Patient primary index management method and device

Abstract

The invention provides a patient primary index management method and device, and relates to the technical field of computers. The method comprises the following steps: acquiring field data of target fields in a plurality of candidate records; determining a matching decision factor of the record pair, and identifying the record pair as a suspected matching pair when the matching decision factor falls in the suspected matching interval; constructing at least one suspected cluster graph model based on the matching decision factor of the suspected matching pair; iteratively optimizing the edge selection in each suspected cluster graph model until the consistency in the suspected cluster graph model is maximized and the conflict edge selection is minimized, and dividing the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be audited based on the merging tendency strength of the current edge; performing depth feature analysis on the second suspected cluster by using a depth discrimination model to obtain an auxiliary decision label; and updating the patient main index database according to an audit confirmation result from the user terminal. The processing efficiency of patient primary index management is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a patient master index management method and device. BACKGROUND

[0002] In the existing medical information system, in order to realize the consistency and intercommunication of patient data between multiple information systems, the "patient master index" (EMPI) technology is generally used to establish a unified patient identity. Usually, the identity card number, medical insurance card number or patient number is used as the primary key to realize the association of patient data between different systems.

[0003] In the management of patient master index, suspected matching refers to the case that two or more patient records have certain field similarity or correlation, but cannot be directly determined as the same patient. At present, the judgment of suspected matching often depends on a single similarity and a fixed threshold, which is easy to cause false merging or missing merging in cross-system data merging. The processing of suspected matching completely depends on manual processing, and lacks systematic management and visual auditing mechanism. Therefore, the processing efficiency of patient master index management is low. SUMMARY

[0004] To solve the above technical problems, the present application provides a patient master index management method and device to solve the problem of low processing efficiency of patient master index management.

[0005] In a first aspect, the present application provides a patient master index management method, used in a processing device, comprising: obtaining field data of a target field in a plurality of candidate records; determining a matching decision factor of a record pair based on the field data of the target field, and identifying the corresponding record pair as a suspected matching pair when the matching decision factor falls within a pre-calculated suspected matching interval, the record pair being composed of any two candidate records; constructing at least one suspected cluster graph model based on the matching decision factor of the suspected matching pair, wherein the nodes of the suspected cluster graph model represent one candidate record in the suspected matching pair, and the edge weight of the suspected cluster graph model represents the matching decision factor of the corresponding suspected matching pair; iteratively optimizing the edge selection in each suspected cluster graph model until the consistency in the suspected cluster graph model is maximized and the conflicting edge selection is minimized, dividing the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be audited based on the current merging tendency strength of the edge; performing deep feature analysis on the second suspected cluster using a preset deep discriminant model to obtain an auxiliary decision label; updating a patient master index according to an audit result from a user terminal, wherein the audit result is generated by the user terminal based on an audit on the candidate record corresponding to the second suspected cluster according to the auxiliary decision label.

[0006] In a possible implementation, the obtaining of the field data of the target field in the plurality of candidate records comprises: obtaining a plurality of patient records and record features of the patient records; when it is detected that a value of any target field in the patient record is missing, obtaining a set of context fields of a field to be completed in the candidate record, an observed value of each context field, a quality confidence of each context field, a temperature coefficient, and a candidate value of the field to be completed, wherein the temperature coefficient is used to control sharpness of a weight distribution of a context field weight; determining a corresponding context field weight based on the quality confidence of the context field and the temperature coefficient; constructing an expected representation of the candidate value in the context field based on a structural relationship between the context field and the candidate value of the field to be completed; performing a squared difference calculation on the expected representation of the candidate value in the context field and the observed value of the context field to obtain a difference measure of the context field; iterating through the set of context fields, and constructing a risk function about the candidate value based on the context field weight, the difference measure of the context field, a preset prior penalty of the context field, and a preset regularization strength; determining a candidate value that minimizes the risk function as a completion value of the field to be completed; filtering, based on the patient record, the patient record containing the completion value of the field to be completed to obtain a plurality of candidate records, wherein the record features include a hash signature, a numerical difference, and a character distance.

[0007] In a possible implementation, the field data comprises a data quality indicator and a field importance prior; and the determining of a matching decision factor of a record pair based on the field data of the target field comprises: determining a field weight of the target field based on the data quality indicator of the target field, the field importance prior, and a preset temperature parameter, wherein the temperature parameter is used to indicate sensitivity of the data quality indicator to the field weight; obtaining a field similarity of the target field and a variance estimate value of the field similarity, wherein the field similarity is used to indicate a matching degree of the record pair on the same target field; determine a matching score of the record pair based on the field similarity of the target field and a corresponding field weight; obtain an uncertainty based on the field weight of the target field and a variance estimation value of the field similarity; obtain a risk adjustment score as the matching decision factor based on the matching score, the uncertainty, and a preset adjustable risk coefficient.

[0008] In a possible implementation, the matching decision factor of the record pair is determined based on the field data of the target field, and when the matching decision factor falls within a pre-calculated suspected matching interval, the corresponding record pair is identified as a suspected matching pair, and the method further includes: obtain a mean value and a standard deviation of the matching score in a specified period of time; determine an upper limit value of the suspected matching interval based on the mean value, the standard deviation, and a preset first adjustment coefficient; determine a lower limit value of the suspected matching interval based on the mean value, the standard deviation, and a preset second adjustment coefficient; when the matching decision factor is greater than the upper limit value of the suspected matching interval, identify the corresponding record pair as a matching pair and generate a corresponding confidence score; when the matching decision factor is less than the lower limit value of the suspected matching interval, identify the corresponding record pair as a non-matching pair.

[0009] In a possible implementation, the suspected matching interval is updated by the following steps, including: within a preset sliding window length, determine a mean value of the matching score at the current time based on the matching score at the current time, a mean value of the matching score at the previous time, and a preset smoothing coefficient; update the upper limit value and the lower limit value of the suspected matching interval using the mean value of the matching score at the current time, and update the data quality indicator at the current time.

[0010] In a possible implementation, the edge selection in each suspected cluster graph model is iteratively optimized until the consistency in the suspected cluster graph model is maximized and the conflicting edge selection is minimized, and based on the current merging tendency strength of the edge, the suspected cluster graph model is divided into a first suspected cluster to be merged and a second suspected cluster to be audited, including: construct a corresponding energy function based on each suspected cluster graph model, wherein the energy function includes a consistency term and a conflict penalty term, the consistency term is composed of the matching decision factor of the selected edge and the merging tendency strength, and the conflict penalty term is composed of a preset conflict penalty coefficient and the merging tendency strength of two conflicting edges selected at the same time; iteratively adjust the merging tendency strength of edges of the suspected cluster graph model until the consistency term is maximized and the conflict penalty term is minimized, to obtain a current merging tendency strength of edges; screen a suspected cluster graph model satisfying a preset merging condition as a first suspected cluster to be merged, generate a corresponding merging confidence, and divide the remaining suspected cluster graph models into second suspected clusters to be audited, wherein the merging condition comprises that a number of target edges in the suspected cluster graph model is greater than a preset number threshold, and a proportion of target conflict edge pairs in the suspected cluster graph model is lower than a preset proportion threshold, the target edges are used to describe edges with a merging tendency strength greater than a preset merging threshold, and the target conflict edge pairs are used to indicate conflict edge pairs that simultaneously satisfy the merging tendency strength greater than the merging threshold.

[0011] In a possible implementation, the deep discriminant model comprises a multi-layer perception classifier and a message passing model in cascade; and the deep feature analysis on the second suspected cluster by using the preset deep discriminant model to obtain an auxiliary decision label comprises: converting candidate records of the second suspected cluster into feature vectors; calculating a feature difference vector based on the feature vectors of any two candidate records in the second suspected cluster; performing feature processing on the feature difference vector by using the multi-layer perception classifier to obtain a merging probability of a record pair; constructing a target graph model based on the second suspected cluster, wherein a target node of the target graph model represents a candidate record in the second suspected cluster, and a target edge of the target graph model represents a merging probability between two connected candidate records; inputting the target graph model into the message passing model, iteratively updating node features of the target nodes by fusing node-related data of neighbor nodes related to any target node in the target graph model, and outputting the node features of the target nodes in the target graph model; obtaining a discriminant evaluation value of the second suspected cluster based on the merging probability of each target edge in the target graph model and the node features of each target node; determining corresponding confidence information based on the merging probability of the record pairs in the second suspected cluster; comparing the discriminant evaluation value with a preset evaluation threshold, and determining a corresponding auxiliary decision label based on a comparison result and the confidence information.

[0012] In a possible implementation, the patient master index library is updated according to an audit confirmation result from a user terminal. For any second suspected cluster, based on a pre-acquired key field coverage, a conflict degree, a preset weight coefficient and a corresponding matching decision factor, a priority score of the second suspected cluster is determined, and a corresponding conflict marker is generated, wherein the key field coverage is used to indicate a complete proportion of candidate records in the second suspected cluster on a specified key field, the conflict degree is used to indicate a conflict degree of records in the second suspected cluster on a specified core field, and the conflict marker is used to indicate an identification of a conflict of records in the second suspected cluster on the key field; The second suspected clusters are processed in turn according to the priority scores to obtain audit confirmation results of the second suspected clusters.

[0013] In a possible implementation, the patient master index library is updated according to the audit confirmation results from the user terminal: Based on a merging confidence threshold of a previous time step, a preset smoothing coefficient and an average confidence, a merging confidence threshold of a current time step is updated, wherein the average confidence is composed of a first average confidence and a second average confidence, the first average confidence is determined by confidence information corresponding to a second suspected cluster whose audit confirmation result is confirmation merging, and the second average confidence is determined by a confidence score and / or a merging confidence corresponding to a record pair that passes the audit in the review queue; When the confidence score of the matching pair is greater than the merging confidence threshold, and the merging confidence of the first suspected cluster is greater than the merging confidence threshold, the matching pair and the first suspected cluster are merged into the patient master index library; The matching pair whose confidence score is less than the merging confidence threshold and the first suspected cluster whose merging confidence is less than the merging confidence threshold are added to the review queue, so that in the review queue, corresponding audit requests are sequentially sent to the user terminal.

[0014] In a second aspect, the present application provides a patient master index management device, comprising: A data acquisition module is configured to acquire field data of a target field in a plurality of candidate records; A matching judgment module is configured to determine a matching decision factor of a record pair based on the field data of the target field, and identify a corresponding record pair as a suspected matching pair when the matching decision factor falls within a pre-calculated suspected matching interval, wherein the record pair is composed of any two candidate records; A graph model construction module is configured to construct at least one suspected cluster graph model based on the matching decision factor of the suspected matching pair, wherein a node of the suspected cluster graph model represents one candidate record in the suspected matching pair, and an edge weight of the suspected cluster graph model represents the matching decision factor of the corresponding suspected matching pair. an optimization module configured to iteratively optimize edge selection in each suspected cluster graph model until consistency within the suspected cluster graph model is maximized and conflicting edge selection is minimized, divide the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be audited based on current edge merging tendency strength; an auxiliary decision-making module configured to perform deep feature analysis on the second suspected cluster by using a preset deep discriminant model to obtain an auxiliary decision-making label; an updating module configured to update a patient master index library according to an audit confirmation result from a user terminal, wherein the audit confirmation result is generated by the user terminal based on the auxiliary decision-making label to audit a candidate record corresponding to the second suspected cluster.

[0015] The patient master index management method and device provided in the application can improve the accuracy of the shunt processing, greatly improve the audit efficiency, and reduce the probability of erroneous or missed merging of data, and then perform deep feature analysis on the second suspected cluster to be audited by using a deep discriminant model to obtain an auxiliary decision-making label to assist the audit, thereby updating the patient master index library according to the audit confirmation result, reducing the amount of manual audit, and reducing the audit error rate. Therefore, the problem of low processing efficiency of patient master index management is solved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a patient master index management method provided in an embodiment of the application; Figure 2 A structural diagram of a patient master index management device provided in an embodiment of the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0018] Figure 1A flowchart of a patient master index management method provided in an embodiment of the present application is shown in FIG. 1. The method is used in a processing device and includes steps S101-S106.

[0019] S101, obtain field data of a target field in a plurality of candidate records.

[0020] In the present application, the processing device is a computing processing device for performing calculation, processing, storage and transmission of data, such as a central processing unit. The candidate records can be patient records from multiple sources (such as different hospitals, departments, and business systems), and the target field can be an ID number, name, gender, date of birth, contact number, address, diagnosis description, medical insurance card number, etc.

[0021] S102, determine a matching decision factor of a record pair based on the field data of the target field, and identify the corresponding record pair as a suspected matching pair when the matching decision factor falls within a pre-calculated suspected matching interval, the record pair being composed of any two of the candidate records.

[0022] In the present application, the matching decision factor is used to judge the relationship classification of the record pair. Specifically, the relationship between the record pair (i.e., two candidate records) can be one of a matching pair, a suspected matching pair, and a non-matching pair. Optionally, for a matching pair, it is considered to have a very high matching degree and a very low uncertainty, and a direct merging operation can be used, for example, the records of the same patient in different departments (such as outpatient and inpatient records), the key fields (ID number, name) are completely consistent, and there are no conflicting fields. For a non-matching pair, it is considered to have a very low matching degree and a relatively high uncertainty, and a direct exclusion operation can be used, for example, records of different patients (such as different names and unrelated ID numbers), or there are obvious conflicting fields (such as inconsistent gender and date of birth).

[0023] However, for a suspected matching pair, it is considered to have uncertainty in the matching degree and needs to be further analyzed, usually completely relying on manual review, for example, a patient with the same name (such as “Zhang San” and “Zhang San” but different ID numbers), and a large difference in cross-hospital data quality (such as missing fields or inconsistent formats). In view of this, the present application classifies the suspected matching pairs into a first suspected cluster to be merged and a second suspected cluster to be reviewed through shunt processing, the first suspected cluster does not need manual review, and the processing efficiency is improved. In addition, the present application performs feature analysis on the second suspected cluster to be reviewed through a deep discriminant model to generate an auxiliary decision label to assist manual review, and the review efficiency is improved.

[0024] S103, construct at least one suspected cluster graph model based on the matching decision factor of the suspected matching pair, wherein the node of the suspected cluster graph model represents one of the candidate records in the suspected matching pair, and the edge weight of the suspected cluster graph model represents the matching decision factor of the corresponding suspected matching pair.

[0025] In the present application, the candidate records in the suspected matching pairs are taken as nodes, and the matching decision factors of the suspected matching pairs are taken as edge weights to construct a suspected cluster graph model. For example, if node A and node B are a suspected matching pair, and node B and node C are a suspected matching pair, then node A, node B and node C form a connected suspected cluster graph model.

[0026] It should be noted that the suspected cluster graph model (or suspected cluster) of the present application can be one or more. Since the patient records can come from different hospitals, departments and time periods (such as outpatient, inpatient and emergency), the records from different sources can form multiple groups of unrelated suspected matching pairs due to incomplete information, format differences or input errors, thereby forming multiple independent suspected clusters (and corresponding suspected cluster graph models). In addition, the suspected matching pairs can naturally form multiple independent suspected clusters due to the matching decision factors being in the middle interval and the lack of cross correlation between different record pairs.

[0027] S104, iteratively optimize the edge selection in each suspected cluster graph model until the consistency in the suspected cluster graph model is maximized and the conflicting edge selection is minimized, divide the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be audited based on the current merging tendency strength of the edge.

[0028] It should be noted that the suspected cluster is a collection of suspected matching pairs. The consistency in the suspected cluster graph model (and the suspected cluster) means that the record pairs in the same suspected cluster exhibit high similarity in the key fields, which can ensure that the record pairs in the cluster belong to the same patient and reduce false positives caused by field fluctuations (such as input errors) or semantic differences (such as “diabetes” and “DM”). The conflicting edge selection means that two or more conflicting edges (such as edges that share the same node but point to different clusters, or edges whose record pairs conflict in key fields) are selected at the same time in the suspected cluster graph model, resulting in conflicting merging decisions.

[0029] Therefore, the merging tendency strength of the edges in the suspected cluster graph model is iteratively adjusted to maximize the intra-cluster consistency and minimize the conflicting edge selection. At this time, based on the value of the merging tendency strength (merging tendency strength) and the cluster structure characteristics, two types of clusters are divided, the high-confidence clusters (first suspected cluster) are automatically merged, and the low-confidence clusters (second suspected cluster) are subjected to deep analysis, balancing efficiency and accuracy, reducing the amount of review and improving processing efficiency.

[0030] S105, performing deep feature analysis on the second suspected cluster using a pre-set deep discriminant model to obtain an auxiliary decision label.

[0031] S106, updating the patient master index database according to the audit confirmation result from the user terminal, wherein the audit confirmation result is generated by the user terminal based on the auxiliary decision label after auditing the candidate record corresponding to the second suspected cluster.

[0032] In the present application, for the second suspected cluster requiring manual audit, the deep discriminant model (such as neural convolution network) is used to extract the deep features (such as name homophony, address transition rule, contact method association, etc.) of the record, and output the auxiliary decision label. The deep discriminant model can capture complex associations (such as the historical address change track of the same patient), improve the analysis depth, and improve the analysis accuracy. Thus, by clearly marking the key information (such as the conflict mark and the matching basis) through the label, intelligent support is provided for manual audit in the user terminal, the audit difficulty is reduced, and the audit efficiency is greatly improved.

[0033] Further, the user terminal audits the second suspected cluster based on the auxiliary decision label, and then feeds back the audit confirmation result of "merge" or "not merge" to the processing device, so as to update the master index database (such as merging the records of the same patient and keeping the independent indexes of different patients).

[0034] In some embodiments, the obtaining of the field data of the target field in the plurality of candidate records comprises: obtaining a plurality of patient records and record features of the patient records; when it is detected that the value of any target field in the patient record is missing, obtaining a set of context fields of a to-be-completed field in the candidate record, an observation value of each context field, a quality confidence of each context field, a temperature coefficient, and a candidate value of the to-be-completed field, wherein the temperature coefficient is used to control the sharpness of the weight distribution of the context field weight; determining the corresponding context field weight based on the quality confidence of the context field and the temperature coefficient; constructing an expected representation of the candidate value in the context field based on the structural relationship between the context field and the candidate value of the to-be-completed field; performing squared difference calculation on the expected representation of the candidate value in the context field and the observation value of the context field to obtain a difference measure of the context field; traversing the set of context fields, and constructing a risk function about the candidate value based on the context field weight, the difference measure of the context field, a preset priori penalty of the context field, and a preset regularization strength; determining the candidate value that minimizes the risk function as the completion value of the to-be-completed field; Based on the patient record, the patient record containing the completion value of the to-be-completed field is filtered to obtain a plurality of candidate records, wherein the record features include hash signatures, numerical value differences, and character distances.

[0035] In this embodiment, for sources such as hospital information systems, laboratory information systems, picture archiving and communication systems, and electronic medical record systems, field semantic mapping, coding and format specification, time and identifier normalization, and other operations can be performed, and quality confidence for each field is generated through format verification and cross-field consistency checking, which is used to depict the reliability of the field in the patient record (such as patient basic information, diagnosis and treatment record, etc.).

[0036] Specifically, the fields in the patient record are completed, and then the completed patient record is filtered by the patient record to obtain a plurality of candidate records. More specifically, by using low-complexity hash signatures, standardized numerical value differences, or simplified character distance calculations, obviously unmatched patient records can be quickly removed, improving the efficiency of subsequent processing of candidate records and reducing invalid calculations.

[0037] In terms of missing completion, the context field set of the to-be-completed field, such as the context fields of “age” which can include “date of birth” and “year of treatment”, the observed value of the context field, i.e. the actual record value of the context field, the quality confidence of the context field, reflecting the reliability of the field data, such as the high confidence of the field passing the verification, the temperature coefficient representing the sharpness (concentration) of the control context field weight distribution, the candidate value of the to-be-completed field, which is a possible missing value candidate, such as a possible value based on historical data, and the context field weight representing the relative importance of the context field in the completion decision.

[0038] Then, the weight of each context field is calculated in combination with the quality confidence (reliability) and temperature coefficient of the context field, taking into account the data quality and reducing the interference of low-quality data. Further, the structural relationship between the context field and the to-be-completed field, for example, the logical relationship of “age” = “year of treatment” - “year of birth”, uses the internal logical association between fields (such as business rules and data association), and makes reasonable inferences that fit the context and actual semantic scenarios. Through the calculation of the squared difference, the difference between the expected representation corresponding to the candidate value and the actual observed value of the context field is determined. The smaller the difference, the more matched the candidate value is to the context, improving the accuracy of the completed field. Then, by traversing the context field to construct a risk function and minimizing the risk function, the completion value of the to-be-completed field is obtained, ensuring that the completion value is consistent with the context and reasonable in business, thereby improving the data integrity.

[0039] Therefore, the embodiment preserves the dominant role of high-quality fields, improves the accuracy and relevance of data completion, and greatly reduces the probability of matching errors caused by missing data.

[0040] In some embodiments, the field data includes data quality indicators and field importance priors; and determining a matching decision factor of the record pair based on the field data of the target field comprises: determining a field weight of the target field based on the data quality indicators, the field importance priors, and a preset temperature parameter, wherein the temperature parameter is used to indicate the sensitivity of the data quality indicators to the field weight; obtaining a field similarity of the target field and a variance estimate value of the field similarity, wherein the field similarity is used to indicate the matching degree of the record pair in the same target field; determining a matching score of the record pair based on the field similarity of the target field and the corresponding field weight; obtaining an uncertainty degree based on the field weight of the target field and the variance estimate value of the field similarity; obtaining a risk adjustment score as the matching decision factor based on the matching score, the uncertainty degree, and a preset adjustable risk coefficient.

[0041] In the embodiment, the data quality indicators are used to indicate the reliability of the field data, the field importance priors (such as the business priority of core fields such as ID number and name) represent the inherent importance of the field in matching, and the temperature parameter is used to control the sensitivity of the quality factor to the field weight. Therefore, when the data quality is high, the field weight is close to the business prior distribution; when the quality of some fields is significantly reduced, the weight will be exponentially attenuated to reduce the interference of low-quality fields on the matching result. By increasing the flexibility of weight calculation through the temperature parameter, the weight is more suitable for the scene requirements, and the accuracy of the field weight is improved.

[0042] Then, the field similarity represents the matching degree of the two candidate records on the same target field, and the variance estimation value reflects the fluctuation of the field similarity in the statistical historical data. Based on the field similarity of the target field and the corresponding field weight, the matching score of the record pair is determined, which reflects the comprehensive score of the overall matching degree of the two candidate records. Further, based on the field weight of the target field and the variance estimation value of the field similarity, the uncertainty is obtained, which reflects the fluctuation risk of the matching score. Therefore, the embodiment considers the uncertainty to reflect the reliability problem of the overall matching score. Further, based on the matching score, the uncertainty and the adjustable risk coefficient (reflecting the tolerance of the business to the risk), the risk adjustment score is obtained, which corrects the original matching score through the uncertainty and enhances the robustness of the judgment. Taking the risk adjustment score as the matching decision factor, considering the field similarity, the weight and the uncertainty, the record pair is divided into "directly merged" (matched pair), "directly excluded" (unmatched pair) and "suspected record" (suspected matched pair), so as to realize data shunting.

[0043] Therefore, the embodiment constructs the matching decision factor through the matching score and the uncertainty, avoids using only the matching score for decision making, considers the matching reliability, and can reduce the probability of false merging or missing merging of data.

[0044] In some embodiments, the matching decision factor of the record pair is determined based on the field data of the target field, and when the matching decision factor falls within a pre-calculated suspected matching interval, the corresponding record pair is identified as a suspected matched pair, and the method further comprises: obtaining the mean and standard deviation of the matching score in the specified period of time; determining the upper limit value of the suspected matching interval based on the mean, the standard deviation and a preset first adjustment coefficient; determining the lower limit value of the suspected matching interval based on the mean, the standard deviation and a preset second adjustment coefficient; when the matching decision factor is greater than the upper limit value of the suspected matching interval, identifying the corresponding record pair as a matched pair and generating a corresponding confidence score; when the matching decision factor is less than the lower limit value of the suspected matching interval, identifying the corresponding record pair as an unmatched pair.

[0045] In the embodiment, the overall level (average) and fluctuation degree (standard deviation) of past matching results are quantified based on historical data, providing data basis for subsequent calculation of suspected matching interval, and reflecting changes in recent data characteristics. Then, based on the average, standard deviation, and first and second adjustment coefficients, the upper and lower limits of the suspected matching interval are flexibly adjusted, improving the accuracy of suspected matching judgment. Thus, through multi-threshold comparison, all record pairs are divided into matching pairs (directly merged), non-matching pairs (directly excluded), and suspected matching pairs (to be further processed), improving processing efficiency. The confidence score represents the matching confidence of the matching pair.

[0046] Therefore, the embodiment dynamically generates the upper and lower limits of the suspected matching interval based on the average and standard deviation of the recent historical matching score, thereby dividing the record pairs, achieving flexibility and accuracy of threshold data, avoiding interference of low-quality fields on the overall result, improving the shunt accuracy, increasing the proportion of high-confidence matching, reducing the number of suspected matching, and thus reducing the burden of manual review.

[0047] In some embodiments, the suspected matching interval is updated by the following steps, including: Within a preset sliding window length, the average of the matching score at the current time is determined based on the matching score at the current time, the average of the matching score at the previous time, and a preset smoothing coefficient. The upper and lower limits of the suspected matching interval are updated using the average of the matching score at the current time, and the data quality indicator at the current time is updated.

[0048] To cope with the matching characteristic drift caused by the change of data stream over time, a sliding window statistics and exponential smoothing update mechanism is introduced to update the average and standard deviation of the matching score and the data quality indicator of each field in real time, ensuring that the rule remains optimal under different time periods and data source combinations. Therefore, the embodiment dynamically adjusts the suspected matching interval according to data changes, improves the accuracy of data classification, and reduces the probability of incorrect or missed merging of data. Through the synchronous update of the data quality indicator, the calculation accuracy of the subsequent field weight is improved, thereby improving the accuracy of the matching decision factor, which is also beneficial to improving the accuracy of data classification.

[0049] In some embodiments, the edge selection in each suspected cluster graph model is iteratively optimized until the consistency in the suspected cluster graph model is maximized and the conflicting edge selection is minimized, and based on the current merging tendency strength of the edge, the suspected cluster graph model is divided into a first suspected cluster to be merged and a second suspected cluster to be audited, including: construct an energy function corresponding to each of the suspected cluster graph models, wherein the energy function comprises a consistency term and a conflict penalty term, the consistency term is composed of the matching decision factor of the selected edge and the merging tendency strength, and the conflict penalty term is composed of a preset conflict penalty coefficient and the merging tendency strength of two conflict edges selected at the same time; iteratively adjust the merging tendency strength of the edge of the suspected cluster graph model until the consistency term is maximized and the conflict penalty term is minimized, and obtain the current merging tendency strength of the edge; screen out the suspected cluster graph model satisfying a preset merging condition as a first suspected cluster to be merged, generate a corresponding merging confidence, and divide the remaining suspected cluster graph models into second suspected clusters to be audited, wherein the merging condition comprises that the number of target edges in the suspected cluster graph model is greater than a preset number threshold, and the proportion of target conflict edge pairs in the suspected cluster graph model is lower than a preset proportion threshold, the target edge is used to describe an edge whose merging tendency strength is greater than a preset merging threshold, and the target conflict edge pair is used to indicate a conflict edge pair that simultaneously satisfies the merging tendency strength greater than the merging threshold.

[0050] In the embodiment, an energy function is constructed as follows: wherein, E represents the energy function, Ei represents an edge of the suspected cluster graph model, E represents an edge set, ti represents the merging tendency strength of the edge, pi represents the matching decision factor, ci represents the conflict penalty coefficient, Eij represents a conflict edge pair set, Eij represents one conflict edge pair, ti,j represents the merging tendency strength of one edge in the conflict edge pair.

[0051] In the embodiment, the consistency term is used to indicate the matching correlation of the records in the cluster (the stronger the merging tendency, the higher the consistency), and the records with high similarity are preferentially merged to form a cluster with internal consistency, and the conflict penalty term is used to specifically inhibit the contradictory relationship and reduce the conflict. Therefore, by minimizing the energy function, the matching correlation in the cluster is maximized and the conflict edge selection is minimized. In the iteration process, the deviation of the initial edge selection can be gradually corrected, that is, the merging tendency strength of the edge is iteratively adjusted until the energy function is minimized, and the accuracy of the cluster division is improved.

[0052] Further, at the end of iteration, if the merging tendency strength of most edges in the cluster is greater than the merging threshold, it is considered that the merging tendency strength of the edge meets the standard, which indicates that the matching consistency between records in the cluster is high, most edges tend to merge, and the proportion of conflict edges in the cluster is low, i.e., there is almost no serious conflict in the cluster, and the risk of merging is low. Therefore, it is considered that the suspected cluster graph model meets the merging condition, and the suspected cluster graph model is divided into a first suspected cluster to be merged. Optionally, the merging condition can also include high overall matching strength, i.e., the average value of all edges in the cluster is large, to ensure the overall matching quality of the cluster.

[0053] In this case, the suspected cluster graph model that does not meet the merging condition is divided into a second suspected cluster to be audited. The second suspected cluster can have the following characteristics: first, there is significant conflict, i.e., the proportion of the target conflict edge pair in the conflict edge pair in the cluster is high; second, the merging tendency is ambiguous, i.e., the merging tendency strength of most edges in the cluster is low, i.e., the matching consistency is moderate, and it cannot be determined; and third, local low matching: the merging tendency strength of some key edges (such as edges involving key fields such as names and ID numbers) is lower than the merging threshold, which can have potential risks.

[0054] Therefore, the embodiment maps the suspected matching pairs to a suspected cluster graph model, constructs an energy function containing consistency and conflict, minimizes the energy function, and divides high-confidence merging clusters (first suspected clusters) and conflict clusters to be audited (second suspected clusters), thereby greatly reducing the amount of manual auditing, achieving shunt processing, improving shunt accuracy, greatly improving auditing efficiency, and reducing the probability of false merging or missed merging of data.

[0055] In some embodiments, the deep discriminant model includes a cascaded ordering multi-layer perception classifier and a message passing model; and the deep feature analysis of the second suspected cluster by using the preset deep discriminant model to obtain an auxiliary decision label includes: converting the candidate records in the second suspected cluster into feature vectors; calculating a feature difference vector based on the feature vectors of any two candidate records in the second suspected cluster; performing feature processing on the feature difference vector by using the multi-layer perception classifier to obtain a merging probability of the record pair; constructing a target graph model based on the second suspected cluster, wherein a target node of the target graph model represents a candidate record in the second suspected cluster, and a target edge of the target graph model represents a merging probability between two connected candidate records; inputting the target graph model into the message passing model, iteratively updating node features of any target node in the target graph model by fusing node-related data of neighbor nodes of the target node, and outputting the node features of each target node in the target graph model; obtaining a discriminant evaluation value of the second suspected cluster based on the merging probability of each target edge in the target graph model and the node features of each target node; determining corresponding confidence information based on the merging probability of the record pairs in the second suspected cluster; comparing the discriminant evaluation value with a preset evaluation threshold, and determining a corresponding auxiliary decision label based on a comparison result and the confidence information.

[0056] In this embodiment, the candidate records are converted into uniform feature vectors containing structured fields and unstructured text features, taking into account the expression ability of different types of data while retaining field importance and semantic information. That is, by converting the candidate records of the second suspected cluster into feature vectors, this embodiment retains the relative importance between fields and can also capture the similarity changes caused by medical entry differences (such as abbreviations, misspellings, language switching, etc.) in the semantic space.

[0057] Further, for any two candidate records in the second suspected cluster, a feature difference vector is calculated. Then, the feature difference vector is input into a multi-layer perceptron classifier (such as containing multiple hidden layers and a nonlinear activation function), and a merging probability is output, which is used to indicate the probability that the two candidate records belong to the same patient. Therefore, by predicting the merging probability through the multi-layer perceptron classifier, complex nonlinear feature interactions are fitted, and the prediction accuracy of the merging probability is improved.

[0058] To further utilize the intra-cluster consistency information, this embodiment introduces a graph neural network structure, treats the second suspected cluster as a weighted undirected graph, takes the candidate records in the second suspected cluster as target nodes, and takes the edge weight of the target edge as the merging probability, to construct a target graph model. Then, the node features of the target nodes in the target graph model are iteratively updated through a message passing mechanism (i.e., a message passing model). After several layers of iteration, the intra-cluster node features will fuse the related information of their neighbor nodes, i.e., the node features of the target nodes fuse their own features and the node features of the neighbor nodes, to realize global information fusion. For the conflicting triangular relationships, the influence of conflicts can be weakened through weighted fusion of neighbor information, to improve the decision robustness.

[0059] In the embodiment, the confidence information is generated by counting the merging probability distribution of all record pairs in the second suspected cluster, quantifying the reliability and volatility of the merging probability, and the confidence information can reflect the matching degree and the stability of the matching result. Secondly, the discriminant evaluation value reflecting the entire second suspected cluster is calculated by comprehensively considering the merging probability of all target edges in the target graph model and the updated node features of all target nodes. The discriminant evaluation value can be used to indicate the matching consistency of the entire cluster. Therefore, the auxiliary decision label is generated by combining the confidence information and the discriminant evaluation value, avoiding misjudgment caused by a single threshold, making the auxiliary label more consistent with the actual matching state of the cluster, and at the same time, the auxiliary decision label clearly indicates the processing priority (priority score) and the focus point (conflict mark) of each cluster, reduces the blindness of manual judgment, speeds up the review process, and thus improves the review processing efficiency.

[0060] Therefore, the embodiment improves the accuracy of the auxiliary decision label by fusing the global correlation features in the second suspected cluster through the deep discriminant model including the cascade ordered multi-layer perception classifier and the message passing model, thereby improving the review processing efficiency.

[0061] In some embodiments, the patient master index database is updated according to the review confirmation result from the user terminal: For any second suspected cluster, a priority score of the second suspected cluster is determined based on a pre-acquired key field coverage, a conflict degree, a preset weight coefficient, and a corresponding matching decision factor, and a corresponding conflict mark is generated, wherein the key field coverage is used to indicate the complete proportion of candidate records in the second suspected cluster on the specified key field, the conflict degree is used to indicate the conflict degree of record pairs in the second suspected cluster on the specified core field, and the conflict mark is used to indicate the identification of the conflict of record pairs in the second suspected cluster on the key field. The second suspected cluster is processed in order according to the priority score to obtain a review confirmation result of the second suspected cluster.

[0062] In the present embodiment, the conflict degree represents the degree of conflict of records in a cluster on core fields (such as identity card number, medical insurance number), the key field coverage degree represents the complete proportion of records in a cluster on key fields (such as name, date of birth), and the preset weight coefficient is used to adjust the relative importance of the key field coverage degree, the conflict degree, and the matching decision factor. Based on the key field coverage degree, the conflict degree, the weight coefficient, and the corresponding matching decision factor, the priority score of the second suspected cluster is determined, which takes into account the importance, urgency, and conflict of the second suspected cluster. Then, according to the generated priority score, all second suspected clusters are sorted in descending order, and the cluster with the highest score is processed first. The reviewer processes them in turn according to the sorting results, and finally generates the review confirmation results such as "merge", "do not merge", "need further supplement information", etc.

[0063] Therefore, the present embodiment takes into account the information integrity, conflict risk, and basic matching strength of the cluster, ensuring that important clusters are processed first. By prioritizing clusters with high priority into the review process, the efficiency of the review and the risk reduction effect are maximized.

[0064] In some embodiments, the patient master index database is updated according to the review confirmation results from the user terminal: Based on the merge confidence threshold of the previous time step, the preset smoothing coefficient, and the average confidence, the merge confidence threshold of the current time step is updated, wherein the average confidence is composed of a first average confidence and a second average confidence, the first average confidence is determined by the confidence information corresponding to the second suspected cluster whose review confirmation result is to confirm merging, and the second average confidence is determined by the confidence score and / or the merge confidence corresponding to the record pair that passes the review in the review queue; When the confidence score of the matching pair is greater than the merge confidence threshold, and the merge confidence of the first suspected cluster is greater than the merge confidence threshold, the matching pair and the first suspected cluster are merged into the patient master index database; The matching pair whose confidence score is less than the merge confidence threshold and the first suspected cluster whose merge confidence is less than the merge confidence threshold are added to the review queue, so that in the review queue, the corresponding review request is sent to the user terminal in turn.

[0065] In the present embodiment, for the record pairs to be merged (i.e. the first suspected cluster and the matching pair), when the corresponding merge confidence (corresponding to the first suspected cluster) and the confidence score (corresponding to the matching pair) are greater than the merge confidence threshold, automatic merging is triggered, otherwise it is transferred to the review queue. The merge confidence threshold is calculated and updated by the following formula, including: ; Wherein, is a smoothing coefficient, is an average confidence, is a merging confidence threshold of a previous time step, is a merging confidence threshold of a current time step. Thus, the merging confidence threshold in the present application is increased with the increase of the average confidence of the last batch of manually confirmed merging record pairs, which can reduce the probability of false merging or missed merging of data.

[0066] Optionally, in actual execution of merging, to prevent multiple records from forming a ring conflict or causing "pollution propagation" due to partial false merging, a structured merging algorithm based on a union-find set is used. For the structured merging algorithm, each primary index corresponds to a patient identifier set, and through path compression and rank-merge strategy, the merging operation is ensured to be completed in approximate time, while maintaining the merging history log record operation order, source module and confidence for subsequent backtracking and revocation. If it is detected in the merging process that two records respectively belong to different primary index clusters that have existed conflict markers, a conflict coordination sub-process is triggered.

[0067] Optionally, the conflict coordination process makes decision optimization based on the principle of minimum information loss. Assuming that two conflict clusters respectively contain field sets , the business importance weight of field is , and the objective function of conflict resolution is: ; wherein is a reserved field set, is the quality confidence of field in the merged version. By using a greedy or dynamic programming method, the field set is selected, which maximizes the reservation of high-quality and high-value information while deleting or marking low-quality conflict fields. This optimization not only solves the field-level conflict, but also bifurcates the entire record version if necessary, to meet the multi-version management needs of conflict data in business.

[0068] After the update is completed, the module generates a unique change transaction ID for each primary index change, and writes the change details (including trigger reason, participating record ID, old index ID, new index ID, change type, confidence, conflict handling strategy, etc.) into the version control library.

[0069] Figure 2 is a structural schematic diagram of a patient primary index management device provided by an embodiment of the present application. The patient primary index management device 200 comprises: a data acquisition module 201 configured to acquire field data of a target field in a plurality of candidate records; The matching judgment module 202 is configured to determine a matching decision factor of a record pair based on field data of the target field, and identify the corresponding record pair as a suspected matching pair when the matching decision factor falls within a pre-calculated suspected matching interval, the record pair being composed of any two candidate records; The graph model construction module 203 is configured to construct at least one suspected cluster graph model based on the matching decision factor of the suspected matching pair, wherein a node of the suspected cluster graph model represents one candidate record in the suspected matching pair, and an edge weight of the suspected cluster graph model represents the matching decision factor of the corresponding suspected matching pair. The optimization module 204 is configured to iteratively optimize edge selection in each suspected cluster graph model until consistency maximization and conflict edge selection minimization are achieved in the suspected cluster graph model, and divide the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be audited based on a current merging tendency strength of the edge. The auxiliary decision module 205 is configured to perform deep feature analysis on the second suspected cluster by using a preset deep discriminant model to obtain an auxiliary decision label. The updating module 206 is configured to update the patient master index library according to an audit confirmation result from a user terminal, wherein the audit confirmation result is generated by the user terminal based on an audit on the candidate record corresponding to the second suspected cluster according to the auxiliary decision label.

[0070] The apparatuses of the embodiments of the present application can perform the methods provided by the embodiments of the present application, and the implementation principles are similar. The actions performed by each module in the apparatuses of the embodiments of the present application are corresponding to the steps in the methods of the embodiments of the present application. The detailed function description of each module of the apparatuses can be referred to the description of the corresponding method in the foregoing description, and will not be repeated here.

[0071] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A patient master index management method characterized by comprising: For use in processing equipment, including: Retrieve field data for a target field from multiple candidate records; Based on the field data of the target field, a matching decision factor for the record pair is determined, and when the matching decision factor falls within a pre-calculated suspected matching interval, the corresponding record pair is identified as a suspected matching pair, wherein the record pair consists of any two candidate records; Based on the matching decision factors of the suspected matching pairs, at least one suspected cluster graph model is constructed, wherein the nodes of the suspected cluster graph model represent a candidate record in the suspected matching pair, and the edge weights of the suspected cluster graph model represent the matching decision factors of the corresponding suspected matching pairs. Iteratively optimize the edge selection in each suspected cluster graph model until the consistency within the suspected cluster graph model is maximized and the selection of conflicting edges is minimized. Based on the current edge merging tendency strength, divide the suspected cluster graph model into a first suspected cluster to be merged and a second suspected cluster to be reviewed. The second suspected cluster is subjected to deep feature analysis using a preset deep discrimination model to obtain auxiliary decision labels; Based on the review and confirmation results from the user terminal, the patient master index is updated, wherein the review and confirmation results are generated by the user terminal after reviewing the candidate records corresponding to the second suspected cluster based on the auxiliary decision-making tags.

2. The patient master index management method according to claim 1, characterized by, The step of obtaining the field data of the target field from multiple candidate records includes: Obtain multiple patient records and the record characteristics of the patient records; When a missing value is detected in any target field in the patient record, the set of context fields of the field to be completed, the observed value of each context field, the quality confidence of each context field, the temperature coefficient, and the candidate values ​​of the field to be completed are obtained in the candidate record. The temperature coefficient is used to control the sharpness of the weight distribution of the context field weights. The corresponding context field weight is determined based on the quality confidence level of the context field and the temperature coefficient. Based on the structural relationship between the context field and the candidate values ​​of the field to be completed, construct the expected representation of the candidate values ​​in the context field; The squared difference is calculated on the expected representation of the candidate value in the context field and the observed value of the context field to obtain the difference measure of the context field; Traverse the set of context fields, and construct a risk function for the candidate values ​​based on the context field weights, the difference measures of the context fields, the preset prior penalties of the context fields, and the preset regularization strength; The candidate values ​​that minimize the risk function are determined and used as the completion values ​​for the field to be completed. Based on the patient records, the patient records containing the completion values ​​of the fields to be completed are filtered to obtain multiple candidate records, wherein the record features include hash signature, numerical difference and character distance.

3. The patient master index management method according to claim 2, characterized by, The field data includes data quality indicators and prior importance of the fields; The determination of matching decision factors for record pairs based on the field data of the target field includes: Based on the data quality index of the target field, the prior importance of the field, and the preset temperature parameter, the field weight of the target field is determined, wherein the temperature parameter is used to indicate the sensitivity of the data quality index to the field weight; Obtain the field similarity of the target field and the variance estimate of the field similarity, wherein the field similarity is used to indicate the degree of matching of the record pair on the same target field; Based on the field similarity of the target field and the corresponding field weight, the matching score of the record pair is determined; The uncertainty is obtained based on the field weights of the target field and the variance estimate of the field similarity. Based on the matching score, the uncertainty, and the preset adjustable risk coefficient, a risk adjustment score is obtained and used as the matching decision factor.

4. The patient master index management method according to claim 3, characterized by, The step of determining a matching decision factor for record pairs based on the field data of the target field, and identifying the corresponding record pair as a suspected matching pair when the matching decision factor falls within a pre-calculated suspected matching interval, further includes: Get the mean and standard deviation of the matching scores over a specified period of time. Based on the mean, the standard deviation, and the preset first adjustment coefficient, the upper limit of the suspected matching interval is determined; Based on the mean, the standard deviation, and the preset second adjustment coefficient, the lower limit of the suspected matching interval is determined; When the matching decision factor is greater than the upper limit of the suspected matching interval, the corresponding record pair is identified as a matching pair, and a corresponding confidence score is generated. When the matching decision factor is less than the lower limit of the suspected matching interval, the corresponding record pair is identified as a non-matching pair.

5. The patient master index management method according to claim 4, characterized by, The suspected matching interval is updated by the following steps: Within a preset sliding window length, the mean of the matching score at the current moment is determined based on the mean of the matching score at the current moment, the mean of the matching scores at the previous moment, and a preset smoothing coefficient. Using the mean of the matching scores at the current moment, update the upper and lower limits of the suspected matching interval, and update the data quality indicators at the current moment.

6. The patient master index management method according to claim 5, wherein The process iteratively optimizes edge selection in each suspected cluster graph model until consistency within the suspected cluster graph model is maximized and conflict edge selection is minimized. Based on the current edge merging tendency strength, the suspected cluster graph model is divided into a first suspected cluster to be merged and a second suspected cluster to be reviewed, including: Based on each of the suspected cluster graph models, a corresponding energy function is constructed, wherein the energy function includes a consistency term and a conflict penalty term. The consistency term consists of the matching decision factor and merging tendency strength of the selected edge, and the conflict penalty term consists of a preset conflict penalty coefficient and the merging tendency strength of two conflicting edges selected at the same time. Iteratively adjust the merging tendency strength of the edges in the suspected cluster graph model until the consistency term is maximized and the conflict penalty term is minimized, and obtain the current edge merging tendency strength; Suspected cluster graph models that meet preset merging conditions are selected as the first suspected cluster to be merged, and corresponding merging confidence scores are generated. The remaining suspected cluster graph models are classified as the second suspected cluster to be reviewed. The merging conditions include: the number of target edges in the suspected cluster graph model is greater than a preset number threshold, and the proportion of target conflict edge pairs in the suspected cluster graph model is lower than a preset proportion threshold. The target edges are used to describe edges whose merging tendency is greater than the preset merging threshold, and the target conflict edge pairs are used to indicate conflict edge pairs that simultaneously satisfy the condition that their merging tendency is greater than the merging threshold.

7. The patient master index management method according to claim 6, wherein The deep discrimination model includes a cascaded multilayer perceptron classifier and a message passing model; the step of using the preset deep discrimination model to perform deep feature analysis on the second suspected cluster to obtain auxiliary decision labels includes: Convert the candidate records of the second suspected cluster into feature vectors; Calculate the feature difference vector based on the feature vectors of any two candidate records within the second suspected cluster; The feature difference vector is processed using the multilayer perceptron classifier to obtain the merging probability of the record pairs; Based on the second suspected cluster, a target graph model is constructed, wherein the target node of the target graph model represents the candidate record in the second suspected cluster, and the target edge of the target graph model represents the merging probability between two connected candidate records; The target graph model is input into the message passing model. By fusing the node-related data of the neighboring nodes related to any target node in the target graph model, the node features of the target node are iteratively updated, and the node features of each target node in the target graph model are output. Based on the merging probability of each target edge and the node characteristics of each target node in the target graph model, the discrimination evaluation value of the second suspected cluster is obtained. Based on the merging probability of the record pairs within the second suspected cluster, the corresponding confidence information is determined; The discrimination evaluation value is compared with a preset evaluation threshold, and the corresponding auxiliary decision label is determined based on the comparison result and the confidence information.

8. The patient master index management method according to claim 7, characterized by, Based on the verification results from the user terminal, the patient master index is updated: For any second suspected cluster, based on the pre-acquired key field coverage, conflict degree, preset weight coefficient and corresponding matching decision factor, the priority score of the second suspected cluster is determined and a corresponding conflict mark is generated. The key field coverage is used to indicate the complete proportion of candidate records in the second suspected cluster on the specified key field, the conflict degree is used to indicate the degree of conflict between record pairs in the second suspected cluster on the specified core field, and the conflict mark is used to indicate the identifier of conflict between record pairs in the second suspected cluster on the key field. The second suspected cluster is processed sequentially according to the priority score to obtain the review and confirmation result of the second suspected cluster.

9. The patient master index management method according to claim 8, wherein, Based on the verification results from the user terminal, the patient master index is updated: Based on the merge confidence threshold of the previous time step, the preset smoothing coefficient, and the average confidence, the merge confidence threshold of the current time step is updated. The average confidence consists of a first average confidence and a second average confidence. The first average confidence is determined by the confidence information corresponding to the second suspected cluster whose review confirmation result is confirmed as merged. The second average confidence is determined by the confidence score and / or merge confidence of the records that have passed the review in the review queue. When the confidence score of the matched pair is greater than the combined confidence threshold, and the combined confidence score of the first suspected cluster is greater than the combined confidence threshold, the matched pair and the first suspected cluster are merged into the patient master index database. The matching pairs whose confidence scores are less than the combined confidence threshold, and the first suspected cluster whose combined confidence scores are less than the combined confidence threshold, are added to the review queue, so that corresponding review requests are sent to the user terminal in the review queue in sequence.

10. A patient master index management apparatus characterized by comprising: include: The data acquisition module is used to acquire field data of the target field from multiple candidate records; The matching judgment module is used to determine the matching decision factor of the record pair based on the field data of the target field, and when the matching decision factor falls within the pre-calculated suspected matching interval, the corresponding record pair is identified as a suspected matching pair, wherein the record pair consists of any two candidate records; The graph model construction module is used to construct at least one suspected cluster graph model based on the matching decision factors of the suspected matching pairs, wherein the nodes of the suspected cluster graph model represent a candidate record in the suspected matching pair, and the edge weights of the suspected cluster graph model represent the matching decision factors of the corresponding suspected matching pairs. An optimization module is used to iteratively optimize the edge selection in each suspected cluster graph model until the consistency within the suspected cluster graph model is maximized and the selection of conflicting edges is minimized. Based on the current edge merging tendency strength, the suspected cluster graph model is divided into a first suspected cluster to be merged and a second suspected cluster to be reviewed. The decision support module is used to perform deep feature analysis on the second suspected cluster using a preset deep discrimination model to obtain decision support labels. The update module is used to update the patient master index database based on the review and confirmation results from the user terminal, wherein the review and confirmation results are generated by the user terminal after reviewing the candidate records corresponding to the second suspected cluster based on the auxiliary decision tags.

Images