generated and used during the data processing, weight parameters determined based on the associated evidence, and clustering records of connected clusters generated based on the weight parameters. The conflict audit table includes records of the determination and modification of consensus profile identifiers.

[0119] In practice, the master data table, merged evidence table, and conflict audit table for each user can be stored in the blockchain for evidence preservation, so as to facilitate subsequent backtracking and querying.

[0120] As can be seen from the above, the cross-institutional health record data processing method based on multi-dimensional features provided in the embodiments of this specification, after obtaining multiple health record texts of different text types generated by different institutions, can first extract the first type of fields of the health record texts based on the multiple health record texts; and obtain the identification indication information of the health record texts based on the first type of fields; then construct the identification blocking key of the health record texts based on the identification indication information; and construct a first node graph based on the identification indication information and the identification blocking key of the health record texts; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the nodes are marked with the identification indication information of the corresponding health record texts; according to preset processing rules, the identification indication information of the health record texts is used to process the data. The system identifies the association evidence data between two connected nodes in the first node graph based on the identification information. Based on this association evidence data, it calculates the weight parameters of the edges between the two connected nodes and labels the corresponding weight parameters on the edges, resulting in a second node graph. Based on the second node graph, it uses clustering to identify multiple connected clusters, where each connected cluster corresponds to a user. Based on these multiple connected clusters, it merges the corresponding health record texts to identify multiple user health record text groups. It then receives and responds to health information query requests for a target of interest, identifying the target of interest and associated objects. Finally, it uses a pre-set health status prediction model to determine the predicted health status of the target of interest based on the health record text groups of the target of interest and associated objects. This system is well-suited for health big data governance scenarios, efficiently and accurately identifying health record data corresponding to the same user from massive amounts of chaotic health record data, thus achieving unified governance and effective integration of multi-source health record data from different users across institutions. Furthermore, the pre-set health status prediction model effectively utilizes the unified and integrated user health record text groups to comprehensively and accurately analyze the health status of the target of interest.

[0121] This specification provides an embodiment of a server, see below. Figure 7 As shown. The server includes a network communication port 701, a processor 702, and a memory 703. These structures are connected by internal cables so that they can perform specific data interaction.

[0122] Specifically, the network communication port 701 can be used to acquire multiple health record texts and extract corresponding first type fields from the multiple health record texts; wherein the first type fields are fields that contain at least identification indication information.

[0123] The processor 702 can specifically be used to: obtain the identification indication information of health record text according to the first type of field; construct the identification blocking key of health record text according to the identification indication information; construct a first node graph according to the identification indication information and the identification blocking key of health record text; wherein the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text; determine the association evidence data between two connected nodes in the first node graph using the identification indication information of the health record text according to the preset association evidence processing rules; calculate the weight parameter of the connection between two connected nodes according to the association evidence data; and mark the corresponding weight parameter on the connection to obtain a second node graph; determine multiple connected clusters through clustering processing according to the second node graph; wherein one connected cluster corresponds to one user; determine multiple user health record text groups by merging the corresponding health record texts according to the multiple connected clusters; and so on. The system receives and responds to health information query requests for targets of interest, identifies the targets of interest, and identifies associated targets. It then queries the corresponding user's health record text group based on the consensus record identifiers corresponding to the targets of interest and associated targets, obtaining multiple health record texts. Based on these multiple health record texts, it extracts second-type fields related to the targets of interest and associated targets, serving as target fields and auxiliary fields, respectively. The second-type fields are fields that contain at least the user's health information. Based on the target fields and auxiliary fields, it constructs corresponding target health feature sequences and auxiliary health feature sequences. A preset health status prediction model is used to process the target health feature sequences and auxiliary health feature sequences to obtain the health status prediction result for the targets of interest. The preset health status prediction model includes at least a first health prediction branch and a second health prediction branch connected in parallel, as well as a joint regression analysis model. The first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0124] The memory 703 can be used to store the corresponding instruction program and related intermediate data.

[0125] Based on the above methods, the relevant structural performance of the server can be effectively utilized to improve the data processing speed of electronic devices and efficiently realize the processing of health record data.

[0126] In this embodiment, the network communication port 701 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0127] In this embodiment, the processor 702 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0128] In this embodiment, the memory 703 may include multiple layers. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0129] This specification also provides a computer-readable storage medium based on the above-described method for processing cross-institutional health record data based on multidimensional features. The computer-readable storage medium stores computer program instructions that, when executed, implement the following: acquiring multiple health record texts; extracting corresponding first-type fields from the multiple health record texts; wherein the first-type fields are fields containing at least identification indication information; acquiring identification indication information of the health record texts based on the first-type fields; constructing identification blocking keys for the health record texts based on the identification indication information; constructing a first node graph based on the identification indication information and the identification blocking keys of the health record texts; wherein the first node graph contains multiple nodes, each node corresponding to one health record text, and the nodes are labeled with the corresponding identification indication information of the health record texts; determining the association evidence data between two connected nodes in the first node graph using the identification indication information of the health record texts according to preset association evidence processing rules; calculating the weight parameters of the edges between the two connected nodes based on the association evidence data; and labeling the corresponding weight parameters on the edges to obtain a second node graph; and determining the second node graph through clustering processing. Multiple connected clusters; where each connected cluster corresponds to one user; based on the multiple connected clusters, by merging the corresponding health record texts, multiple user health record text groups are determined; a query request for health information of a target being monitored is received and responded to, the target being monitored and associated objects are determined; and the corresponding user health record text groups are queried based on the consensus record identifiers corresponding to the target being monitored and associated objects, resulting in multiple health record texts; based on the multiple health record texts, second-type fields related to the target being monitored and associated objects are extracted as target fields and auxiliary fields, respectively; the second-type fields are fields that contain at least the user's health information; and based on the target fields and auxiliary fields, corresponding target health feature sequences and auxiliary health feature sequences are constructed; the target health feature sequences and auxiliary health feature sequences are processed using a preset health status prediction model to obtain the health status prediction result of the target being monitored; wherein the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0130] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0131] In this embodiment, the specific functions and effects implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other embodiments, and will not be repeated here.

[0132] This specification also provides a computer program product, comprising at least a computer program, which, when executed by a processor, implements the following method steps: acquiring multiple health record texts; and extracting corresponding first-type fields from the multiple health record texts; wherein the first-type fields are fields that at least contain identification indication information; acquiring identification indication information of the health record texts based on the first-type fields; and constructing identification blocking keys for the health record texts based on the identification indication information; constructing a first node graph based on the identification indication information and the identification blocking keys of the health record texts; wherein the first node graph contains multiple nodes, one node corresponding to one health record text, and the nodes are marked with the corresponding identification indication information of the health record texts; determining the association evidence data between two connected nodes in the first node graph using the identification indication information of the health record texts according to preset association evidence processing rules; calculating the weight parameters of the edges between two connected nodes based on the association evidence data; and marking the corresponding weight parameters on the edges to obtain a second node graph; and determining multiple connected clusters based on the second node graph through clustering processing; wherein one connected cluster corresponds to one The system identifies multiple users and their health record text groups by merging the corresponding health record texts based on the multiple connected clusters. It receives and responds to health information query requests for objects of interest, identifies the objects of interest and associated objects, and queries the corresponding user's health record text group based on the consensus record identifiers corresponding to the objects of interest and associated objects, obtaining multiple health record texts. Based on the multiple health record texts, it extracts second-type fields related to the objects of interest and associated objects as target fields and auxiliary fields, respectively. The second-type fields are fields that contain at least the user's health information. Based on the target fields and auxiliary fields, it constructs corresponding target health feature sequences and auxiliary health feature sequences. It processes the target health feature sequences and auxiliary health feature sequences using a preset health status prediction model to obtain the health status prediction result of the objects of interest. The preset health status prediction model includes at least a parallel first health prediction branch, a second health prediction branch, and a joint regression analysis model. The first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0133] See Figure 8 As shown in the embodiments of this specification, an apparatus for processing cross-institutional health record data based on multi-dimensional features is also provided. This apparatus may specifically include the following structural modules: The first acquisition module 801 is specifically used to acquire multiple health record texts; and extract corresponding first type fields from the multiple health record texts; wherein, the first type fields are fields that contain at least identification indication information; The first construction module 802 can be specifically used to obtain the identification indication information of the health record text based on the first type of field; and to construct the identification blocking key of the health record text based on the identification indication information. The second construction module 803 can be specifically used to construct a first node graph based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text. The first determining module 804 can be used to determine the associated evidence data between two connected nodes in the first node diagram based on the preset associated evidence processing rules and the identification indication information of the health record text. The first processing module 805 can be used to calculate the weight parameters of the edge between two connected nodes based on the associated evidence data; and to mark the corresponding weight parameters on the edge to obtain the second node graph. The second processing module 806 is specifically used to determine multiple connected clusters based on the second node graph through clustering processing; wherein, one connected cluster corresponds to one user; The second determining module 807 can be specifically used to determine multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters. The second acquisition module 808 can be used to receive and respond to health information query requests for objects of interest, identify objects of interest and associated objects associated with objects of interest, and query the health record text group of the corresponding user according to the consensus record identifier corresponding to the objects of interest and associated objects to obtain multiple health record texts; The extraction module 809 can be specifically used to extract second-type fields about the object of interest and the associated object from the multiple health record texts, respectively, as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and construct corresponding target health feature sequences and auxiliary health feature sequences based on the target fields and auxiliary fields; The third processing module 8010 is specifically used to process the target health feature sequence and the auxiliary health feature sequence using a preset health status prediction model to obtain the health status prediction result of the object of interest; wherein, the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

[0134] In some embodiments, the identification information may specifically include: attribute information of the health record text and user attribute information, etc. The attribute information of the health record text may specifically include at least one of the following: the health record text's file number, the physical examination number associated with the health record text, the physical examination date associated with the health record text, the name of the institution that generated the health record text, and the text generation date of the health record text. The user's attribute information may specifically include at least one of the following: user's ID card number, user's name, user's date of birth, user's mobile phone number, user's age, name of the organization to which the user belongs, user's gender, etc.

[0135] In some embodiments, when the second construction module 803 is specifically implemented, a first node graph can be constructed according to the identification indication information and identification blocking key of the health record text in the following manner: Based on the identification indication information of the health record text, a node corresponding to the health record text is set in the initial node graph; and the identification indication information of the health record text is marked on the node; based on the identification blocking key of the health record text, two potentially related health record texts are determined; wherein, at least one identical identification blocking key exists between the two potentially related health record texts; and the two nodes corresponding to the two potentially related health record texts are connected by connecting edges to obtain the corresponding first node graph.

[0136] In some embodiments, when the first determining module 804 is specifically implemented, it can determine the associated evidence data between the connected first node and the second node in the following manner: using a preset feature matching model to process the identification indication information of the first node and the identification indication information of the second node, and calculating the information matching degree of the corresponding identification indication information of the first node and the second node; based on preset associated evidence processing rules, and according to the information matching degree of the corresponding identification indication information of the first node and the second node, determining and configuring the identification indication information of the first node and the second node that meet the requirements, so as to construct one or more associated evidences; combining the associated evidences to obtain the associated evidence data between the first node and the second node.

[0137] In some embodiments, when the first processing module 805 is specifically implemented, the weight parameters of the connection between the connected first node and the second node can be calculated in the following manner: the evidence type of the associated evidence is determined according to the preset associated evidence processing rules; the similarity score between the first node and the second node is determined according to multiple associated evidence and the evidence type of the associated evidence; and the weight parameters of the connection between the connected first node and the second node are calculated using the similarity score according to the preset associated evidence processing rules.

[0138] In some embodiments, the device may further be used to: detect whether key indication information in the identification indication information of the first node and the identification indication information of the second node is empty; wherein the key indication information includes: user ID number, user gender, and user date of birth; when the key indication information in the identification indication information of the first node and the identification indication information of the second node is not empty, detect whether the key identification indication information of the first node and the key identification indication information of the second node are the same; when the key identification indication information of the first node and the key identification indication information of the second node are different, determine the weight parameter of the connection between the first node and the second node as 0.

[0139] In some embodiments, when the second determining module 807 is specifically implemented, it can determine a group of health record texts for multiple users by merging the corresponding health record texts according to the multiple connected clusters in the following manner: determining multiple health record texts for multiple users according to the multiple connected clusters; determining a consensus file identifier for the multiple health record texts for each user according to a preset file merging rule; and merging and combining the multiple health record texts for each user using the consensus file identifier to obtain a group of health record texts for multiple users.

[0140] In some embodiments, when the second determining module 807 is specifically implemented, the consensus file identifier for multiple health record texts of each user can be determined in the following manner: detecting whether the file number identifier field of the multiple health record texts of the user is empty; when at least one of the file number identifier fields of the multiple health record texts of the user has a non-empty file identifier, extracting existing file identifiers based on the file number identifier field of the user's health record texts; detecting whether the number of types of existing file identifiers is greater than 1; when the number of types of existing file identifiers is greater than 1, determining the quantity and generating organization of each type of existing file identifier; determining existing file identifiers that meet the requirements as candidate file identifiers based on the quantity and generating organization of each type of existing file identifier; detecting whether the candidate file identifiers overlap with the consensus file identifiers of other users; when the candidate file identifiers overlap with the consensus file identifiers of other users, generating consensus file identifiers for multiple health record texts of that user using the candidate file identifiers according to preset identifier generation rules.

[0141] In some embodiments, after determining multiple user health record text groups by merging the corresponding health record texts according to the multiple connected clusters, the device may further be used to: extract a second type of field from the user's health record text group; wherein the second type of field is a field that contains at least the user's health information; determine and determine the reliability weight coefficient of the second type of field according to the institution level and record type of the health record text generating institution; determine the target object indicated by the second type of field by performing semantic recognition on the second type of field; detect whether the target objects indicated by the second type of field overlap; when it is determined that the target objects indicated by the second type of field overlap, filter out the second type of field with overlapping target objects as a first candidate field; filter out the second candidate field whose text generation date meets the requirements according to the text generation date of the health record text to which the first candidate field belongs; detect whether the number of second candidate fields is greater than 1; when the number of second candidate fields is greater than 1, according to the preset record data fusion rules, use the reliability weight coefficient to fuse multiple second candidate fields to generate a health information consensus field about the user's target object's health information; wherein the health information consensus field is used to generate the user's health data fusion summary report.

[0142] It should be noted that the units, devices, or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this specification, the functions of each module can be implemented in one or more software and / or hardware, or the module that implements the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or units shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0143] As can be seen from the above, the processing of health record data based on the embodiments of this specification can be well adapted to the health big data governance scenario. It can efficiently and accurately find the health record data corresponding to the same user from massive and chaotic health record data, so as to realize the unified governance and effective integration of multi-source health record data of different users across institutions. Furthermore, it can also effectively utilize the unified and integrated user health record text group through a preset health status prediction model to comprehensively and accurately analyze the health status of the subject of interest.

[0144] In a specific scenario example, the multi-dimensional feature-based cross-institutional health record data processing method provided in this manual can be applied to achieve unique identification and correction of health examination record numbers based on cross-institutional multi-feature fusion and clustering algorithms. The specific implementation process may include the following:

[0145] In this scenario example, considering the widespread use of information systems in physical examination centers and hospitals across the country, residents often undergo physical examinations at different institutions and in different years, resulting in a large number of scattered records. Due to the different file number coding rules and large differences in input habits among institutions, the following common problems have arisen: multiple files for one person: the same person has duplicate file numbers created by different hospitals; multiple people in one file: the same file number is shared by multiple examinees; missing or non-standard ID numbers; abnormal names or names containing decorative characters (such as "12 Zhang San", "Zhang San physical examination", "Li Xiao (Da)"); inconsistent gender, birthday, and mobile phone number; conflicting numbers across hospitals (such as "52970" for Zhang San at Hospital A, and "52970" for Li Si at Hospital B).

[0146] Traditional methods typically rely on single-field matching (ID number or mobile phone number) or simple fuzzy comparison, lacking cross-institutional identification and multi-feature fusion capabilities, and cannot achieve global consistency of "one person, one file number".

[0147] To address the aforementioned issues, this scenario example presents a method for uniquely identifying and correcting health checkup file numbers across institutions and using multi-feature fusion and clustering algorithms. By constructing a multi-layered blocking mechanism, weighted evidence matching, disjoint-set clustering, and cross-institutional conflict resolution, it achieves: one file per person (globally unique file number); one checkup per number (unique checkup number); cross-institutional data consistency and traceability; and ultimately, the establishment of a unified system for examinee identification across hospitals, years, and departments. For detailed implementation details, please refer to [reference needed]. Figure 9 As shown, it includes the following steps.

[0148] Step 1: Data reading and preprocessing.

[0149] In specific implementation, fields (e.g., the first type of fields) can be read from data files exported from multiple physical examination centers: archive number (id_patientarchive), physical examination number (id_patient), name, gender, ID card number, date of birth, age, mobile phone number, entity name, physical examination date, etc. (e.g., identification indication information).

[0150] And perform the following standardization processing (e.g., preprocessing): convert full-width characters to half-width characters, and remove redundant symbols; remove content in parentheses (e.g., "Zhang San (Da)" → "Zhang San"); clear prefixes and suffixes of numbers in names (e.g., "12Zhang San", "Zhang San12"); verify and repair ID card numbers, automatically upgrade 15-digit ID cards to 18-digit ones, perform check code verification based on the GB 11643 standard, and extract date of birth and gender information from ID cards; extract main numbers and last digits of mobile phone numbers; unify date formats and support parsing of multiple formats; special examination identification: if the name contains "mammography", "thyroid and breast", or "breast", mark it as a special examination and exclude it from clustering; intelligent conversion between traditional and simplified Chinese characters, establish and use a mapping table of traditional and simplified Chinese characters to uniformly convert text fields such as names and entity names; clean leading garbage characters, identify and remove non-semantic characters such as numbers and special symbols leading names; extract main telephone numbers, extract 11-digit main mobile phone numbers from complex formats through regular expressions. The purpose of the above processing is to ensure unified data structure, complete fields and comparable texts across different institutions.

[0151] Step 2: Perform hierarchical classification processing on records.

[0152] In specific implementation, the processing includes: bad name identification, identifying invalid names such as pure numbers, pure symbols, and alphanumeric mixtures, and performing clustering only through ID cards; whitelist identification, identifying special identifiers combining English letters and numbers (such as "A123"), and keeping the original archive number unchanged; normal records, performing a complete linking process on records with valid Chinese names.

[0153] Step 3: Calculate and use multi-feature blocking keys to construct Blocking (e.g., identify blocking keys), and perform multi-dimensional blocking key streaming clustering.

[0154] In specific implementation, in order to reduce computational complexity, the full database data is grouped according to feature combinations, and comparison is performed only within each group. Specifically, five types of block keys can be constructed. In specific operations, when the number of records in a single blocking bucket exceeds 800, the bucket is automatically subdivided by birth year or name prefix to prevent O(n²) explosion.

[0155] In practice, different weights can be set for different blocking keys. First, adaptive block segmentation is performed: when the size of the blocking block exceeds the threshold (e.g., 800 records), it is further subdivided according to dimensions such as birth year, last digit of mobile phone number, and hospital; then, union-query clustering is performed: pairwise comparisons are performed within the blocking block, and records with a union weight score ≥ 60 are merged using a union-query set (to obtain the first node graph).

[0156] Step 4: Cross-cluster bridging mechanism.

[0157] In practice, signature extraction can be performed first: extract signature features for each cluster, including sets of ID cards, dates of birth, names, and phone numbers; then, a security check can be performed: determine whether two clusters can be safely merged (ID cards do not conflict, dates of birth do not conflict); then, strong evidence bridging can be performed: for example, when the similarity between the same ID card and name is ≥95%, bridging is determined, or when the same file number and strong evidence are obtained, bridging is determined; finally, signature merging can be performed: after bridging, the cluster signature is updated, and new bridging opportunities are recursively checked.

[0158] Step 5: Intelligent allocation of unified file numbers.

[0159] In practice, the following principles can be followed to allocate unified file numbers: frequency priority principle, select the original file number that appears most frequently for each cluster as the unified file number; time priority principle, when multiple file numbers appear the same number of times, select the file number corresponding to the earliest physical examination date; new number generation rule, for clusters without original file numbers, give priority to using the physical examination number, and generate an 8-digit random number when it is not possible to use it.

[0160] Step 6: Conflict detection of multidimensional evidence strength score.

[0161] In practical implementation, a conflict score calculation model can be designed: calculate the conflict score for each pair of records within a cluster, with the following scoring rules: For hard conflicts (immediately determined), clearly different ID cards (not empty and unequal): +100 points. For moderate conflicts, clearly different genders: +60 points; clearly different birth dates: +60 points; name differences: name similarity <70%: +40 points; name similarity <85%: +20 points. For secondary conflicts, unit name similarity <70%: +20 points; different mobile phone number segments (first 3 digits): +15 points; different names with no common evidence: +25 points. The conflict determination threshold is: when the cumulative score is ≥80 points, it is determined to be a truly different person, and the file number needs to be split. In addition, there is a default conservative principle: when the score is <80, it is assumed that it may be the same person (data entry difference), and the file number remains unchanged.

[0162] Step 7: Conflict resolution based on disjoint set.

[0163] In specific implementation, the conflict relationship graph is constructed (e.g., the second node graph): records within a cluster are used as nodes, and conflict edges are established between record pairs with a conflict score ≥ 80; non-conflict connectivity is achieved: all non-conflict record pairs (score < 80) are connected using a disjoint-set data structure; group splitting is performed: each connected component of the disjoint-set data structure is treated as an independent subgroup; file number reallocation is performed: the largest subgroup retains the original file number, or other subgroups are assigned new file numbers, or the physical examination number is used first, otherwise a random number is generated.

[0164] Step 8: Results Export and Quality Check.

[0165] In practice, multiple export formats can be supported: generating standard CSV and secure Excel CSV (numeric fields enclosed in single quotes). Quality metrics monitoring includes: the number of ID cards corresponding to one file number (should be 0); file number retention rate (assessing system stability); and the number of conflict resolutions (assessing detection accuracy).

[0166] In this scenario example, a multi-dimensional evidence accumulation scoring mechanism can also be introduced. Unlike the binary judgment (yes / no) of existing technologies, a quantitative evidence strength accumulation model is proposed for the first time: hierarchical weight design: hard conflict (ID card) 100 points, medium conflict (gender / birth) 60 points, soft conflict (name / company) 20-40 points; multiple evidence superposition: multiple independent evidence combinations must reach the threshold to determine conflict, avoiding misjudgment due to a single field error; dynamic threshold: the conflict threshold can be adjusted according to business needs (80-100 points recommended) to balance accuracy and stability.

[0167] In this scenario example, a default conservative conflict detection strategy can also be introduced. Existing technologies tend to "split if there is a difference", while this solution adopts the principle of "default correct, strong evidence to overturn": score < 80: defaults to the same person (may be due to data entry differences), keep the file number; score ≥ 80: judged to be different people, split the file number; special protection: when the ID card is the same (not empty), never split, the highest priority.

[0168] In this scenario example, a hierarchical record processing architecture can be introduced. To address the complexity of medical data, a three-layer processing strategy is designed: a bad name layer (pure numbers / symbols): clustering only ID card numbers to avoid invalid comparisons; a whitelist layer (special encoding): preserving the original file number and not participating in clustering; and a normal record layer: completing the entire connection process for precise matching. Based on this hierarchical processing, system efficiency is improved by over 300%.

[0169] In this scenario example, an adaptive multi-dimensional blocking algorithm may be specifically introduced. Traditional blocking methods use fixed key values, while this solution adopts multi-dimensional sequence blocking + adaptive subdivision: Sequence blocking: applying 4 types of blocking keys in sequence to gradually cover different matching scenarios; Adaptive subdivision: when a blocking block is too large (>800), it is automatically subdivided by secondary dimensions (birth year, mobile phone tail number); Hard upper limit protection: forcibly subdividing extra-large blocks (>5000) to prevent memory overflow; the time complexity is reduced from O(n²) to O(n k), where k is the average size of the blocking block (usually <100).

[0170] In this scenario example, a cluster-level unified archive number allocation mechanism may be specifically introduced. Existing methods allocate archive numbers record by record, and the present invention proposes a cluster-level unified allocation strategy for the first time: Frequency priority: selecting the archive number that appears most frequently in the cluster as the representative; Time backtracking: selecting the earliest used archive number when the frequencies are the same; Global consistency: ensuring that all records in the same cluster use the same archive number. Based on this mechanism, the archive number retention rate is increased from 60% to more than 85%.

[0171] In this scenario example, a dedicated cleaning pipeline for Chinese medical texts may be specifically introduced. A dedicated cleaning process is designed targeting the characteristics of Chinese medical data: Traditional and simplified character mapping table: covering commonly used Chinese characters and surnames (88 traditional Chinese characters); Leading garbage cleaning: using regular expressions to identify and remove the "number+" prefix in entries such as "\d+、Zhang San"; Special recognition for breast-related items: specially processing names of special physical examination items such as "breast color ultrasound"; Scientific notation filtering: identifying and removing abnormal formats such as "1.23E+17". Based on this cleaning pipeline, the name matching accuracy after cleaning is increased by 15%.

[0172] In this scenario example, a multi-feature weighted matching model (Evidence Weighted Matching) (for example, a preset feature matching model) may be specifically introduced and used.

[0173] In specific implementation, a comprehensive similarity score, evidence (for example, associated evidence) and corresponding weight (for example, the weight parameter of the edge) may be calculated between two records in each blocking bucket.

[0174] In specific implementation, when the cumulative score is ≥60, the two records are determined to belong to the same examinee. Among them, Age-Date Coherence: if the age difference between two physical examinations is close to the difference between the physical examination dates (the error does not exceed 1.2 years), it indicates that the possibility that the two records belong to long-term reexaminations of the same person is extremely high, which is used as strong auxiliary evidence when no ID card is available.

[0175] In this scenario example, a Negative Guard mechanism may be specifically introduced and executed.

[0176] In practice, if the ID card number, gender, or date of birth fields in two records are all non-empty and inconsistent, merging will be blocked directly. This mechanism prevents highly similar names from being mistakenly identified as the same person.

[0177] In this scenario example, a clustering algorithm based on disjoint-set data structure (Union-FindClustering) can also be introduced for clustering.

[0178] In practice, each record can be treated as a node, and edges can be connected based on the weighted relationship of evidence. By merging all connected components using a disjoint-set data structure algorithm, several "clusters" are obtained, each representing all the physical examination records of the same examinee.

[0179] Among them, the Union-Find is an efficient data structure that supports quickly determining whether two elements belong to the same set, and is used for dynamic merging and searching.

[0180] In this scenario example, cross-organizational consistency governance can also be introduced.

[0181] In practice, cross-agency governance can be divided into two scenarios: Scenario 1. Same file number for different people (conflict splitting): When the same file number appears in different hospitals but belongs to different clusters, the system retains the original number of the main hospital according to the "primary hospital priority" principle, and the file numbers of the other hospitals are automatically derived as cid_xxxxx.

[0182] In practice, if the primary hospital is prioritized (e.g., West China Hospital > Shangjin Hospital > others), and if there is a tie, the primary cluster is selected based on the majority vote of the number of records.

[0183] Scenario 2. Different ID numbers for the same person (merging of strong evidence): If cross-court records meet one of the following strong evidence criteria, they shall be merged.

[0184] Strong evidence 1: The ID numbers match; Strong evidence 2: Names are ≥95% similar and birthdays are the same (phone numbers are the same or age-dates are consistent). After merging, the original file number from the hospital with the highest degree of support or the primary hospital will be used.

[0185] Among them, the aforementioned Cross-Institution Conflict Resolution: when the same file number is shared by different institutions and different people, it is automatically split to eliminate global conflicts.

[0186] Strong Evidence Merge: Merging across institutions is only performed when there is conclusive evidence of identity consistency between the institutions, ensuring accuracy.

[0187] In this scenario example, we can also introduce the uniqueness of the file number and physical examination number.

[0188] In practice, a unique archive number (archive_corrected) and a patient number (patient_corrected) can be generated for each cluster.

[0189] The specific generation rules are as follows: 1) If there is a valid and unique file number in the cluster, the original number is retained; 2) If there is a conflict with the same number, the primary cluster retains the number, and the other clusters derive a new number (such as cid_10000001); 3) If there is no file number in the cluster but there is a physical examination number, the first physical examination number is taken; 4) If there is no available number, the system automatically generates a new number.

[0190] Among them, the Derived ID is a new unique number generated from the conflict number, with the prefix cid_ or pid_, while retaining the original number mapping relationship.

[0191] In this scenario example, consistent closing and traceable output can also be introduced.

[0192] In practice, the following three types of result files can be output and saved: id_corrected_full.csv: the master data table, containing the corrected file number and medical examination number; merge_evidence.csv: the merge evidence table, recording the evidence and weight of each merge relationship; and conflict_audit.csv: the conflict audit table, recording the reasons for splitting conflicts with the same number and the logic for selecting the primary number. This ensures that each merge or split is supported by evidence, allowing for manual review and model optimization.

[0193] Through the above scenario examples, the processing method for cross-institutional health record data based on multi-dimensional features provided in this manual has the following advantages: 1) Cross-institutional consistency identification: Supports unified governance of multi-source physical examination data across hospitals and years; 2) Strong multi-dimensional fusion and robustness: Simultaneously utilizes multi-dimensional features such as ID card, name, gender, birthday, mobile phone number, unit, and time pattern, and can correctly identify even missing fields; 3) High efficiency and scalability: Adopting blocking algorithms and disjoint-set data structures, millions of data points can be processed within hours; 4) Traceable and interpretable results: Outputs evidence tables and conflict tables to achieve transparent algorithm management; 5) Strong engineering feasibility: A Python prototype has been implemented, supporting GPU acceleration, and can be directly embedded into hospital physical examination information systems.

[0194] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0195] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0196] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer-readable storage media, including storage devices.

[0197] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0198] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. This specification can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0199] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations and modifications are possible without departing from the spirit of this specification, and it is intended that the appended claims cover such variations and modifications without departing from the spirit of this specification.

Claims

1. A method for processing cross-institutional health record data based on multidimensional features, characterized in that, include: Retrieve multiple health record texts; And extract the corresponding first type of fields from the multiple health record texts; wherein, the first type of fields are fields that contain at least identification and indication information; Based on the first type of field, obtain the identification indication information of the health record text; and construct the identification blocking key of the health record text based on the identification indication information; A first node graph is constructed based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text. Based on the preset rules for processing related evidence, the identification information of the health record text is used to determine the related evidence data between two connected nodes in the first node diagram; Based on the associated evidence data, the weight parameters of the edges between two connected nodes are calculated; and the corresponding weight parameters are marked on the edges to obtain the second node graph. Based on the second node graph, multiple connected clusters are identified through clustering; each connected cluster corresponds to one user. Based on the multiple connected clusters, the health record text groups of multiple users are determined by merging the corresponding health record texts. Receive and respond to health information query requests for targets of interest, identify targets of interest and associated objects associated with targets of interest; and query the health record text groups of the corresponding users based on the consensus record identifiers corresponding to targets of interest and associated objects to obtain multiple health record texts; Based on the multiple health record texts, second-type fields related to the objects of interest and associated objects are extracted respectively, serving as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and based on the target fields and auxiliary fields, corresponding target health feature sequences and auxiliary health feature sequences are constructed. A pre-defined health status prediction model is used to process the target health feature sequence and the auxiliary health feature sequence to obtain the health status prediction result of the object of interest. The pre-defined health status prediction model includes at least a first health prediction branch and a second health prediction branch connected in parallel, as well as a joint regression analysis model. The first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

2. The method according to claim 1, characterized in that, The identification information includes: attribute information of the health record text and attribute information of the user; The attribute information of the health record text includes at least one of the following: the health record text's file number, the physical examination number associated with the health record text, the physical examination date associated with the health record text, the name of the institution that generated the health record text, and the text generation date of the health record text. The user's attribute information includes at least one of the following: user ID number, user name, user date of birth, user mobile phone number, user age, name of the organization to which the user belongs, and user gender.

3. The method according to claim 1, characterized in that, The first node graph is constructed based on the identifier indication information and identifier blocking key of the health record text, including: Based on the identification information of the health record text, set up a node corresponding to the health record text in the initial node diagram; and mark the identification information of the health record text on the node. Based on the identifier blocking key of the health record text, two potentially related health record texts are identified; wherein, at least one identifier blocking key is shared between the two potentially related health record texts. By connecting the two nodes corresponding to two potentially related health record texts with edges, the corresponding first node graph can be obtained.

4. The method according to claim 1, characterized in that, The step of determining the association evidence data between two connected nodes in the first node graph based on preset processing rules and using the identifier information of the health record text includes: The evidence data linking the first and second connected nodes is determined in the following manner: The identification information of the first node and the identification information of the second node are processed using a preset feature matching model, and the information matching degree of the corresponding identification information of the first node and the second node is calculated. Based on the preset rules for processing related evidence, and according to the information matching degree of the corresponding identifier indication information of the first node and the second node, the identifier indication information of the first node and the identifier indication information of the second node that meet the requirements are determined and configured to construct one or more pieces of related evidence. By combining the aforementioned related evidence, we obtain the related evidence data between the first node and the second node.

5. The method according to claim 4, characterized in that, The step of calculating the weight parameters of the edge between two connected nodes based on the associated evidence data includes: The weight parameters of the edge connecting the first and second nodes are calculated as follows: Based on the pre-defined rules for handling related evidence, determine the type of evidence for related evidence; Based on multiple pieces of related evidence and the types of evidence associated with them, a similarity score is determined between the first node and the second node. Based on the preset rules for processing related evidence, the weight parameters of the connection between the first and second connected nodes are calculated using the similarity score.

6. The method according to claim 5, characterized in that, The method further includes: The system checks whether the key indication information in the identification indication information of the first node and the identification indication information of the second node is empty; wherein, the key indication information includes: user ID number, user gender, and user date of birth; When the key indication information in the first node's identification indication information and the second node's identification indication information is not empty, check whether the key identification indication information of the first node and the key identification indication information of the second node are the same; When the key identifier information of the first node is different from that of the second node, the weight parameter of the edge connecting the first node and the second node is set to 0.

7. The method according to claim 1, characterized in that, The step of determining multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters includes: Based on the multiple connected clusters, multiple health profile texts of multiple users are determined; Based on the preset file merging rules, a consensus file identifier is determined for multiple health record documents for each user; Using consensus archive identifiers, multiple health record texts from various users are merged and combined to obtain a health record text group for multiple users.

8. The method according to claim 7, characterized in that, The process of determining consensus file identifiers for multiple health record files of each user based on preset file merging rules includes: Check if the file number identifier field of multiple health record texts of a user is empty; When at least one of the health record texts has a non-empty file identifier in its file number identifier field, the existing file identifier is extracted based on the file number identifier field of the user's health record text. Check if the number of existing file identification types is greater than 1; When the number of existing archival identifiers is greater than 1, determine the number of existing archival identifiers and the generating institution for each type; Based on the quantity and generating organization of existing archival identifiers of each type, existing archival identifiers that meet the requirements are identified as candidate archival identifiers. Detect whether the candidate file identifier overlaps with the consensus file identifier of other users; When a candidate profile identifier overlaps with the consensus profile identifier of other users, a consensus profile identifier for multiple health profiles of that user is generated using the candidate profile identifier according to the preset identifier generation rules.

9. The method according to claim 1, characterized in that, After determining multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters, the method further includes: Based on the user's health record text group, the second type of fields of the health record text are extracted; wherein, the second type of fields are fields that contain at least the user's health information; Determine the reliability weight coefficient of the second type of field based on the institution level and record type of the institution that generated the health record text; By performing semantic recognition on the second type of field, the target object indicated by the second type of field is determined; Detect whether there is overlap between the target objects indicated by the second type of field; When it is determined that the target objects indicated by the second type of field overlap, the second type of field in which the indicated target objects overlap is selected as the first candidate field; Based on the text generation date of the health record text to which the first candidate field belongs, filter out the second candidate fields whose text generation dates meet the requirements; Check if the number of second candidate fields is greater than 1; When the number of second candidate fields is greater than 1, multiple second candidate fields are merged according to the preset archive data fusion rules and the reliability weight coefficient to generate a health information consensus field about the health information of the user's target object; wherein, the health information consensus field is used to generate a health data fusion summary report for the user.

10. A processing device for cross-institutional health record data based on multi-dimensional features, characterized in that, include: The first acquisition module is used to acquire multiple health record texts; And extract the corresponding first type of fields from the multiple health record texts; wherein, the first type of fields are fields that contain at least identification and indication information; The first construction module is used to obtain the identification indication information of the health record text based on the first type of field; and to construct the identification blocking key of the health record text based on the identification indication information. The second construction module is used to construct a first node graph based on the identification indication information and identification blocking key of the health record text; wherein, the first node graph contains multiple nodes, one node corresponds to one health record text, and the node is marked with the identification indication information of the corresponding health record text. The first determination module is used to determine the associated evidence data between two connected nodes in the first node diagram based on the preset associated evidence processing rules and the identification information of the health record text. The first processing module is used to calculate the weight parameters of the edge between two connected nodes based on the associated evidence data; and to mark the corresponding weight parameters on the edge to obtain the second node graph. The second processing module is used to determine multiple connected clusters based on the second node graph through clustering processing; wherein, one connected cluster corresponds to one user; The second determining module is used to determine multiple user health record text groups by merging the corresponding health record texts based on the multiple connected clusters. The second acquisition module is used to receive and respond to health information query requests for objects of interest, identify objects of interest and associated objects associated with objects of interest, and query the health record text groups of the corresponding users according to the consensus record identifiers corresponding to the objects of interest and associated objects to obtain multiple health record texts; The extraction module is used to extract second-type fields about the object of interest and the associated object from the multiple health record texts, respectively, as target fields and auxiliary fields; the second-type fields are fields that contain at least the user's health information; and construct corresponding target health feature sequences and auxiliary health feature sequences based on the target fields and auxiliary fields; The third processing module is used to process the target health feature sequence and the auxiliary health feature sequence using a preset health status prediction model to obtain the health status prediction result of the object of interest; wherein, the preset health status prediction model includes at least a first health prediction branch and a second health prediction branch in parallel, as well as a joint regression analysis model; the first health prediction branch includes at least a first feature processing module based on an LSTM structure and a first hidden information processing module based on a dual GRU structure.

Images